Keywords: Control Theory and Applications
Abstract: 本文探讨了 switched 的绝对稳定性 Lurie 系统在模式依赖最小停留条件下 时间（MDMDT）策略。绝对的充分条件 稳定性通过二次方位建立 离散化的李雅普诺夫–卢里函数。与 传统的线性离散化李雅普诺夫函数 方法中，所提方法能够实现更准确的拟合， 从而显著降低了 稳定性标准。理论的有效性 结果通过数值示例进行展示。

Keywords: Control Theory and Applications, Autonomous Vehicle Systems
Abstract: In this paper, we study a zero-order-hold (ZOH) triggered gain-scheduling controller for hovering control of a quadrotor under time-varying mass. The controller is designed by treating the time-varying mass as the scheduling variable. Due to the ZOH mechanism, the proposed controller is updated intermittently, which provides the advantage of reducing communication resource usage and computational burden. The stability of the closed-loop quadrotor system is established using Lyapunov theory and the Razumikhin theorem. It is shown that the closed-loop system remains bounded, and the ultimate bound is explicitly characterized as a function of the interexecution time. Finally, the effectiveness and validity of the proposed method are demonstrated through simulation studies.

Keywords: Control Theory and Applications, Autonomous Vehicle Systems
Abstract: Motorcycle manoeuvring at low speeds is particularly demanding due to the reduced stability, which increases the rider’s effort during cornering. This paper presents a steering-torque assistance strategy designed to support the rider by reducing effort while maintaining stable and accurate roll tracking. Modelling the rider as a PD controller acting on the roll angle of the bike, we investigate a co-operative control framework in which an inverse rider model is used to estimate the rider's internal tracking error, which is fed to a controller providing torque assistance on the steering column of the bike. Simulation results using a validated linearised model from the motorcycle simulation software FastBike demonstrate that the proposed approach reduces the rider’s peak torque from 3.5 Nm to 1.5 Nm, while also improving stability during low-speed cornering. To reflect realistic variations in rider behavior, ±20% changes in rider parameters are considered, and results show that the error signal is still estimated accurately with this model mismatch.

Keywords: Control Theory and Applications, Autonomous Vehicle Systems, Industrial Applications of Control
Abstract: This paper presents a parametric analysis of scenario-based conditional value-at-risk model predictive control (CVaR-MPC) for adaptive cruise control under stochastic leading-vehicle behavior modeled by Gaussian mixture models (GMMs). We study the effect of the CVaR confidence level &alpha: under profile-matched and mismatched plant/controller GMM libraries, benchmarking CVaR-MPC against nominal MPC and RSS-soft, a soft-penalty baseline based on Responsibility-Sensitive Safety (RSS). We tested 1,000 rollouts under an under-prepared mismatch, where the controller used a milder GMM profile than the plant. Increasing alpha from 0 to 0.99 reduced the fraction of rollouts violating the velocity-dependent time-headway boundary $$d_risk(v_e) from 48.4 to 40.8. Over the same range, the 5th percentile of the minimum inter-vehicle distance increased from 1.30 to 2.07m. Even at alpha=0, CVaR-MPC reduced the d_risk violation rate from 62.5% for nominal MPC to 48.4% because its empirical CVaR term remains active as a positively weighted sample-mean risk penalty. RSS-soft yielded larger minimum inter-vehicle distances, but these did not translate into a lower violation rate for $$d_risk. This discrepancy arises from the different velocity dependence of the RSS and time-headway boundaries, which can reverse their ordering. Together, these results clarify how alpha tunes lower-tail protection within CVaR-MPC and why comparisons with RSS-soft depend on the safety metric used for evaluation.

Keywords: Control Theory and Applications, Industrial Applications of Control
Abstract: This paper establishes a converse theorem for strict control barrier functions. For merely continuous control-affine systems, we formulate safety as forward completeness of the closed-loop system on the safe set, requiring every maximal solution to be complete. We prove that the existence of a strict control barrier function is equivalent to the existence of a continuous safe feedback. The key ingredients of the proof are a partition-of-unity feedback construction and a Lyapunov-like characterization of forward completeness. We also show that, whenever a continuous safe feedback exists, a smooth safe feedback can be chosen.

Keywords: Control Theory and Applications
Abstract: We applied the Mean Exponential Growth Factor of Nearby Orbits (MEGNO) to the gain-space of a PD-controlled two-axis coupled gimbal system. The MEGNO is a “fast-indicator” that has been developed to identify chaotic from quasi-periodic structures of dynamical systems (originally in astrodynamics), yet to be applied to PID-controlled systems. MEGNO serves as the main engine for the analysis of our system, used with an additional external classification layer (later introduced as Eigenvalue-Comparison Classification), based on a newly established relationship that states that the closed-loop Jacobian’s dominant eigenvalue λ* predicts the asymptotic MEGNO slope as λ*/2; allowing us to use it in systems with an isolated dissipative equilibrium. In the used apparatus the Eigenvalue-Comparison Classification (ECC) method successfully identified the absence of bounded chaotic regimes across 927 stable gain points with the maximum residual between the predicted and the measured slope being 0.024, meaning the system is cleanly bimodal.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Control Theory and Applications
Abstract: Two-wheeled legged robots combine the efficiency of wheeled locomotion with the posture regula- tion capability of legged mechanisms. However, stabilizing body height and attitude remains challenging under disturbances and varying contact conditions because wheel-ground contact and leg joint dynamics are coupled. Although reinforcement learning has been applied to such robots, position-based control supports stable policy learning but limits direct torque regulation, whereas torque-based control enables direct torque regulation but makes learning more difficult. To address this trade-off, this paper proposes a reinforcement learning-based ac- tion structure that generates joint torques through an MIT impedance controller rather than directly outputting them from the policy. The policy outputs target position offsets and feedforward torques for the leg joints, and target velocities and feedforward torques for the wheel joints. The controller combines feedback torques with policy-generated feedforward torques, preserving a stable feedback structure while providing additional torque for post-disturbance recovery. The proposed policy was trained in an IsaacLab-based simulation environment and compared with position-based and torque-based control policies under vertical disturbances. Experimental results show that the proposed MIT impedance control achieved up to 13.9% lower body height drop and 36.5% shorter recovery time than position control, and up to 15.4% lower body height drop and 58.2% shorter recovery time than torque control.

Keywords: Robotic Applications, Sensors and Signal Processing, Robot Vision
Abstract: Tracked vehicles provide stable traction and high mobility on rough terrain, but their dynamic behavior is difficult to predict because many track links, sprockets, idlers, and terrain contacts interact simultaneously. This study presents an experimental validation procedure for a dynamic simulation model of a small tracked robot designed for laboratory rough-terrain mobility tests. The main contribution is to incorporate experimentally identified track-chain backlash and stiffness into a Project Chrono multibody dynamics model and to validate the resulting model directly against the physical robot using a vision-based three-dimensional measurement system. A compact tracked robot, an adjustable driving testbed, a chain stiffness measurement testbed, and a camera-based measurement system were developed. In the simulation, adjacent chain links are connected by 6-DOF bushing elements that represent hinge motion, backlash, and stiffness characteristics. Validation focuses on slope-driving and obstacle-climbing behavior by comparing velocity, slip ratio, climbability, and selected 6-DOF pose peak values. The proposed model is also compared with a baseline simple-hinge track-chain model.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: Reference-conditioned humanoid locomotion can benefit from future motion information, but the representation of this information is critical under motion transitions. Directly flattening a future reference trajectory gives the actor high-dimensional input and can propagate abrupt reference changes to the policy. This paper proposes a time-structured future motion embedding that combines explicit time tokens, temporal weighting, and a lightweight MLP encoder. Each future reference sample is augmented with normalized time and sampling-interval information, processed by a shared step-wise MLP, reweighted to emphasize near-future samples, and fused into a compact latent vector. The method is evaluated in an Isaac-based RoK-4 humanoid simulation using three direction-varied walking clips randomly cropped from LaFAN1 and retargeted to the robot. Under a three-seed transition-survival protocol, the proposed representation achieves the highest survival rate among current-reference-only, raw-future, no-time-token, and no-weighting baselines. The results suggest that future preview is most effective when represented as a temporally structured compact embedding rather than as a raw trajectory.

Keywords: Robot Mechanism and Control, Robotic Applications, Navigation, Guidance and Control
Abstract: This paper presents the development and system integration of a front double-flipper tracked mobile robot for stair and rough-terrain traversal. The platform was designed for heavy-duty outdoor operation while reducing the mechanical complexity of the flipper mechanism. System requirements were derived with reference to the NIST/ASTM E54.09 response robot test methods, including traversal of 175 mm standard stairs, unstructured terrain, and payload capacity for perception and computation modules. Based on these requirements, the robot was configured with a central body, main tracks, and continuously rotating front flippers for obstacle negotiation and self-righting. The robot integrates a ROS2-based upper computer, a TI controller, CAN communication, motor drivers, and encoder feedback into a unified control architecture. In addition, camera, LiDAR, and IMU sensors were considered for future autonomous navigation and terrain perception. Gazebo simulation and preliminary hardware tests were conducted to verify the feasibility of the proposed platform and its system-level integration.

Keywords: Robotic Applications, Robot Vision, Artificial Intelligence Systems
Abstract: This work targets pick-and-place in extremely cluttered tabletop scenes in which many objects rest in mutual contact, observed by a single wrist-mounted RGB-D camera. Existing graph-based methods describe inter-object inter- actions through purely spatial predicates, which capture where objects sit relative to one another but not which object physically supports which. We propose an interaction-aware graph neural network that, from a single view, jointly infers a per-object structural role (INDEPENDENT, DEPEND, or SUPPORT), an occlusion state, and a target-conditioned stability score that estimates how the remaining structure would respond if a given object were removed. The three heads share a PointNet-based encoder and are trained jointly in simulation. Closed-loop evaluation under matched recovery dispatch shows that the proposed policy attains the highest first-attempt success rate across all evaluated clutter densities and consistently outperforms a ground-truth-contact heuristic and a depth-priority baseline, while also achieving the smallest displacement of the remaining structure during clearing.

Keywords: Human-Robot Interaction, Autonomous Vehicle Systems, Process Control Systems
Abstract: Human-following mobile robots require local navigation strategies capable of simultaneously maintaining collision-free motion, stable following geometry, and tracking continuity. Conventional local planners such as Rapidly-exploring Random Trees (RRT) and the Dynamic Window Approach (DWA) primarily emphasize obstacle avoidance and motion feasibility, often resulting in unstable following behavior or delayed recovery when obstacles interfere with the desired path or temporary target occlusion occurs. This paper proposes a Tracking Continuity-Aware Reactive Planner (TCARP) for human-following mobile robots. TCARP integrates dynamic follow-point generation, condition-aware reactive planning, and predictive continuity recovery within a unified local-planning framework. Experiments conducted using an Autonomous Mobile Robot platform under obstacle-interfered environments demonstrated that TCARP reduced follow-point tracking error by approximately 47–49% and target-centering error by approximately 75–81% compared with DWA while maintaining comparable navigation feasibility. These results indicate that explicitly incorporating tracking continuity into local planning improves human-following robustness and navigation consistency.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Artificial Intelligence Systems
Abstract: This paper proposes a network-aware path planning method for autonomous mobile robots in construction sites equipped with Wi-Fi mesh networks. The method addresses the trade-off between efficient navigation and reliable wireless communication through an integrated framework. First, an empirical relationship between relay hop count and throughput degradation is established. Second, a machine learning based radio propagation model predicts the received signal strength (RSS) heatmap from floor plan information. Wireless communication quality cost map is created based on these two components. Finally, A* path planning is applied to the cost map with configurable weighting parameters, enabling flexible trade-offs between travel distance and wireless communication quality. The proposed method was validated through experiments in a real indoor environment equipped with Wi-Fi mesh network, where an autonomous mobile robot was navigated planned paths while measuring actual throughput. The results demonstrate that the proposed method effectively predicts wireless communication quality and enables network-aware navigation.

Keywords: Biomedical Instruments and Systems, Robotic Applications, Human-Robot Interaction
Abstract: Neurogenic lower urinary tract dysfunction (NLUTD) patients require stable bladder function and effective residual urine management to prevent complications and improve quality of life. Although clean intermittent catheterization (CIC) is widely used, female patients often experience difficulty locating the urethral opening and aligning the catheter due to anatomical constraints, resulting in low adherence and high discontinuation rates. Existing assistive devices provide only limited support, such as securing the field of view or stabilizing the device. To address these limitations, this study proposes a semi-automated self-catheterization assistive device that integrates camera-based urethral recognition and active catheter alignment. The proposed system identifies the urethral position using a vision-based algorithm and employs a two-stage framework in which user-guided tracking is initially applied under limited data conditions, followed by gradual transition to an AI-based automatic detection model using accumulated personalized data. In addition, two pan–tilt motors actively align the catheter tip toward the detected urethral opening through real-time feedback control. A prototype system and urethral dummy model were developed to evaluate the feasibility and quantitative performance of the proposed approach. Experimental results demonstrated the feasibility of the proposed system for supporting independent female self-catheterization.

Keywords: Robotic Applications, Robot Mechanism and Control, Artificial Intelligence Systems
Abstract: Increasing parallelism in reinforcement learning (RL) can destabilize training and reduce sample efficiency, especially at higher update-to-data (UTD) ratios, limiting the scalability of massively parallel, high-throughput RL pipelines. This work investigates why standard RL training becomes unstable in this regime. Our analysis reveals that instability is driven by early-stage value divergence: during the initial phase, high-frequency updates on highly correlated data amplify bootstrapping errors, destabilizing the critic and triggering policy collapse. To counteract this mechanism, we propose Bellman Target Clipping (BeTC), a minimally invasive stabilization technique that clips Bellman targets only during the early phase of training. After this early-phase intervention, training proceeds with the original baseline algorithm. Notably, this technique is simple and avoids potentially harmful algorithmic interventions, yet mitigates a fundamental source of instability in parallel RL throughout the entire training run. Experiments on six continuous-control benchmarks in Brax with 2048 parallel environments show that BeTC significantly stabilizes training at UTD = 1.0, achieving a 2.6× higher normalized return after 1 million environment steps than the non-clipped baseline.

Keywords: Navigation, Guidance and Control, Robot Mechanism and Control, Control Theory and Applications
Abstract: This paper proposes an extended Model Predictive Control (MPC) cost design to induce yielding behavior in autonomous mobile robots navigating narrow passages. To resolve deadlocks without explicit mode-switching, we introduce two specific cost terms: a forward-suppression term that discourages motion toward the goal, and a lateral-evacuation term that promotes movement into available side spaces. By integrating these terms into the continuous optimization process, robots generate complex behavior such as waiting, retreating, and side-evacuation naturally as optimal solutions. Simulation results across five scenarios demonstrate that this approach consistently improves task completion and reduces travel time compared to basic goal-tracking MPC, effectively resolving deadlocks in constrained environments. Additionally, the study identifies performance limitations regarding evacuation space and input constraints, providing insights for practical robot navigation.

Keywords: Robotic Applications, Sensors and Signal Processing, Artificial Intelligence Systems
Abstract: Autonomous drones are used in a wide range of fields, including military, security, and agriculture, and their applications are expanding. An autonomous drone is a drone that performs tasks such as route planning, control, perception, and decision-making without direct human intervention. A power supply is essential for drones to carry out their missions, and the power consumed by drones depends on flight characteristics such as mission conditions and flight patterns. Accordingly, to ensure the stable operation of drones, it must be possible to detect anomalies in the current status based on power consumption, taking into account mission conditions and flight patterns. In this paper, we propose an anomaly detection method for identifying abnormal power consumption patterns. The proposed method utilizes a CNN-based autoencoder to detect power anomalies based on the relationships among input features through a process of compression and reconstruction. Experimental results confirm that the proposed method is capable of reliably detecting anomalies.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Autonomous Vehicle Systems
Abstract: This paper presents a fully data-driven 3-D path-following framework for autonomous underwater vehicles (AUVs), a representative class of underwater field robotics, based on Data-Enabled Predictive Control (DeePC). The approach eliminates explicit hydrodynamic modeling by exploiting measured input-output trajectories to predict and optimize future system behavior. Classic DeePC is employed for heading control, while a cascaded DeePC architecture with loop-frequency separation is proposed for depth regulation, extending DeePC to plants whose dominant output evolves significantly slower than the actuator bandwidth. For 3-D waypoint path following, the Adaptive Line-of-Sight (ALOS) guidance law is extended to a predictive multistep formulation (PALOS) that supplies the horizon-consistent reference required by receding-horizon predictive controllers. All methods are validated in high-fidelity 6 degrees of freedom simulation on the REMUS 100 AUV under nominal operation, ocean-current disturbances, operation beyond the data regime, and 3-D waypoint path following, consistently outperforming the corresponding state-of-the-art benchmarks. In 3-D waypoint path following, the framework reduces cross-track error by approximately 28% relative to the ALOS-PI/PID baseline.

Keywords: Navigation, Guidance and Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: Autonomous aerial monitoring has become important for applications such as infrastructure (a.k.a.'target') inspection and surveillance, where consistent and efficient coverage is critical. However, reactive strategies ensuring collision avoidance in cluttered, dynamic environments may degrade target visibility, thereby resulting in poorer coverage and less effective monitoring. Thus, ensuring motion safety while simultaneously achieving sufficiently high visual coverage of the target remains a major challenge. To address this, a unified framework that amalgamates velocity obstacles (VO)-based avoidance strategy with a coverage recovery mechanism for enhancement of target visibility in 3-D environment is presented in this paper. The novelty of the developed local reactive motion planner, named as '3-D Coverage-Aware Velocity Obstacles' (3-D CAVO), lies in coupling geometric visibility constraints with collision avoidance and introducing a superimposed local sweep strategy on a vertical surface based on the VO-governed avoidance path and resulting in coverage loss. This enables velocity-level geometric collision avoidance and bounded velocity commands under the single-integrator model, while improving projected target-surface coverage. Software-in-the-Loop (SITL) simulation results are presented to illustrate the effectiveness of the developed reactive motion planner in maintaining safety of UAV motion, while enhancing target visibility and coverage in an aerial monitoring mission.

Keywords: Navigation, Guidance and Control, Sensors and Signal Processing, Autonomous Vehicle Systems
Abstract: Accurate localization is essential for autonomous mowers in agricultural environments. However, LiDAR-based scan matching can degrade when grass-height variations cause discrepancies between a prior map and current observations. This study proposes an Error-State Kalman Filter (ESKF)-based localization method that models this effect as a grass-height-induced bias and estimates it online using ground-position observations obtained from the prior map by RANSAC-based ground-plane estimation. The estimated bias is incorporated into the localization process to compensate for systematic localization errors. Experiments in a sloped environment with different grass-height conditions between mapping and localization showed that the proposed method reduced the MAE from 0.16 m to 0.08 m and the RMSE from 0.17 m to 0.09 m, demonstrating the effectiveness of the proposed method without requiring map updates. The results indicate that the approach improves robustness against environmental changes and is suitable for long-term operation in agricultural environments.

Keywords: Artificial Intelligence Systems, Navigation, Guidance and Control, Information and Networking
Abstract: Human-centred ship operation system design requires designers to translate distributed maritime standards, procedures, and human-factors knowledge into task-specific system requirements. However, direct large language model prompting lacks reliable evidence provenance, while conventional retrieval-augmented generation may retrieve topically relevant knowledge without sufficiently representing the operator, task condition, information need, and operational risk. This paper presents a task-driven RAG framework that decomposes a natural-language operation scenario into seven task elements and uses them to support query expansion, metadata-guided retrieval, and source-gated requirement generation. A prototype based on a 55-fragment seed knowledge base was examined through two preliminary cases: bridge alarm interpretation under low visibility and navigation monitoring in high-traffic waters. Direct prompting produced only uncertain suggestions, while Naive RAG generated two and five evidence-linked requirements in the two cases. Task-Driven RAG generated six evidence-linked preliminary requirements in each case. These case-based results suggest that explicit task modelling can strengthen the alignment between operation scenarios, retrieved evidence, and preliminary design requirements. Further work will expand the maritime knowledge base and conduct expert evaluation across more representative operation scenarios.

Keywords: Sensors and Signal Processing, Control Theory and Applications, Information and Networking
Abstract: In recent years, advances in technologies such as security cameras and drones have prompted growing demand for multi-view systems that process multiple images. Thereby, we develop a distributed Robust Principal Component Analysis (RPCA) method for object detection using images obtained from multiple images. RPCA is an algorithm that separates multiple images captured at different times into background and moving objects. We propose a multi-agent system that performs object detection by separating the background from the foreground using RPCA. However, the optimization problem in RPCA is generally solved through centralized updates. This approach raises issues such as communication bandwidth constraints and　increased computational resource requirements, and it is unsuitable for application in a multi-agent system. Therefore, we formulate RPCA as a distributed optimization problem at the agent level. We then derive a distributed algorithm by solving　this problem and verify the effectiveness of these algorithms through simulation. As a result, we have successfully developed a distributed system capable of object detection as a multi-agent system.

Keywords: Human-Robot Interaction
Abstract: This study presents a kinesthetic haptic interface capable of independently constraining 3-DOF finger flexion using miniaturized electrostatic clutches. Conventional interfaces suffer from coupled joint control, where a single actuator constrains multiple joint angles, preventing complex interactions. We overcome this limitation by proposing a clutch-on-a-joint paradigm, placing modules directly on each finger joint. Key innovations enable this approach: (1) module miniaturization achieved via curvature-induced rigidity in the electrode film; (2) skin anchoring challenges resolved through a kinematically modeled rolling joint mechanism; and (3) a passive variable-length structure resolving kinematic overconstraint to restore range of motion. To validate the architecture, we conducted a blind user study using bending sensors. Results confirm the rolling joint successfully suppresses kinematic lash within the human proprioceptive limit (5°–9°). Participants accurately perceived independent joint constraints with up to 76% accuracy. These findings demonstrate this clutch-on-a-joint paradigm effectively decouples multi-joint constraints, delivering stable 3-DOF kinesthetic feedback.

Keywords: Robot Vision, Artificial Intelligence Systems, Robotic Applications
Abstract: Vision-Language-Action (VLA) models can suffer substantial performance degradation when task-irrelevant visual contexts change at test time. This paper proposes CARVE, a training-free inference-time attention reweighting method for frozen VLA policies. CARVE identifies task-relevant regions using language-conditioned source/goal grounding and temporal gripper-centered cues, while preserving all visual tokens and suppressing persistent task-irrelevant attention sinks through soft attention reweighting. CARVE consistently improves performance across three representative VLA backbones in SIMPLER simulation and also demonstrates improved performance in real-world manipulation environments.

Keywords: Robotic Applications, Robot Vision, Industrial Applications of Control
Abstract: Cellulose filament production based on Interfacial Polyelectrolyte Complexation (IPC) requires stable liquid handling and wet-drawing operations, where manual execution often introduces variability that affects filament quality. This paper presents a robotic automation system for cellulose filament production using a standardized gripper interface and vision-based labware detection. The proposed system integrates a 6-DOF robot arm, a depth camera, and a customized parallel gripper to perform automated laboratory procedures including pipetting, droplet handling, and filament drawing. A standardized gripper interface is designed to enable stable manipulation of multiple laboratory tools without modifying the gripper structure. To support autonomous operation, a vision-based framework is developed for labware localization through point cloud processing and OBB estimation, while lightweight visual recognition methods are employed for pipette tip recognition and droplet detection. Experimental evaluations are conducted through robotic pipetting and IPC filament generation. The system achieves reliable autonomous pipetting, with aspiration and dispensing times comparable to those of a human operator and a commercial pipetting device, although additional time is required for robotic tip attachment. In IPC experiments, the robotic system generates longer filaments with smaller mean diameters, higher mean Young’s modulus, and improved diameter uniformity compared with manual operation, demonstrating the potential of robotic execution to improve the repeatability and consistency of IPC-based cellulose filament production.

Keywords: Robot Vision, Artificial Intelligence Systems, Human-Robot Interaction
Abstract: This paper addresses unknown-aware few-shot novel place mapping for semantic and topological robot maps. Existing place mapping systems often assume that every place belongs to a predefined closed set of categories or can be retrieved from a pretrained open-vocabulary embedding space. This assumption is insufficient for real-world robot deployments, where environment-specific concepts such as a robot charging station or a user's study space may be absent from the current map. We formulate a mapping problem in which a robot detects places outside the known class set, keeps ambiguous observations unresolved when necessary, and registers learnable novel place classes from only a few support observations. The proposed formulation connects unknown detection and few-shot adaptation with semantic/topological mapping, enabling maps to represent base-known, novel-known, and unknown places in a unified structure.

Keywords: Robotic Applications, Autonomous Vehicle Systems, Navigation, Guidance and Control
Abstract: This work proposes a local environmental representation for safe mobile robot navigation based on a continuous Gaussian Process (GP) approximation of a discrete Signed Distance Field (SDF). The representation is constructed from local sensor measurements and provides a smooth, differentiable approximation of the surrounding environment. The proposed representation is evaluated in terms of computational performance, approximation accuracy, and navigation safety. In simulation, the Gaussian Process Signed Distance Field (GP-SDF) is embedded as a Control Barrier Function (CBF) into a Model Predictive Controller (MPC) and compared against the Dynamic Window Approach (DWA) and Model Predictive Path Integral (MPPI) controller, both of which rely on local information for obstacle avoidance. The results show that the proposed controller achieves lower failure rates, reducing failures from 43% with DWA and 25% with MPPI to 12% with the proposed solution in the most challenging scenario with 10 dynamic obstacles. The evaluation is conducted in a simulated service robot navigation scenario with static obstacles and moving pedestrians. The code repository is available at https://gitlab.ub.uni-bielefeld.de/dibami/gp-sdf-cbf_mpc .

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Robot Vision
Abstract: Detecting defects in manufactured products is difficult because defects vary widely in scale and appearance: a fine scratch disturbs only the local surface texture, a misaligned component alters the structural pattern, and a discolouration shifts the overall colour uniformly across a region.

We present PRISM-AD (Precision-weighted Reconstruction with Illumination-conditioned Spectral Mixture), a reconstruction-based anomaly detector that addresses this variation by training three expert decoders to specialise on different frequency bands of the feature representation. Each expert learns to reconstruct normal patterns within its designated band; regions where reconstruction fails indicate anomalies.

A router conditioned on scene illumination and frequency energy weights the experts, and a precision-weighted uncertainty fusion step down-weights locations where an expert's reconstruction is unreliable, suppressing false positives on chronically difficult texture; we measure the trained router and find its weights stay close to uniform, so the uncertainty term carries the fusion in practice.

A frozen DINOv3 Vision Transformer provides the feature backbone; only normal images are required for training. On the full Real-IAD multi-view benchmark covering 30 product categories, PRISM-AD achieves S-AUROC 93.4, I-AUROC 90.8, and P-AUPRO 93.9, the strongest scores on all three metrics among the reconstruction-based baselines tabulated in the Real-IAD benchmark paper.

A second stage uses a Vision Language Model to explain detected anomalies in structured natural language, so an inspector is shown where the flagged region is and what it looks like without any fine-tuning.

Keywords: Control Theory and Applications, Navigation, Guidance and Control, Robot Mechanism and Control
Abstract: This paper proposes a learning method to design a neural-network-based nonlinear H-infinity control law for a two-dimensional quadcopter. In the proposed method, training data for the nonlinear H-infinity control law are generated by executing backward time integration of Hamilton's canonical equations associated with the Hamilton-Jacobi-Isaacs (HJI) equation. In this process, the setting of initial states for the backward integration is critical to ensuring that the generated data are uniformly distributed across the state space. To address this challenge, we introduce and compare two terminal data generation strategies: level-surface sampling (Strategy A) and linearized-system prediction (Strategy B). By strategically deploying the terminal states along the system's stable modes, Strategy B effectively suppresses immediate numerical divergence and significantly extends the backward trajectory durations. This extended duration facilitates the learning of a more comprehensive and robust control law. Flight simulations under wind velocity disturbances demonstrate that the neural network controller trained via Strategy B achieves the lowest L2 norm ratio compared to both Strategy A and a conventional linear H-infinity control. These results confirm that the proposed method successfully captures a smooth, robust control law that enables extended horizontal translations without triggering abrupt attitude variations of the vehicle.

Keywords: Control Theory and Applications, Robot Mechanism and Control, Robotic Applications
Abstract: Sampling-based model predictive control such as Model Predictive Path Integral (MPPI) provides derivative-free optimization and robustness for complex robotic systems, but cannot guarantee hard-constraint satisfaction since constraints are handled as soft penalties. This paper proposes Manifold-Constrained MPPI (MC-MPPI), a real-time framework that strictly enforces manifold-based equality constraints by decoupling constraint handling into two stages. At the planning stage, a Variational Autoencoder (VAE) learns a low-dimensional latent representation of the constraint manifold so that MPPI can sample near-feasible trajectories without per-sample projection. At the execution stage, a single-step QP projection corrects residual constraint violations on the selected solution before tracking by an inverse-dynamics controller. Closed-chain dual-arm manipulation experiments show that MC-MPPI reduces equality-constraint violation by an order of magnitude over baseline MPPI variants. Detailed experiment results including videos, additional ablations, and real-world deployment are available at https://rcilab.github.io/mcmppi.

Keywords: Robotic Applications, Artificial Intelligence Systems, Human-Robot Interaction
Abstract: Bimanual object handover is a fundamental capability for collaborative robotic manipulation, particularly in precision-demanding surgical subtasks. This paper investigates the application of Context-Based Echo State Networks (CESNs) as a computationally efficient Learning from Demonstration (LfD) mechanism for autonomous bimanual needle handover. We develop a dual-arm simulation environment in NVIDIA Isaac Sim featuring two Franka Panda manipulators equipped with patient-side manipulator (PSM) tools. Demonstration trajectories are generated programmatically utilizing the cuMotion planner across four varying spatial needle initializations. Addressing high-dimensional control, the CESN architecture receives external task contexts and robot joint feedback to directly predict continuous 11-degree-of-freedom (DoF) joint-position commands for closed-loop execution. We evaluate the model's performance through offline joint trajectory reconstruction and online simulation rollouts. The results indicate that CESNs can accurately reproduce context-dependent bimanual trajectories with low computational overhead, achieving millimeter-level precision and offering a lightweight alternative to high-capacity deep learning models for encoding structured manipulation subtasks.

Keywords: Robotic Applications, Artificial Intelligence Systems, Robot Mechanism and Control
Abstract: Routing in a Robotic Mobile Fulfillment System (RMFS) is normally computed by one central planner. A perrobot world model could remove that dependency; this paper evaluates one learning robot among planner-driven peers. The model’s weakest channel is the position of other robots: on a four-camera 360◦ model the predicted neighbour position is 6.2 m off after 2.0 s in a 16 m wide warehouse, and storage pods hide 47.9% of nearby robots from the cameras. This paper adds an Integrated Sensing and Communication (ISAC) channel: each neighbour slot gains five noisy, range-gated sensing channels and two communicated destination channels with packet loss. Two world models are trained on one dataset, encoder and latent set, differing only in those seven channels, and scored by the rank correlation ρ between dream and simulator rankings of 20 candidate action sequences. At 5 robots the effect is null (ρ 0.723 vs 0.681, 95% confidence interval (CI) on the difference [−0.023, +0.106], n=16). At 16 robots the ISAC arm reaches ρ=0.602 against 0.449 (∆=+0.152, CI [+0.028, +0.277], permutation p=0.025, n=10), a pattern consistent with congestion making neighbour information decision-relevant. A controller trained in the camera-only dream runs in RAWSim-O with throughput inside the range of six centralized planners but follows the A* direction on every hop; neither dream reaches the ρ ≥ 0.90 gate needed to depart from A*.

Keywords: Robotic Applications, Autonomous Vehicle Systems
Abstract: Counter Unmanned Aerial Systems (CUAS) surveillance-evasion studies commonly assume an adversarial UAS with perfect knowledge of the defender's sensor network, an assumption unlikely to hold after imperfect reconnaissance. This paper investigates how uncertainty in sensor position, pan speed, and initial scan phase affects attacker evasion performance and which information components the defender should prioritize denying or degrading. We extend a bilevel surveillance-evasion framework by separating the attacker's believed defender state from the true state, allowing the attacker to optimize against a perturbed belief while its trajectory is evaluated against the true sensor configuration. A Monte Carlo study over 100 randomized maps, five uncertainty levels, and five trials per map compares the three information components under matched normalized degradation. Positional and pan-speed uncertainty produce comparable increases in true detection cost, with no statistically significant difference between them, while both have significantly larger effects than scan-phase uncertainty. These results suggest that defender information-denial efforts should prioritize the attacker's estimates of sensor position and pan speed.

Keywords: Autonomous Vehicle Systems, Robot Mechanism and Control, Sensors and Signal Processing
Abstract: Unmanned Surface Vessels (USVs) require trajectory generation that keeps a safe distance from surrounding traffic while remaining robust and efficient, yet conventional model-based planners generalize poorly and single-paradigm learning methods suffer from covariate shift or lowsample efficiency. This paper proposes iTransformer-GRPO, a two-stage learning framework for USV trajectory generation that combines imitation learning with deep reinforcement learning. An iTransformer backbone captures the long-term temporal dependencies of multivariate navigational states and outputs a three-dimensional action consisting of one immediate waypoint displacement and a normalized curvature, so that the learning signal is aligned with the curvature-based steering used at deployment. Stage 1 distills demonstration trajectories into the policy by behavioral cloning, and Stage 2 refines it in closed loop with value-function-free Group Relative Policy Optimization (GRPO), in which the policy is re-applied to the state its own previous action produced and a group is a set of such trajectories sampled from one navigational state, so that the advantage follows from group standardization alone. The safety term of the reward is built on a Fujii-type ship-domain geometry, which keeps the safety requirement well defined for an arbitrary number of targets, while the ncounter classification of Rules 13–15 and 17 of the International Regulations for Preventing Collisions at Sea (COLREGs) is used to label encounter types. The scope of this paper is the formulation and design of the framework; a quantitative closed-loop evaluation of the trained policies is left to a dedicated study.

Keywords: Artificial Intelligence Systems, Multimedia Systems, Robot Vision
Abstract: Plant diseases threaten agricultural productivity and food security, especially in developing regions with limited access to plant pathology expertise. Deep learning–based disease detection offers a scalable solution for precision agriculture and agricultural robots, enabling real-time monitoring while reducing reliance on human labor. This study benchmarks four backbones—ResNet-50, EfficientNet-B0, ViT-Base/16 (AugReg), and ReXNet-150—using a controlled two-stage transfer learning protocol with adaptive augmentation, including the proposed Random Shadow technique, and weighted random sampling for class imbalance. Two training regimes were evaluated: (A) standard cross-entropy and (B) MixUp/CutMix with label smoothing (ε=0.1). Experiments used a 65-class dataset combining twelve public collections (115,059 images across 20 crop species), with results reported as mean ± standard deviation over three seeds and paired significance testing. MixUp/CutMix + label smoothing did not improve ReXNet-150 (95.01 ± 0.15% vs. 95.37 ± 0.16% Top-1; Δ = −0.36 ± 0.30, p = 0.175). Ablation analysis identified CutMix as detrimental; removing it significantly improved balanced accuracy (95.36 ± 0.21% vs. 94.78 ± 0.13%; Δ = +0.58 ± 0.08, Holm-adjusted p = 0.018). The findings support practical deployment in mobile crop monitoring and robotic disease scouting while highlighting the need for field-image validation.

Keywords: Artificial Intelligence Systems, Multimedia Systems, Sensors and Signal Processing
Abstract: Industrial assembly operations require continuous traceability through structured reporting. However, manual data entry remains challenging in practice, as operators work with protective equipment, both hands engaged, and under strict safety and productivity constraints. In such conditions, interacting with keyboards or touchscreens is inconvenient, leading to delayed, incomplete, or inaccurate reporting. Voice input provides a natural alternative, but spoken descriptions alone are often noisy, ambiguous, and insufficient for reliable form completion. This paper proposes a multimodal approach for voice-based hazard reporting and missing-object identification, formulated as a semantic form-filling problem. The pipeline combines speech enhancement and automatic speech recognition using Whisper with context-aware semantic interpretation and visual reasoning based on a unified Vision-Language Model (Qwen3.6-27B). The model integrates transcription, assembly context, and optional visual input to generate structured, schema-compliant reports and assess component presence. It provides pre-filled forms subject to final human validation. Under simulated industrial conditions, the proposed pipeline achieved a 84.5% F1-score for semantic extraction and 83.4% accuracy for missing object detection. The results demonstrate the feasibility of multimodal AI assistance for reducing operator input effort while maintaining safety, reliability and traceability in industrial workflows.

Keywords: Artificial Intelligence Systems, Navigation, Guidance and Control
Abstract: This study proposes an AR-AI verification framework that enables physical AI systems to be verified in an AR environment. In the proposed framework, an LLM recognizes the positions and poses of target objects and a robot from AR images. The use of AR technology enables operation verification in a real environment without using an actual robot. In this study, a coordinate estimation method that draws grid lines on AR images to assist the spatial reasoning of the LLM is proposed. The effects of the presence or absence of grid lines, grid-line colors, capture conditions, target objects, and capture environments on the coordinate estimation accuracy of the LLM were evaluated. The experimental results showed that the relative error in straight-line distance was reduced from 56.7% to 8.3% by displaying grid lines. In addition, the difference in the relative error of straight-line distance due to the grid-line color was small. On the other hand, under capture conditions in which the grid lines appeared slanted, the relative error increased to a maximum of 92.8%. Furthermore, even for different target objects and capture environments, a tendency was observed for the relative error of straight-line distance to be smaller when grid lines were displayed.

Keywords: Artificial Intelligence Systems, Navigation, Guidance and Control, Robotic Applications
Abstract: Large language models (LLMs) are increasingly used as natural-language interfaces for robotic systems, yet their integration with Robot Operating System (ROS)-based navigation remains limited by two gaps. First, navigation data such as occupancy grids are represented as raw geometric messages that are difficult for LLMs to use directly as spatial or semantic context. Second, adding LLM-driven capabilities often requires custom wrappers or robot-specific interfaces, limiting reuse across systems. To address these challenges, we propose a non-invasive framework that connects LLM reasoning with ROS-based navigation through a navigation-oriented representation layer, exposed through the Model Context Protocol (MCP) as standardized, reusable tools so that any MCP-compatible LLM can access them without robot-specific wrappers. The visual map modules transform occupancy grids into metric, pose-aware images for goal reasoning, while the semantic annotation modules record waypoint-level observations with robot poses. We evaluate the framework on three tasks: autonomous mapping, spatial reasoning-based navigation, and semantic reasoning-based navigation. The results show that the evaluated LLM backends use these representations to achieve over 97% map coverage and select spatial or semantic navigation targets from natural-language instructions in a simulated indoor environment. This demonstrates representation-mediated LLM navigation without modifying the existing ROS navigation stack.

Keywords: Autonomous Vehicle Systems, Artificial Intelligence Systems, Control Theory and Applications
Abstract: Rescue robots equipped with multiple joints and sensors impose a significant operational burden on human operators, particularly in rough terrain. This study proposes a reinforcement learning-based flipper control method for a tracked rescue robot using the Proximal Policy Optimization (PPO) algorithm. The controller was trained in Isaac Sim and Isaac Lab using only a 45° staircase environment. A reward function was designed to encourage upward movement while suppressing abrupt flipper motions and excessive vertical velocity, enabling stable and smooth locomotion. To evaluate the generalization capability of the learned policy, stair-climbing experiments were conducted on staircases with inclination angles of 15°, 30°, 45°, and 60°, including conditions not used during training. Experimental results showed success rates of 90%, 100%, 100%, and 90% for reaching a height of 1.0 m, and 0%, 95%, 95%, and 85% for climbing 2.0 m, respectively. These results demonstrate that the proposed method successfully learned an effective flipper control policy and showed promising generalization capability to unseen staircase inclinations.

Keywords: Artificial Intelligence Systems, Rehabilitation Robot, Exoskeleton Robot
Abstract: Generating physically plausible human gait with explicit control over biomechanical parameters is motivated by applications such as clinical gait analysis, rehabilitation, wearable robot control, and biomechanical simulation. Existing human motion generation models are typically conditioned on text descriptions or action labels, which provide limited control over quantitative gait variables such as walking speed and cadence. They also commonly generate only kinematics, without ground reaction forces (GRFs) needed for biomechanical analysis. In this work, we propose PhyGait, a diffusion-based framework that generates full-body human gait motion and six-channel GRFs conditioned on target speed and cadence. The framework uses a transformer-based denoising network with continuous gait conditioning, adaptive layer normalization, joint motion-GRF generation, and gait-specific physical consistency losses. Trained on 8,953 laboratory motion-capture gait clips collected from two treadmill datasets, PhyGait was evaluated across 13,800 generated samples spanning 115 speed-cadence conditions over speeds of 0.75-1.7 m/s and cadences of 1.5-2.1 steps/s. It accurately followed the target gait conditions, with speed and cadence mean absolute errors of 0.005 m/s and 0.049 steps/s, respectively, while generating realistic motions, stable foot-ground contacts, and biomechanically plausible GRF profiles. These results demonstrate the potential of diffusion models for generating controllable and biomechanically plausible synthetic gait data.

Keywords: Sensors and Signal Processing, Navigation, Guidance and Control, Robot Vision
Abstract: This paper proposes a posterior existence-guided selective temporal accumulation framework for sparse LiDAR-based drone tracking. Conventional LiDAR clustering-based and Deep Learning (DL)-based tracking methods remain vulnerable to track loss in long-range sparse-observation environments. To address this problem, the proposed method transforms raw LiDAR points into a pseudo-likelihood map to represent weak target evidence. A distance-dependent empty-observation likelihood is introduced to update the Bernoulli posterior even when no return is observed. In addition, posterior-guided selective temporal accumulation is used to accumulate only high-confidence past observations. Experimental results demonstrate that the proposed method improves recall, reduces tracking position error, and decreases track loss compared with conventional clustering-based and DL-based tracking methods. These results show that the proposed framework provides a robust tracking solution under sparse and intermittent LiDAR observations.

Keywords: Robotic Applications, Navigation, Guidance and Control, Process Control Systems
Abstract: Indoor service robots are increasingly expected to find and reach objects described by their users in natural language. Existing map based navigation methods often rely on compact semantic maps that retain only a fixed set of categories and discard the texture needed to tell similar objects apart, while recent Gaussian Splatting navigation can ground an image goal but cannot accept open vocabulary language. We present a framework that builds a language embedded Gaussian map of an indoor scene, in which each object instance is associated with a long text vision language feature so that the scene can be queried by free form text or images. Given a user prompt, the target object is extracted and matched against the map to retrieve candidate instances. Because language goals are ambiguous, we render views of the candidates from the map and use a vision language model to verify which one satisfies the user requirement before the robot navigates to it. Experiments against representative baselines show that our pipeline supports a range of language driven navigation tasks for indoor service robots.

Keywords: Robot Vision, Artificial Intelligence Systems, Robotic Applications
Abstract: 3D Gaussian Splatting (3DGS) is a powerful scene representation, but extending it to instance-level and open-vocabulary understanding remains difficult. Point-level methods learn low-dimensional instance features on Gaussians and then discretize them, but they rely on multi-level k-means codebooks with predefined cluster counts that cannot adapt to scene complexity, and they attach language features by averaging CLIP embeddings over every view a cluster appears in, letting poorly visible or mis-associated views corrupt the result. We address both steps in post-processing. First, we discretize the features with a single density-based hierarchical clustering pass over a balanced fusion of 3D position and instance features, so a cluster forms only where points are spatially close and feature-similar and the instance count emerges from the scene rather than a fixed k. Second, we add a robust language association step that rejects unreliable views and averages only those agreeing with the cluster consensus. Experiments on indoor scenes show consistent improvements in semantic segmentation and instance clustering over the Gaussian-based baseline, and ablations confirm each component's contribution.

Keywords: Control Theory and Applications, Industrial Applications of Control, Autonomous Vehicle Systems
Abstract: Performance optimization for quadrotor control requires different control efforts during transient and steady-state conditions. Satisfying these requirements is difficult with a single Model Predictive Control (MPC). This paper proposes a switching-based hybrid control system for quadrotor position control using Koopman MPC (KMPC) and Nonlinear MPC (NMPC). The proposed method includes a selector that evaluates the switching cost of each controller and chooses the input from the controller with the lower cost. By applying this selector, the method exploits advantages of each MPC and achieves effective position control. The proposed method was verified on Quanser 3 degrees of freedom (DOF) Hover platform and in MATLAB/Simulink environment. We also conducted simulations to evaluate the proposed method, and analyzed the characteristics of each controller in terms of tracking performance and input consumption. Experimental result shows that NMPC was selected more frequently, while KMPC tended to be selected during transient intervals. As a result, the proposed method can select the appropriate controller depending on the flight conditions.

Keywords: Control Theory and Applications, Navigation, Guidance and Control, Robot Mechanism and Control
Abstract: Tracked mobile robots (TMRs) are advantageous for rough-terrain navigation because they provide high traction and stable mobility. However, skid-steering locomotion induces model uncertainty through track--terrain interaction, thereby degrading path-tracking accuracy. This paper proposes a slip-aware path-tracking framework for TMRs based on a disturbance-augmented kinematic model and a DOB-EKF-based estimator. The robot motion is modeled using persistent slip parameters and transient additive disturbances. The slip parameters represent terrain-induced response degradation, whereas the additive disturbances represent short-duration effects such as impulsive impacts. To estimate these quantities, a dual-layer estimator is designed. A disturbance observer (DOB) estimates fast additive disturbances, while an extended Kalman filter (EKF) estimates persistent slip parameters. A persistence detector determines whether the model residual should be retained as a transient disturbance or reflected in the EKF-based slip parameter estimation. The estimated parameters are then used to compensate the linear and angular velocity commands in the path-tracking controller. The proposed method is evaluated through simulations including terrain transitions and impulsive disturbances.

Keywords: Autonomous Vehicle Systems, Artificial Intelligence Systems, Control Theory and Applications
Abstract: This paper presents the design and field validation of an artificial intelligence (AI)-assisted autonomous navigation system for a low-speed electric golf cart. The proposed system integrates a voice AI agent for destination and stop-command input, LiDAR-Inertial Odometry via Smoothing and Mapping (LIO-SAM)-based 3D mapping, OpenMP-accelerated Normal Distributions Transform (NDT_OMP) scan-matching localization, and adaptive Pure Pursuit path-following control. A light detection and ranging (LiDAR)-inertial map is generated offline from a manual teaching run, while online scan-to-map matching provides the vehicle pose in the map frame for path tracking. Field experiments were conducted on a campus route using an instrumented electric golf cart. The voice AI agent achieved intent-recognition accuracies of 98%, 78%, and 90% under indoor, outdoor, and moving-vehicle conditions, respectively. The localization module achieved a mean planar error of 0.514m and a root-mean-square error (RMSE) of 0.614m relative to a real-time kinematic Global Navigation Satellite System (RTK-GNSS) reference trajectory. In three adaptive Pure Pursuit runs, the controller achieved an overall mean absolute cross-track error (CTE) of 0.207m and a root-mean-square CTE of 0.261m over 1626.6m of autonomous path-following operation. These results demonstrate that the integrated pipeline can support practical low-speed autonomous navigation in a campus environment while keeping high-level AI interaction separate from deterministic localization and control.

Keywords: Autonomous Vehicle Systems, Artificial Intelligence Systems, Control Theory and Applications
Abstract: Learning-based controllers offer data-driven adaptivity for autonomous systems operating under uncertain and time-variant conditions, but their deployment in safety-critical control applications lacks formal guarantees for stability and robustness. Training data cannot cover all operating conditions, and learned models may degrade silently through concept drift. This paper proposes a runtime assurance reference architecture that places a safety shell around the learning-based controller. The shell combines a novelty detector on the input space of the learning-based controller, a drift detector on the input-output behavior of the system, a fallback controller restricted to analyzable model classes, a supervisor coordinating mode transitions, and a cloud-based learning module that keeps batch model updates out of the safety-critical control path. Because the switching evidence is statistical, the shell reduces risk instead of proving safety, and it concentrates the need for formal evidence in the fallback path and the surrounding safety case. A sensitivity study using a multibody model of an autonomous rail vehicle shows that none of the three representative PI tunings is simultaneously safe and efficient across 27 operating points, motivating adaptive control combined with runtime safety assurance. The shell is specified at the level of functional roles and interfaces and has not yet been instantiated or validated in closed loop.

Keywords: Autonomous Vehicle Systems, Artificial Intelligence Systems, Control Theory and Applications
Abstract: Driving style adaptation has attracted increasing attention as a means of improving user acceptance and comfort in autonomous vehicles. This paper proposes a progression-aware residual learning framework for driving style adaptation while preserving an existing baseline path-tracking controller. The proposed method learns the residual steering command between a baseline controller and a target driving style controller using vehicle states, a normalized lane-change progression variable τ, and baseline steering commands. To generate diverse driving styles without requiring real driver data, lane-change style profiles with different lane-change durations and steering characteristics are designed using model predictive control (MPC). For each driving style, an individual residual learning model is trained and integrated with the baseline controller in a closed-loop CarSim–Simulink environment. The variable τ encodes maneuver-phase information, enabling the network to distinguish different stages of a lane-change event and reproduce style-dependent steering behaviors throughout the maneuver. Experimental results demonstrate that the proposed framework successfully reproduces the lane-change characteristics of target driving styles while preserving the stability and structure of the baseline controller. Furthermore, the trained models show satisfactory generalization to unseen double lane-change scenarios without additional retraining, indicating that the proposed framework captures style-dependent steering patterns rather than memorizing a specific lane-change trajectory.

Keywords: Autonomous Vehicle Systems, Control Theory and Applications
Abstract: Modal analysis helps engineers understand complex dynamical systems by representing their responses as combinations of simpler modal components. Vehicle-trailer (VT) lateral dynamics exhibit strongly coupled lateral motion between the towing vehicle and trailer, while retaining modal characteristics associated with each subsystem. Therefore, a modal viewpoint can provide useful insight for VT system evaluation, design optimization, stability assessment, and controller design. However, VT systems have rarely been investigated from a systematic modal-analysis perspective. This paper proposes an anchor-based modal family classification framework for the lateral dynamics of VT systems and vehicle-trailer-driver (VTD) systems, where the VTD system represents a driver-in-the-loop closed-loop extension of the VT system. For the VT system, vehicle- and trailer-origin modes are defined using a near-zero-trailer-mass anchor and continued to the nominal trailer mass. For the VTD system, vehicle-, trailer-, and driver-origin modes are classified using a near-zero-driver-feedback anchor and continued to the target driver feedback condition. A continuation tracking algorithm combining pole proximity and the modal assurance criterion is proposed to maintain modal-family labels across operating speeds. Simulation results show the speed-dependent evolution of VT and VTD modal families, demonstrate the consistency of the proposed modal classification method, and support the physical interpretation of the classified modes through modal contribution analysis.

Keywords: Autonomous Vehicle Systems, Control Theory and Applications, Artificial Intelligence Systems
Abstract: This paper proposes an online friction-adaptive reinforcement learning (RL) framework for autonomous racing under spatially varying tire-road friction. Such spatial friction variation hinders reliable model-based control, whereas RL can learn racing policies through direct interaction with the racing environment. However, conventional RL policies neither explicitly account for the current friction condition nor inherently prevent actions from exceeding the friction limit. To address these challenges, we develop an action-mapping RL framework that uses online estimated friction information both to condition the policy and to constrain action execution. Specifically, a nonlinear least-squares estimator identifies the tire-road friction coefficient from recent state-action transitions. The estimated friction is then included in the input of the policy trained by proximal policy optimization (PPO). The estimated friction is further used to define , whereby action mapping (AM) mechanism maps each policy output to a physically feasible action. Simulation results under spatially varying friction demonstrate that the proposed framework reduces mean lap time by up to 31.42% of baseline, confirming improved friction-adaptive racing performance.

Keywords: Autonomous Vehicle Systems, Control Theory and Applications, Information and Networking
Abstract: In vehicle platooning, wireless information exchange allows vehicles to coordinate and follow each other closely, reducing the aerodynamic drag and increasing fuel savings. However, wireless transmissions are subject to failures due to packet losses and when such failures happen at the time of a sudden brake, they may cause safety issues. The safety performance under packet losses can be improved by choosing the vehicle control parameters effectively. In particular, the minimum inter-vehicle distance during a sudden brake can be made large enough to avoid crashes. On the other hand, such control parameters typically increase aerodynamic drag on vehicles and in turn reduce fuel savings. In this paper, we formalize the control parameter search as constrained multi-objective optimization problems. Under different wireless channel settings, we use NSGA-II with repeated simulation runs to search for control parameters that perform well in terms of both safety and fuel efficiency. For a vehicle platoon that contains 10 vehicles, we demonstrate the approximate Pareto front with solutions showing trade-off relationship between safety and fuel efficiency.

Keywords: Robot Mechanism and Control, Biomedical Instruments and Systems, Robotic Applications
Abstract: This paper presents a needle insertion robotic end-effector for percutaneous epidural neuroplasty. The system employs a parallelogram-based remote center of motion (RCM) mechanism to maintain a fixed skin entry point while enabling two degrees of freedom (DOF) for needle orientation. Needle insertion and axial rotation are independently actuated via a syringe-based hydraulic transmission, providing a total of 4-DOF. A prototype was developed to validate the kinematic design. Experimental results confirmed independent 4-DOF operation without mutual interference and demonstrated angular positioning accuracy within ±1° over a ±30° range. However, limitations of the hydraulic transmission were identified, including susceptibility to nozzle clogging, transmission variability, and maintenance complexity. To address these issues while preserving the RCM kinematics, we propose a belt–pulley-based transmission mechanism. The redesigned system aims to improve mechanical stiffness, reduce response delays, and enhance repeatability. Future work will focus on quantitative performance comparison and precision control validation for clinical applicability.

Keywords: Biomedical Instruments and Systems, Robotic Applications
Abstract: This study presents a handheld SMART microsurgical system designed to improve precision and safety in vitreoretinal surgery. Ophthalmic microsurgery requires micrometer-scale manipulation of delicate retinal tissues, but physiological hand tremor and unintended instrument drift can limit surgical accuracy and increase the risk of tissue damage. To address these challenges, the proposed system integrates a common-path swept-source optical coherence tomography distance sensor with a compact PZT actuator. The OCT sensor measures the real-time distance between the surgical tool tip and the target tissue, while the actuator compensates for unwanted motion at the distal end of the instrument. This approach preserves the intuitive operation of handheld surgical tools while providing active tremor cancellation. The study also introduces a bimanual SMART platform that enables one stabilized tool to grasp and tension tissue while another performs precise dissection or peeling. Recent developments, including OCT-based predictive stabilization, magnetic sensing, multimodal optical imaging, and multi-degree-of-freedom control, further demonstrate the potential of SMART systems as clinically practical sensor-fusion platforms for future robot-assisted microsurgery.

Keywords: Robot Vision, Robotic Applications
Abstract: Fluoroscopy-guided spinal interventions require accurate alignment between preoperative CT images and intraoperative X-ray images. However, this alignment is challenging because the lumbar spine may not behave as a single rigid structure during patient positioning. Individual vertebrae can undergo local rotation and translation, which may reduce the accuracy of conventional rigid 2D/3D registration methods. This study proposes a learning-based framework for estimating the local six-degree-of-freedom pose of individual lumbar vertebrae from multi-vertebra digitally reconstructed radiographs. Isolated CT volumes of L1-L5 were used to generate synthetic DRR images with known per-vertebra pose variations and a global patient-pose. A ResNet-34-based regression network was trained to predict 30 local pose parameters corresponding to the five lumbar vertebrae. The method was evaluated on 100 test DRR images, representing 500 individual vertebra predictions. The model achieved an overall mean rotation L2 error of 2.5898^circ and a mean translation L2 error of 2.1400 mm. Per-vertebra analysis showed better performance for the middle lumbar levels compared with boundary vertebrae. These results suggest that modeling the lumbar spine as multiple locally moving vertebrae is a promising approach for improving patient-specific 2D/3D spine registration.

Keywords: Navigation, Guidance and Control
Abstract: Pain treatment typically relies on fluoroscopic imaging or computed tomography (CT), which has disadvantages such as requiring two-dimensional image interpretation, inducing repetitive radiation exposure, and lacking soft tissue visualization. This study implements a 3D reconstruction and multi-mode CT/MRI fusion pipeline and proposes an augmented reality (AR) navigation architecture integrated with a robot-assisted needle guidance function. The developed pipeline fuses CT (anatomical bone structures) and MRI (neuroroots, intervertebral discs, dural sacs) images using deformable registration, segments key structures through a deep learning workflow, and generates a simplified 3D surface model optimized for real-time AR rendering. To address occlusion in crowded treatment areas, a robot needle tracking framework was designed based on robot kinematics and reference point registration attached to the skin. The planned AR interface provides real-time trajectory guidance, distance feedback between the needle tip and the target point, and proximity warning functions to a head-mounted display. Furthermore, a step-by-step validation protocol was developed, including lumbar phantom accuracy, mannequin/cadaver usability, and pilot randomized clinical trials. This detailed summary describes the patient-specific multimodal modeling process and, in particular, details the overall system architecture and preclinical validation pathways. Physical hardware integration and clinical trials remain as future validation steps.

Keywords: Biomedical Instruments and Systems, Robot Mechanism and Control
Abstract: Recently, a double-layer five-bar robot designed for epidural pain intervention has been proposed. Its procedure is typically performed by clinicians under X-ray guidance. In traditional workflows, controlling the robot requires acquiring multiple intraoperative images, which is highly time-consuming and causes significant radiation hazards on medical staff. To mitigate these limitations, we propose the initial phase of an image-guided automatic alignment method. This method processes the captured X-ray images to automatically recognize a metallic marking ring, which serves as a reference to determine the spatial scale ratio between image pixels and physical dimensions (mm/pixel). Tested on real X-ray images from animal test, the proposed image recognition algorithm achieved a success rate of 80% in detecting the marking ring. As a preliminary study prior to the actual automatic alignment, this work successfully validates the feasibility of spatial scale calibration, establishing a critical foundation for subsequent autonomous robot-source alignment.

Keywords: Control Theory and Applications, Robot Mechanism and Control
Abstract: Tendon-sheath mechanisms are widely used for flexible surgical and endoscopic robots, but friction, elastic deformation, and slack cause hysteresis between motor input and distal motion. This paper presents a preliminary tension-aware hysteresis compensation framework for an antagonistic tendon-sheath mechanism with hybrid feedback control. Unlike a purely offline curve-fitting approach, the proposed framework considers both prediction accuracy and over-tension risk. The pulling-side actuator follows motor-position feedback with a real-time measured-signal-based compensation term, while the releasing-side actuator uses load-cell tension feedback to maintain a reference tension and suppress slack-induced dead-zone behavior. Candidate compensation models are identified using tension, velocity, and acceleration terms, and are evaluated by offline prediction error, high-tension response, and tension sensitivity. A 23 N evaluation point, corresponding to approximately 5% elastic input loss in the present setup, is used as an experimental operating-risk criterion. Preliminary switching experiments show that asymmetric model selection for the two antagonistic motors reduces hysteresis width across multiple paths while keeping the observed peak tension below the 23 N criterion. Future work will further analyze switching transients, holding-section effects, and predictive over-tension mitigation.

Keywords: Sensors and Signal Processing, Autonomous Vehicle Systems
Abstract: Object detection is a key perception task for safe autonomous driving and reliable path planning. LiDAR sensors are widely used in autonomous vehicles due to their accurate range measurements. However, LiDAR-based perception can be degraded by adverse weather conditions. Particularly, snowfall is challenging because snow particles generate noise points and reduce object points in measured signals. This paper proposes a snow particle filtering algorithm that combines LiDAR intensity characteristics and spatiotemporal consistency in sequential frames. First, the proposed method selects snow particle candidates using the low intensity characteristics of snow particles and a LiDAR signal model. Second, the remaining snow candidates are removed using frame-to-frame registration and a neighborhood-based density filter. The registration step uses vehicle kinematics to correct point-wise motion distortion caused by the sequential scanning operation of LiDAR. As a result, the proposed method removes sparse and inconsistent snow particles while preserving object points. The proposed algorithm improves the robustness of LiDAR perception such as pole-like object detection under heavy snowfall conditions with low computation time compared with existing filtering methods, including DROR and LIOR.

Keywords: Control Theory and Applications, Robot Mechanism and Control, Artificial Intelligence Systems
Abstract: This paper reports a reinforcement-learning redesign for knee assistance with a nonlinear torsional series elastic actuator (SEA). The goal is to exploit a fxed nonlinear defection-torque curve under limited runtime sensing. The runtime policy receives only gait-phase and motor-side signals; it does not receive the dataset torque target, knee angle, SEA defection, or delivered assist torque. Unlike an earlier absolute position-target formulation, the fnal controller uses PPO to output a normalized, rate-limited motor-equilibrium increment. A no-teacher PPO run on the canonical Fukuchi overground young-adult profle reduces deterministic evaluation RMSE from 6.46 Nm for a PD gain-scheduling baseline to 0.706 Nm over 25 evaluation episodes. The delivered SEA assist torque reaches 0.988 correlation with the hidden target profle, 0.54/1.88 Nm quiet-zone mean/max assist, 0.51% gait-cycle peak timing error, zero unsafe termination, and 8.46% action saturation. The result demonstrates deployable preload-command learning in a simplifed simulation while identifying remaining limits: prescribed knee motion, a single canonical profile, assumed spring parameters, and a high-speed actuator assumption

Keywords: Robotic Applications, Sensors and Signal Processing, Industrial Applications of Control
Abstract: In this study we propose a method for recognizing small tableware, such as cutlery, on a table using three-dimensional (3D) point cloud data, aiming to automate tableware collection using robots in the restaurant industry. In the previous study, we proposed cutlery recognition method using a glass table to clarify the differences in height between the surface of the table and the cutlery. However, its versatility for deployment in actual restaurants remained issues. Therefore, we propose a more practical method that utilizes a table with infrared-absorbing material surface and received light intensity. Cutlery made from various materials was placed on simulated table surfaces—constructed from plywood and aluminum plates—that had been covered with light-absorbing sheets or coated with an anti-reflective agent; the cutlery was then recognized. The results confirmed that the light-absorbing surface reduced the received light intensity, creating a distinct contrast between the table surface and the cutlery. While metal cutlery exhibited fluctuations in received light intensity due to specular reflection, the shape data for wooden and plastic cutlery could be extracted clearly and with high uniformity. These findings demonstrate that cutlery placed on a table surfaced with an infrared-absorbing material can be recognized by observing the difference in received light intensity, thereby confirming the effectiveness of the method proposed in this study.

Keywords: Robot Mechanism and Control
Abstract: In this study, we develop an assistive arm module for electric wheelchairs to support daily activities. The proposed arm combines a cam–spring gravity compensation mechanism with a belt–pulley transmission, enabling independent gravity compensation at each joint. This design aims to realize a lightweight manipulator with sufficient payload capability even when using low-power actuators.

Keywords: Robot Mechanism and Control
Abstract: Anti-corrosion coating of cage-type railway catenary poles is often performed manually by workers climbing the poles, which involves significant risks associated with working at height. In addition, labor shortages and the decline in the number of skilled workers in the coating industry have increased the demand for automation of painting operations. This study focuses on a hybrid robotic arm that combines the high rigidity of a delta mechanism with the large workspace of a serial mechanism while providing sufficient dexterity for painting tasks. In this paper, a hybrid-structured follower arm was designed and fabricated to satisfy the coating workspace required for actual railway catenary poles. The robotic arm dimensions were determined through workspace simulations, and the mechanism was designed with a hollow structure that allows a paint supply tube to pass through the robotic arm. Furthermore, a downsized leader arm was designed and fabricated based on the dimensions of the follower arm, and a leader–follower teleoperation system using simple joint-angle communication was implemented.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Navigation, Guidance and Control
Abstract: Snake-like robots exhibit high mobility on rough terrain; however, they consume significant power and tend to tip over in the roll direction on undulating surfaces. To achieve improved propulsion and posture stabilization, this study investigated a novel snake-like robot equipped with active roll joints. We constructed a hierarchical control system combining a Central Pattern Generator (CPG) with Proximal Policy Optimization (PPO) to adapt to unknown terrains. Through large-scale parallel simulations using the physics engine Genesis, we discovered an insightful phenomenon: naively incorporating active roll joints severely degrades locomotion speed and maneuverability. Analysis of the joint angle trajectories revealed that the rolling motion causes the rigid link edges to dig into the ground, physically locking the pitch joints and preventing the lifting motion necessary for sidewinding. This study highlights that expanding the degrees of freedom in snake-like robots strictly requires morphological co-design, such as cross-sectional shape optimization, to prevent physical interference with the environment.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robot Vision
Abstract: To streamline the simulation and design of microrobots with multi-articulated structures, constructing accurate kinematic models of complex micro-mechanisms is essential. However, manually defining joint positions and link connections is highly time-consuming.In this study, we propose a method to automatically generate a robot description model (URDF) of micro-articulated objects by extending the AutoURDF framework and estimating the joint structure from time-series point cloud data using unsupervised learning. We adopted a high-precision 3D grasshopper model as a benchmark to facilitate rigorous quantitative evaluation of the proposed method and the acquisition of ground truth data. To resolve challenges in the part segmentation of fine structures, we introduce point cloud preprocessing techniques such as model scale enlargement, local high-density sampling around moving parts, and the application of an expanded initial pose to explicitly separate parts. Furthermore, we directly extract key kinematic parameters, such as joint positions, rotation axes, and motion limits, from displacements across time-series frames. By synthesizing the sub-models extracted with high precision through these processes, we successfully generated a fully functional 12-joint model, demonstrating the effectiveness of the proposed method.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: This paper proposes a disentangled state reconstruction framework for humanoid locomotion. Proprioceptive humanoid locomotion over diverse terrains remains a challenging problem due to instability arising from bipedal contact. While quadrupedal robots exhibit high stability and can robustly traverse rough terrains, humanoid locomotion has a limitation in that it is difficult to handle such terrains without accurate terrain estimation. To address this issue, various methods have been proposed to estimate extrinsic states of the robot, such as terrain information, solely from observation history. However, since intrinsic and extrinsic states are estimated jointly, the representations become entangled, causing interference between the information required for each. To overcome this limitation, we propose Disentangled State Reconstruction (DSR), which separates state reconstruction based on temporal scales. Specifically, intrinsic states, such as the robot’s linear velocity, are reconstructed using short-term observation history, whereas extrinsic states, such as terrain information, are reconstructed using longer-term observation history. The proposed DSR demonstrates improved locomotion performance on most evaluated terrains in simulation, particularly on terrains involving discontinuous or irregular contact interactions.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: Wheel-legged quadruped robots need to combine three behaviorally distinct skills: rolling at speed, crossing stairs, and recovering from falls. Because their training objectives can conflict, a single end-to-end policy can struggle to learn all three at once. HieMoE-WL is a three-stage hierarchical mixture-of-experts framework that decouples skill acquisition from composition. Dedicated experts are learned, a privileged heightmap-gated teacher composes them via soft gating, and DAgger distills a proprio-only student. In IsaacLab on the Unitree Go2-W, the student matches its teacher and completes 97% of the course, versus 77% for a tuned rule-based switch over the same experts. A solo Recovery expert achieves 100% isolated get-up but does not locomote afterward; composition achieves 83% and keeps moving. A directly trained monolithic baseline recovers from only a third of mid-course falls and finishes 35% of runs; under the tested configurations it did not retain all three skills at once.

Keywords: Robotic Applications, Autonomous Vehicle Systems, Navigation, Guidance and Control
Abstract: This paper presents a situation aware locomotion framework for two mobile collaborative robots operating in shared industrial environments. The proposed method models situation awareness through perception, comprehension, and projection to support locomotion decisions. Robot pose, load state, manipulator state, shared zone occupancy, obstacle state, and predicted inter robot conflict were used to select safe locomotion actions. The framework was implemented in simulation and evaluated in simulated industrial scenarios designed to match a feasible 4 m by 4 m physical test area. The proposed method was compared with two other baselines over multiple trials and randomized seeds. The results show that the situation aware method achieved 100% task success across all scenarios, while the independent and fixed priority baselines each achieved 33.3% overall success. The proposed method eliminated shared zone conflicts and safety stops, maintained the largest average minimum inter robot distance, and completed the tasks with the lowest average completion time. These results indicate that situational awareness can improve the locomotion of dual robots by combining load state, manipulator state, reasoning about the shared zone, and prediction of short-horizon conflicts.

Keywords: Robotic Applications, Autonomous Vehicle Systems, Navigation, Guidance and Control
Abstract: Quadruped robots are a promising platform for search and rescue missions because they can navigate cluttered indoor environments that may be restrictive for wheeled systems. However, in unknown rescue scenarios, autonomous exploration must balance map expansion with the likelihood of finding victims, which is not explicitly addressed by classical frontier selection strategies. This paper presents the Situation Aware Frontier Prioritization (SAFP) method for single robot quadruped search and rescue. The proposed approach preserves the frontier exploration framework, but extends frontier ranking with information gain, observation deficit, rescue relevance, terrain penalty, and travel cost. The method is evaluated in Gazebo simulation with a quadruped robot in two indoor rescue scenarios with different levels of difficulty. The first scenario is used as a sanity check, while the second introduces stronger clutter and frontier ambiguity. Experimental results show that all methods perform reliably in a simple scenario, whereas in a complex scenario is different. In that setting, the proposed method achieves the highest completion rate and the highest victim recovery among the evaluated approaches. These results indicate that SAFP is beneficial when frontier choice becomes nontrivial and rescue utility must be balanced against generic exploration objectives.

Keywords: Robotic Applications, Control Theory and Applications
Abstract: This study addresses state mismatch in the sequential execution of independently learned robot task policies. In sequential manipulation tasks, the terminal state of a preceding policy may vary across executions, leading to joint position and velocity mismatches between two policies. To reduce this mismatch, we propose a sliding-surface-based transition generation method that selects a target state from the first action chunk of the subsequent policy and generates an intermediate trajectory using joint position and velocity errors. The proposed dynamics is designed to provide trajectory-level convergence toward the selected target state. MuJoCo simulations with a Unitree G1 humanoid robot on sequential drawer-opening and pick-and-place tasks show that the proposed method achieves balanced performance in task success rate, transition smoothness, and terminal state error.

Keywords: Robotic Applications, Human-Robot Interaction
Abstract: Risk assessment for collaborative robot applications is still largely a manual, paper-based process that lacks formal traceability and scales poorly with system complexity. This paper maps the iterative ISO 12100 risk assessment process onto a MAPE-K (Monitor-Analyze-Plan-Execute-Knowledge) loop whose Knowledge layer is a directed graph with typed nodes and constrained edges. Each iteration corresponds to one ISO~12100 revision cycle, and all intermediate states are persisted, which keeps the reduction path from initial to residual risk traceable. Hazards are localized in time by normalized progress values and in space by the spatial extent of an entity. Each hazard records the validity band of its evidence, which reduces change detection to a comparison and enables incremental re-evaluation. A collaborative pick-and-place case study covers three iterations for hazard identification, countermeasure modeling, and selective re-evaluation. The architecture provides a methodological foundation for digital risk assessment tools whose granularity scales with the available data.

Keywords: Robotic Applications, Human-Robot Interaction, Artificial Intelligence Systems
Abstract: This research proposes a constrained motion planning framework for robot manipulators in human-robot interaction (HRI). For a non-redundant manipulator with a fully specified end-effector pose, additional requirements such as collision avoidance and self-collision avoidance are difficult to handle as simple null-space secondary tasks. This limitation makes it challenging to generate feasible joint-space trajectories in HRI environments where safety and kinematic constraints must be considered simultaneously. To address this limitation, collision- and self-collision-aware trajectories are generated using Rapidly-exploring Random Tree (RRT) and RRT* algorithms, and the resulting dataset is used to train a diffusion model that generates constraint-satisfying trajectories through guided sampling. To reduce the inference time required for iterative diffusion sampling, consistency distillation is applied, and a joint-weighted jerk regularization term is incorporated into the loss function to promote smoother trajectories. Simulation results show that the consistency model generates 150 trajectory candidates in less than 100 ms, maintains a high episode success rate, and substantially reduces joint and end-effector jerk when jerk regularization is applied.

Keywords: Robotic Applications, Human-Robot Interaction, Robot Vision
Abstract: Due to the rapid expansion of the elderly population and the growing scarcity of caregivers, there is an urgent need for safe, comfortable assistive robotic devices. This study develops a cloud-based robotic bathing system for elderly and bedridden people. The proposed system combines compliant robotic interaction with anatomy-aware semantic bodyregion perception and cloud-supported task processing, thereby enhancing both safety and operational flexibility. The system uses a 5-DoF robotic arm with a soft-contact end-effector to support gentle physical interaction during bathing. To support accurate body-region understanding, a SegFormer-based semantic segmentation model is used. The model was trained on a 20-class anatomical dataset collected from full-scale male and female mannequins. The dataset included different body postures, lighting conditions, and viewpoints to improve perception robustness. Experimental evaluation showed good segmentation performance with a mean Intersection over Union (mIoU) of 76.33%, pixel accuracy of 96.63%, and mean class accuracy of 92.95%. The robotic platform successfully reached 16 of 18 predefined anatomical target regions, achieving 88.9% workspace coverage with an average trajectory execution time of 16 s. Optimization of the perception pipeline and cloud-assisted processing reduced the total operational latency to approximately 2.443 s. These results demonstrate the prototype-level feasibility of integrating anatomy-aware perception, cloud coordination, and compliant robotic execution for adaptive robotic bathing assistance in elderly-care applications.

Keywords: Artificial Intelligence Systems, Robot Mechanism and Control, Human-Robot Interaction
Abstract: Simulation-based demonstration augmentation reduces the cost of collecting robot manipulation data by synthesizing training trajectories from a few human demonstrations. Prior augmentation methods add every generated trajectory to training, on the assumption that more generated data yields a better policy. We test this assumption in a resource-constrained regime — a low-cost manipulator with five joints and a gripper (SO-101) and a small Vision-Language-Action model (SmolVLA), fine-tuned on nine human demonstrations plus 50 generated ones — and find that it does not hold. The augmented policy succeeded in 28.3% of episodes, below the 40.0% human-only baseline. The failure took an unusual form: the policy froze mid-episode in 60% of episodes rather than manipulating poorly. The harmful episodes were nearly indistinguishable from clean ones under every static check we applied. The generation process, however, provides a usable signal: generation reliability decreases monotonically with distance from the source demonstrations. Selecting candidates by a single distance threshold derived from these failure logs raised the success rate to 65.0%, eliminated the freezing entirely, and produced a policy that outperformed the baseline beyond its training coverage. In low-data fine-tuning of small VLA models, augmentation succeeds or fails by data selection, not by data volume.

Keywords: Artificial Intelligence Systems, Robot Vision, Industrial Applications of Control
Abstract: Post-harvest mango grading needs lesion-coverage estimation beyond a disease label, yet dense lesion masks are expensive and exist for only some classes. This paper presents an empirical study of a coverage-based severity pipeline, focusing on reliability, supervision, and failure modes rather than architectural novelty. The pipeline (Disease-Aware Mango Coverage, DAMC) feeds a ResNet-50 classifier into a Feature Pyramid Network (FPN) segmentor, with optional rule-based post-processing. On a 126-image SenMangoFruitDDS test set, the central finding is that the raw FPN mask should be retained: against human polygon annotations it scored a higher mean Dice (0.6605) than the rule-gated (0.6448) or always-on rule (0.5357) output. The rule stage is therefore best kept only as a reliability flag, and a Stage-2-only pipeline suffices for deployment. The FPN coverage reached a quadratic-weighted Cohen’s κ of 0.8514 against expert grades, but this exceeded a coverage-blind class-prior baseline by only about 0.07 κ, so coverage mainly resolves within-class severity rather than replacing the disease label. Ground-truth masks alone achieved comparable agreement, indicating that pseudo-masks mainly serve as a practical fallback for unannotated classes. Coverage remains an imperfect proxy for severity: for Stem-End Rot, ripening false positives and possible label inconsistencies impose a ceiling that coverage alone cannot overcome. Because the evidence comes from a single small test set, one annotator, and one favorable split, the guidance is dataset-specific and calls for multi-annotator ground truth and external validation before it is generalized.

Keywords: Artificial Intelligence Systems, Robot Vision, Industrial Applications of Control
Abstract: Automated visual inspection in manufacturing requires high defect recall and human-interpretable rejection rationale—properties that existing score-based anomaly detectors seldom provide simultaneously. This paper presents a two-stage inspection framework and systematically investigates whether compact vision-language models (VLMs) can serve as semantic arbiters for borderline anomaly decisions. In the first stage, a PatchCore memory bank on a frozen DINOv2 ViT-B/14 backbone scores each image and routes it through a three-zone policy—direct pass, direct reject, or a VLM zone for borderline cases—with thresholds fixed on a held-out validation split. In the second stage, a compact VLM inspects the region-of-interest crop and returns a structured PASS/REJECT verdict, defaulting to REJECT for malformed responses (fail-safe). On MVTec AD the DINOv2 backbone reaches a mean AUROC of 0.947, matching a WideResNet-50 PatchCore baseline with category-dependent trade-offs. Our central finding is negative: every compact VLM tested fails as a Stage-2 arbiter. The three free-generation models (Moondream2, Phi-3.5-Vision, LLaVA-1.5-7B) collapse to an explicit reject-all baseline, while Qwen2.5-VL passes every borderline defect from 3B to 32B—reporting no defect at all, a grounding rather than a decision failure that constrained decoding does not fix. Cross-dataset evaluation on VisA shows degraded performance that a matched-ratio ablation attributes predominantly to reduced coreset density rather than domain shift.

Keywords: Artificial Intelligence Systems, Robotic Applications, Human-Robot Interaction
Abstract: This paper presents a Multilingual Speech Emotion Recognition (Multilingual SER) system using a Per-Language ResNet architecture trained separately for each of five languages: Thai, Chinese, Japanese, Korean, and English. A key finding from prior research is that a single Unified Multilingual Model suffers from Prosody Mismatch — each language has distinctly different pitch contours, speech rates, and energy patterns for the same emotion, causing the model to confuse language characteristics with emotion characteristics. The Per-Language approach resolves this by allowing each ResNet model to learn emotion patterns within the prosodic context of its own language. The system employs an automatic Language Detector (SVM pipeline: StandardScaler → PCA(80) → SVM RBF) that identifies the spoken language from 128×130 Mel Spectrogram features before routing audio to the appropriate emotion model. Experimental results on held-out test sets demonstrate per-language accuracy of: Thai 88.40%, Chinese 86.62%, Japanese 84.36%, English 82.84%, and Korean 72.62% (SVM deployed model). The overall deployed system accuracy is 82.97%, surpassing the 75% paper target. The Language Detector achieves 98.54% accuracy on a 5-class language classification task. The paper additionally includes a Gradio-based Web Demo and a cross-lingual feature analysis module for studying emotion pattern similarities across languages.

Keywords: Artificial Intelligence Systems, Robotic Applications, Human-Robot Interaction
Abstract: In this work, we present a test-driven, agentic framework for synthesizing a deployable robot controller for navigation tasks. Our approach uses large language models (LLMs) to generate controller code and evaluates it through user defined PyTest suites that capture task-specific performance and deployment requirements. To improve reliability and executability, we introduce a dual-tier repair strategy that alternates between prompt refinement and direct code editing, enabling iterative correction of reasoning and implementation errors. We evaluate the framework on (i) 2D map-based navigation, and (ii) 3D Webots simulation. Experimental results show that test-driven synthesis substantially improves controller reliability and robustness over one-shot controller generation, especially when the initial prompt is underspecified.

Keywords: Sensors and Signal Processing, Artificial Intelligence Systems, Multimedia Systems
Abstract: Automatic pronunciation assessment (APA) gives second-language learners fine-grained, multi-aspect feedback at the phone, word, and utterance levels. State-of-the-art systems increasingly fine-tune large self-supervised speech encoders. Updating hundreds of millions of parameters, however, is computationally expensive and tends to over-fit the few thousand expert-scored utterances typical of APA datasets. We show that once the encoder is frozen, the dominant lever is not trunk size but complementary per-phone feature engineering and training objectives. Our model, Qwen3-APA, keeps a pre-trained Qwen3-ASR encoder frozen and trains only a compact 1.26M-parameter hierarchical head, comparable in trainable size to the HMamba baseline. A learned gated fusion combines eight complementary per-phone streams: four goodness-of-pronunciation (GOP) variants, namely canonical, class-normalized, second-recognizer, and SSL-template, together with articulatory, prosodic, and two self-supervised views. The head then refines the phone-, word-, and utterance-level representations with a dilated temporal-convolution branch and a residual graph-convolution module, all in a single forward pass. Qwen3-APA ranks first among the compared systems on eight of the ten SpeechOcean762 metrics, including a sentence-level total PCC of 0.839. It also reaches 0.746 on completeness, well above the 0.16–0.51 of prior systems.

Keywords: Process Control Systems, Artificial Intelligence Systems, Robotic Applications
Abstract: This study proposes a two stage robust scheduling framework for 24 h scheduling of a renewable energy integrated reversible solid oxide cell (RSOC) system. The proposed framework aims to coordinate the operation of the RSOC, BESS, utility grid, and hydrogen storage under renewable generation and process response uncertainties. Mode specific process responses of the SOFC and SOEC modes are generated from IDAES based steady state process models and converted into affine surrogate models for integration into a MILP scheduling formulation. The scheduling model jointly considers RSOC mode selection, electricity and hydrogen storage balances, BESS operation, hot standby operation, and a thermal cycling proxy. The resulting scheduling problem is formulated as a two stage robust MILP and solved using a column and constraint generation framework. The 24 h robust scheduling results show that the proposed model provides stable operation under the identified worst case renewable scenario while satisfying electricity and hydrogen balance constraints. These results demonstrate the applicability of the proposed framework to robust operation of RSOC systems under renewable uncertainty.

Keywords: Artificial Intelligence Systems, Process Control Systems, Sensors and Signal Processing
Abstract: Although unsupervised domain adaptation (UDA) has emerged as a promising solution to mitigate data distribution shifts caused by varying operating conditions, conventional UDA methods typically assume the availability of all fault classes in the target domain. In practical chemical processes, however, collecting a comprehensive set of target fault data is highly challenging due to the scarcity of fault occurrences, whereas normal operating data is easily accessible. To address this extreme class-asymmetric challenge, this paper proposes a generative UDA fault diagnosis framework tailored for scenarios where the target domain fault data is entirely unavailable. The proposed methodology executes a multi-level statistics alignment coupled with a style transfer mechanism. This approach dynamically injects channel-wise target normal statistics into source fault features, thereby synthesizing pseudo-fault samples that reflect target-specific process dynamics without disrupting the latent space topology. Additionally, a supervised contrastive learning objective is integrated to prevent feature collapse and effectively expand the generalization margins of fault decision boundaries toward the target domain. Evaluated on the Tennessee Eastman process benchmark, the proposed framework overwhelmingly outperforms existing UDA and domain generalization techniques, demonstrating robust diagnostic performance even with minimal target normal samples.

Keywords: Process Control Systems, Artificial Intelligence Systems, Industrial Applications of Control
Abstract: Most existing reinforcement learning (RL) controllers assume its operation on stationary Markov decision process (MDP), which limits their applicability to chemical processes operating under continuously evolving conditions and disturbances. Factors such as catalyst deactivation, external disturbances and equipment fouling induce variations in process dynamics, necessitating adaptive control strategies with justifications. To address this issue, we develop an adaptive RL control framework for non-stationary MDPs that learns a latent context belief from recent state-action trajectories and augments it to the state representation. Based on meta-RL, the controller is exposed to multiple operating conditions, training additional context encoder network to quickly estimate current drift status with minimum amount of samples. The learned latent context also enables analysis of underlying MDP variations throughout process evolution. Experimental results on a continuous drift scenario within CSTR reactor benchmark demonstrate the adaptivity and robustness of our method to process drift. Moreover, the latent vector has been well aligned to the parameter changes, providing operators with additional insight into the model's decision under process drift.

Keywords: Robot Mechanism and Control, Robotic Applications, Robot Vision
Abstract: Vision-Language-Action models promise general purpose robot manipulation, yet existing benchmarks are dominated by household pick-and-place tasks and hardly evaluate the wet lab manipulation skills required for chemistry automation such as transferring glassware, pouring liquids between glassware, and rotating valves. We present a factorial study that compares one specialist imitation policy, Action Chunking with Transformers, with two generalist Vision-Language-Action models, π0 and π0.5, across three chemistry lab manipulation tasks and three teleoperation dataset sizes of 50, 100, and 200 demonstrations, trained on identical hardware. Each of the resulting nine conditions defined by model and dataset size is evaluated on every task by five trials of two consecutive attempts, for a total of 270 attempts across the three tasks. We report a 50-point rubric score, computed as 10 points per trial summed over the five trials of each condition, as the primary metric, and time-to-success as a secondary indicator. The results characterize the data efficiency knee of each model family on chemistry tasks and inform practical model selection for laboratory deployments.

Keywords: Industrial Applications of Control, Artificial Intelligence Systems, Process Control Systems
Abstract: Reinforcement learning (RL) offers a powerful data-driven paradigm for chemical process control, yet safety constraints and biased operation logs pose significant practical barriers to its deployment. We present GRAM-QT (Graph-based Regime-Aware Mixture-of-Q-Transformers), an offline RL framework that addresses these challenges through self-supervised semantic regime partitioning. Raw operational data are encoded as directed variable graphs and processed by a graph attention transformer, which captures complex multivariate dependencies while identifying distinct operating regimes in an interpretable manner. Rather than relying on a single monolithic policy, the framework deploys a specialized expert Q-transformer for each discovered regime and synthesizes control actions by aggregating regime-specific Q-values, weighted by the membership of the current process state in each regime. This avoids the unreliability of global policy averaging and sustains performance even under data-sparse, out-of-distribution conditions. On the industrial-scale IndPenSim fed-batch penicillin benchmark, GRAM-QT recovers the biological growth phases without supervision and exhibits smooth, interpretable expert switching aligned with the underlying phase transitions. As a result, it improves penicillin yield by 43% over a single global Q-transformer while outperforming representative offline RL baselines.

Keywords: Autonomous Vehicle Systems, Control Theory and Applications, Navigation, Guidance and Control
Abstract: This paper presents a disturbance observer–based model predictive control (DOB–MPC) framework with heterogeneous actuation dead-time compensation for precise trajectory tracking of autonomous surface vessels (ASVs). ASVs pose significant control challenges due to underactuation, environmental disturbances such as wind, waves, and currents, and different actuator dynamics, where command transmission, engine response, and propeller slip introduce heterogeneous input-to-thrust time delay. Ignoring these heterogeneous delay characteristics leads to inaccurate disturbance estimation and degraded tracking performance. The proposed framework integrates a dead-time compensator (DTC) and a DOB within MPC: the DTC predicts future vessel states over the dead-time horizon to provide a delay-compensated initial condition for MPC while synchronizing heterogeneous actuator inputs for consistent disturbance estimation, while the DOB estimates disturbances and model uncertainties for real-time compensation within the MPC prediction model. The effectiveness of the proposed framework is validated through numerical simulations and full-scale sea trials on an 8 m ASV, demonstrating substantial improvements in trajectory-tracking accuracy compared with conventional MPC methods.

Keywords: Autonomous Delivery Vehicles, Leader-Velocity Prediction, Flocking Control, Cooperative Control
Abstract: A two-layer cooperative control strategy for multiple autonomous delivery vehicles (ADVs) is proposed in this paper. In the upper layer, a sequence model is trained offline using historical driving data to predict the velocity of the leader ADV, and a Transformer is selected according to a predefined validation criterion. In the lower layer, a leader-velocity-guided flocking controller based on the Olfati-Saber model is adopted for online cooperative control of the follower ADVs. The predicted leader velocity is introduced into the lower-layer controller as a time-varying reference, enabling the follower ADVs to respond in advance to changes in leader motion. Simulations covering the complete route are conducted with different leader velocity inputs to verify the proposed strategy. The results show that the predicted velocity reference reduces tracking error compared with the current and delayed leader velocity inputs. These results demonstrate the effectiveness of integrating offline leader-velocity prediction with online multi-ADV flocking control under the considered simulation conditions.

Keywords: Autonomous Vehicle Systems, Navigation, Guidance and Control
Abstract: Aerial-ground robots improve energy efficiency through ground driving while retaining the ability to traverse obstacles by flight. However, global planning for such robots must jointly decide path geometry and ground-versus-aerial mode selection while remaining computationally efficient in large-scale 3D environments. Existing aerial-ground navigation planners can generate feasible energy-aware paths, but grid-based or decoupled planning methods often struggle to achieve short path geometry, mode-aware behavior, and low online search time simultaneously. To address this gap, we propose a multi-resolution any-angle global planner for aerial-ground robots. The planner combines hierarchical free-space expansion, Lazy Theta*-style direct connection, and a mode-aware edge cost for ground travel, aerial travel, altitude change, and mode transition. Because this cost is evaluated inside the same search that produces the any-angle geometry, the planner generates platform-aware paths without separating geometric path generation from platform-mode selection. In simulation across open, wall-passage, and cluttered environments, the proposed planner solves all queries and achieves the lowest search time and shortest path length among the evaluated baselines. In the hardest environment, where the grid-based energy-aware baseline times out under the shared 5 s budget, it reduces mean online search time by 66.2% and path length by 5.4% relative to the only remaining baseline that solves all queries. The proposed method therefore provides fast global path generation and the shortest evaluated paths, with energy remaining comparable by using aerial segments selectively where they provide useful obstacle crossings for aerial-ground robots.

Keywords: Autonomous Vehicle Systems, Navigation, Guidance and Control, Control Theory and Applications
Abstract: This paper proposes a fully decentralized and time-extended task-allocation framework for heterogeneous agricultural multi-robot systems in fruit-harvesting applications. A Bayesian Network with an integrated priority mechanism is developed to evaluate the success probability of each task and the success utility of each task bundle by jointly considering robot hardware capability, task requirements, energy fulfillment, and task priority. Based on these inferred probabilities, each robot independently evaluates feasible task bundles and submits bids through a decentralized auction process. A genetic algorithm is applied to determine the optimal execution order within each task bundle, and mixed integer linear programming is used to identify the allocation solution that maximizes the overall success probability. Experiments conducted in an indoor simulated agricultural environment using two heterogeneous mobile robots demonstrate that the proposed framework can reliably allocate tasks according to robot hardware capabilities and task information while maintaining a high overall task success probability, reliability, and adaptability under varying harvesting conditions and task-priority settings.

Keywords: Sensors and Signal Processing, Robotic Applications
Abstract: Conventional seismic data acquisition involves several labor-intensive stages, including delivery, placement of sensors, data recording, and sensor retrieval, which are accompanied by significant time and labor costs, as well as operational challenges in difficult terrain and extreme climatic conditions. For this reason, there is a need to automate field operations. One of the developing areas is the use of UAVs for sensor delivery, deployment, data acquisition, and return to base. This paper presents experimental results obtained during the modernization of these UAVs for seismic data acquisition by switching to recording seismic data using accelerometers only. Two experiments were conducted: the first was aimed at validating the performance of the new UAV configuration under field conditions, while the second focused on comparing the quality of the recorded seismic signal among different versions of Autonomous Seismic Acquisition Devices (ASADs). The results showed that the updated version of the robot is superior to the previous generation in terms of signal quality, while also providing greater reliability and ease of use.

Keywords: Autonomous Vehicle Systems, Navigation, Guidance and Control, Control Theory and Applications
Abstract: This paper presents a fully integrated guidance and control framework for autonomous unmanned surface vessels (USVs) that achieves high-precision trajectory tracking in real maritime environments. The framework integrates two tightly coupled modules: (1) a Lyapunov-based adaptive RBF nonlinear controller with online parameter estimation, whose stability is rigorously established via Lyapunov stability analysis together with Barbalat's Lemma; and (2) a Lagrange multiplier-based constrained optimal control allocation scheme for rotatable waterjet propulsors, augmented with dedicated RBF networks for accurate actuator modeling. The proposed architecture has been successfully implemented and validated on the Lab611 USV. Extensive real-sea experiments demonstrate exceptional performance, achieving sub-meter tracking accuracy with average tracking errors of 0.26 m over an approximately 2 km ring-port trajectory, 0.14 m on straight-line-plus-curve missions in constrained narrow channels, and 0.42 m on serpentine cruising paths, all while strictly respecting actuator saturation constraints. These results confirm the practical feasibility, robustness, and superior tracking accuracy of the integrated approach in realistic sea states (0-3).

Keywords: Human-Robot Interaction, Artificial Intelligence Systems, Robotic Applications
Abstract: This paper proposes a software-defined context-aware Virtual Fixture (CAVF) system that leverages a Vision-Language Model (VLM) to automatically guide robot end-effector motion during teleoperation. The proposed system preserves task-relevant motion while suppressing unnecessary translational and rotational components, enabling haptic-free assistance without handcrafted rules. Experiments in simulation and real-world environments show that CAVF improves task success rates, reduces completion times, and enhances trajectory efficiency compared with conventional teleoperation.

Keywords: Human-Robot Interaction, Robot Mechanism and Control, Robotic Applications
Abstract: Reading aloud plays an important role in children’s language development and reading engagement. Recent reading companion robots have explored verbal interaction to support children’s reading activities; however, most existing systems are designed as standalone platforms positioned separately from the book, making it difficult to naturally integrate robot interaction into children’s reading space. This study proposes Tokki-Rang, a book-attached reading companion robot designed to provide nonverbal companionship interaction during children’s read-aloud activities. By directly attaching to a book, Tokki-Rang is designed to remain within the child’s reading field of view and support interaction while minimizing additional demands during reading. The robot expresses listening and engagement behaviors through pneumatic soft ear motions and vertical body movements. Interaction scenarios were developed based on children’s interpretations of reading-related situations and robot motions, enabling the robot to convey listening, attention, and engagement through nonverbal embodied behaviors. This work illustrates the potential of book-attached, nonverbal reading companion robots as an interaction design approach for supporting children’s shared reading experiences.

Keywords: Human-Robot Interaction
Abstract: In this paper, we propose an imitation-based grasp pose generation method that combines human motion guidance with point-cloud sensing of target objects for robotic teleoperation. The operator provides intuitive grasping intent through an approaching vector and hand direction vector extracted from human motion. Based on this guidance, an initial human-guided grasp pose is obtained and subsequently refined using the point cloud acquired by a depth camera mounted on the robot end-effector. Unlike learning-based grasp generation methods, the proposed approach preserves real-time human intent while requiring no training data. Experimental results demonstrate the feasibility of the proposed imitation-based grasping strategy, achieving a grasp success rate of 83.6% across multiple object categories and grasping directions.

Keywords: Human-Robot Interaction, Robot Vision, Robotic Applications
Abstract: Speed and separation monitoring reduces robot speed before hazardous human-robot contact can occur, through a protective separation distance that combines human motion, robot reaction and stopping behavior, sensing uncertainty, and geometric margins. In many vision based deployments the human approach speed and the sensing margin stay fixed at conservative values, so the monitored boundary cannot follow the direction of human motion or the quality of the perception stream. This paper presents a markerless multi RGB-D monitor that updates these two perception dependent terms online from a fused skeleton. Human keypoint motion is projected toward the robot link geometry to obtain a geometry-aware approach speed, and a cell specific RGB-D calibration residual is added to the sensing margin while the standard additive form of the separation distance is retained. A recorded UR5e workcell evaluation compares the monitor against a fixed-speed, link aware baseline under identical geometry. The proposed monitor lowers the separation distance and the robot interruption while observed robot link clearance is maintained throughout the replay. Markerless perception is positioned as an online input source for dynamic safety boundaries rather than as a safety certified controller.

Keywords: Human-Robot Interaction, Industrial Applications of Control, Robotic Applications
Abstract: This paper presents a leader-follower teleoperation framework with whole-body collision avoidance for heterogeneous robot manipulators. The proposed system transfers the end-effector motion of an Open-Manipulator-X leader to a Franka Emika Panda follower while enforcing whole-body collision avoidance through an operational space control barrier function (OSCBF)-based safety filter. The follower first computes a nominal operational space torque command for tracking the desired end-effector pose and then solves a quadratic program that minimally modifies the command to satisfy safety constraints, with emphasis on maintaining separation between robot collision spheres and obstacles. To evaluate the safety behavior under time-varying constraints, a virtual spherical obstacle is generated in the follower workspace and moved along the lateral axis. The resulting controller enables intuitive teleoperation while reacting to both static and moving obstacle constraints.

Keywords: Human-Robot Interaction, Autonomous Vehicle Systems, Robot Mechanism and Control
Abstract: As autonomous-driving regulations, such as UN Regulation No. 171 and Euro NCAP, become increasingly stringent, robust verification technologies for high-reliability Driver Monitoring System (DMS) are critical. However, conventional evaluations relying on human participants are limited in generating repeatable reference data due to significant gaze dispersion and ocular fatigue. To address these limitations, this study quantitatively verified the gaze reproduction accuracy of an in-house developed driver behavior simulation dummy through a one-to-one comparison with human subjects. Experiments were conducted in a driving simulator under two Euro NCAP protocol-based behavior patterns: owl-like and lizard-like. The dummy system demonstrated lower standard deviations than human participants, proving high mechanical repeatability. Although noticeable gaze errors were observed in the peripheral regions of the lizard-like behavior (Yaw -20° and +30°) due to eyeball surface material differences and initial geometric alignment constraints, stable tracking performance was validated within the primary effective gaze region. These findings confirm the dummy’s validity as a high-reliability evaluation alternative. Future work will focus on minimizing errors by improving artificial-eyeball materials and enhancing alignment precision.

Keywords: Human-Robot Interaction, Robotic Applications, Control Theory and Applications
Abstract: Intuitive hand guiding is a fundamental feature in physical human-robot interaction, enabling operators to teach trajectories without complex programming. However, uncompensated motor-side friction substantially degrades operational transparency. While a motor-side friction observer can effectively reduce interaction resistance, its high sensitivity makes the hand-guiding loop vulnerable to various sources of uncertainty, such as dynamic modeling errors and the inherent bias and thermal drift of the joint torque sensor (JTS). When the manipulator is released and expected to remain stationary, these residual uncertainties are misinterpreted as a persistent external input, leading to unintended continuous motion, or drift. In this paper, we present a hand-guiding framework that applies a motor-side L1 Adaptive Friction Observer (L1AFO) integrated with a dual-stage stabilization strategy. By incorporating a linearly interpolated dead-zone and an online moving-average residual compensation mechanism during stationary phases, the proposed method systematically suppresses the residual uncertainty that induces the drift. Experimental validations on a multidegree-of-freedom collaborative robot demonstrate that the proposed framework reduces the average interaction force by 44–64% during hand guiding while suppressing unintended drift throughout a 10-minute stationary evaluation.

Keywords: Human-Robot Interaction, Artificial Intelligence Systems, Robot Mechanism and Control
Abstract: Recent latent world models such as V-JEPA learn predictive representations by forecasting future latent states. Although such latent representations are effective for visual prediction, they are difficult to interpret and often contain redundant information irrelevant to downstream robot planning. This paper proposes Concept-JEPA, an idea-level framework that replaces latent-space prediction with semantic concept dynamics. The proposed method extracts latent tokens using a frozen V-JEPA encoder, projects them into automatically generated semantic concepts using vision-language supervision, and selects task-relevant concepts through an adaptive Top-K mechanism. A lightweight concept dynamics model predicts future semantic concepts, which are then used for robot action prediction. By modeling semantic state transitions rather than full latent vectors, Concept-JEPA provides an efficient and explainable direction for future robot world models.

Keywords: Human-Robot Interaction
Abstract: Haptic feedback is essential for achieving immersive and precise interactions. Specifically, multi-directional shear forces applied to the fingertip play a crucial role in the perception of object slippage, surface textures, and directional movements. Herein, we propose a 2.8 g ultra-lightweight wearable haptic multi-directional soft pneumatic device, generating both normal and multi-directional shear force. The proposed system integrates a forearm-worn origami pump and valve module to achieve independent control of three pouches using only two origami pumps and a single solenoid valve through a pneumatic method. With this designed device, the performance of each pouch was evaluated to show that it can deliver shear forces in eight directions as well as normal. This novel haptic device demonstrates the potential to provide various tactile feedback to the fingertip using the independent pneumatic method, thereby enabling lightweight and wireless devices.

Keywords: Human-Robot Interaction, Robotic Applications, Industrial Applications of Control
Abstract: This study proposes an intuitive single-master control framework for dual-arm teleoperation to alleviate operator cognitive workload and simplify task complexity in bimanual manipulation. The proposed scheme features two coordinated kinematic mapping modes: a synchronized translation mode for the stable co-transportation of heavy payloads, and a symmetric mirror mode to facilitate intuitive end-effector alignment during dual-arm grasping. To prevent operational drift and enhance positioning accuracy, active virtual fixtures are integrated into the bilateral control loop to provide reactive haptic guidance. The effectiveness of the proposed framework was evaluated through a user study involving nine participants categorized by proficiency levels (experts, intermediates, and novices) performing a bimanual pipe transportation task. Experimental results demonstrated that the proposed haptic-assisted scheme achieved an average trial-wise task time reduction of 25.1% and an average subjective workload reduction of 45.1%, validating its practical utility in reducing bimanual coordination fatigue.

Keywords: Industrial Applications of Control, Process Control Systems
Abstract: As the nuclear industry shifts toward Small Modular Reactors (SMRs), the adoption of multi-module operations and high levels of automation has become essential to ensure economic feasibility. However, increased automation can lead to the “Out-of-the-Loop (OOTL)” phenomenon. This phenomenon degrades operators’ Situation Awareness (SA) of plant system states during automated operations and hinders effective operator responses to automation failures. In multi-module SMRs, this issue is further exacerbated by the need to monitor and manage multiple units simultaneously, leading to dispersed attention and reduced system understanding. To address these challenges, this study proposes a conceptual HSI design framework consisting of three strategies: integrated overview display, system-level information integration and automation transparency with mode visualization. The integrated overview display supports attention allocation across multiple modules; system-level information integration helps operators understand interactions among modules, shared systems, and plant-level operations; and automation transparency with mode visualization supports operators in understanding automation status, rationale, and expected behavior. These strategies are derived from prior studies on OOTL, SA and regulatory guidance for advanced reactor HSI design. The proposed framework is expected to help operators maintain SA and remain in the control loop during automated operation in multi-module SMR.

Keywords: Industrial Applications of Control, Robot Mechanism and Control, Robotic Applications
Abstract: We present an incremental joint-wise tuning protocol for Active Disturbance Rejection Control (ADRC) on a 6-DOF Neuromeka Indy7 manipulator, releasing joints one at a time from the end-link toward the base so that inter-link coupling is absorbed stage by stage rather than as a single multi-joint search. Against a strong PD + feed-forward (FF) baseline that is given the full nominal robot model M(q), C(q,˙q), G(q) and a friction model, and that shares the identical PD feedback as ADRC, ADRC outperforms the baseline in both scenarios using only one gain bn per joint as model knowledge: by 12 % RMS in the nominal scenario, and by 76 % RMS when a 2.5 kg payload — unknown to the nominal model — is rigidly attached to the end-effector, at essentially the same control effort.

Keywords: Industrial Applications of Control, Control Devices and Instruments
Abstract: This paper presents the development and verification of a simulation testbed for a piezo-actuated nanopositioning control system, specifically tailored for the Z scanner servo system of an atomic force microscope (AFM). Nanopositioning systems driven by piezo actuators are essential in high-precision instruments to achieve sub-nanometer resolution. However, evaluating control performance and implementing advanced control algorithms directly on hardware can be time-consuming and carries the risk of device damage. To address this, we construct a Simulink-based testbed incorporating a discretized proportional-integral (PI) controller, high-voltage amplifier dynamics, stacked piezo actuator characteristics with flexure structures, and cantilever sensor dynamics. The closed-loop frequency response and step response are evaluated, and the testbed's ability to mirror physical conditions is validated through topography-tracking scenarios under white Gaussian noise. Simulation results demonstrate that the proposed testbed effectively captures tracking speed and noise amplification when adjusting controller gains. This virtual framework provides a valuable and safe platform for control engineers to optimize nanopositioning control systems and evaluate newly designed control logic.

Keywords: Industrial Applications of Control, Control Theory and Applications
Abstract: This paper proposes an automatic calibration method for the input gain of active disturbance rejection control (ADRC) applied to permanent magnet synchronous motor (PMSM) position control. In ADRC, the input gain depends on mechanical parameters such as rotor inertia and flux linkage, which are difficult to identify precisely and vary with operating conditions. A gain mismatch causes the closed-loop bandwidth to deviate from the desired value, resulting in degraded tracking performance. The proposed method applies a sinusoidal calibration signal and constructs a reference model using a second-order low-pass filter whose bandwidth matches the ADRC control bandwidth. Rather than relying on instantaneous tracking error, the method compares the steady-state sinusoidal amplitudes of the closed-loop output and the reference model output to identify the gain mismatch. The amplitude comparison is performed by integrating the absolute value of each signal over one half-cycle using resettable integrators triggered by zero-crossing instants, yielding a phase-independent amplitude measure. A Lyapunov-based convergence analysis shows that the gain estimate converges to the true value in finite time under ideal conditions. The effectiveness of the proposed method is verified through simulations in MATLAB/Simulink.

Keywords: Industrial Applications of Control, Control Theory and Applications
Abstract: Direct-drive rotary permanent magnet synchronous motors are widely used in industrial motion systems requiring high-speed and high-precision position control. For such systems, an accurate plant model is important for controller design, bandwidth selection, stability analysis, and performance prediction. This paper presents a frequency- response-function-based parameter identification procedure for an industrial DD rotary PMSM under closed-loop position control. A position-command exponential chirp experiment is performed, and the closed-loop FRF is estimated from the measured position response. The open-loop and plant FRFs are then sequentially reconstructed using the known controller and feedback model, and particle swarm optimization is applied to identify the effective mechanical parameters. A no- delay model and a position-control-period delay model are compared to analyze the influence of the delay term on FRF fitting. Experimental results show that the delay-included PSO model reduces the phase RMSEs of the closed-loop, open- loop, and plant FRFs by approximately 34.6%, 35.3%, and 35.4%, respectively, compared with the no-delay PSO model under the same magnitude-and-phase objective function. The magnitude RMSEs of all evaluated FRFs remain close to approximately 2.1 dB

Keywords: Industrial Applications of Control
Abstract: This paper proposes an adaptive internal model control design method for uncertain linear time invariant first order process with uncertain high speed actuator. Adaptive rules for the targeted systems are derived and the process and actuator nominal models are updated through adaptive gains. To acquire the control parameters meeting requirements, the performance index based parameter selection guideline is proposed. The effectiveness of the proposed method is evaluated through the gas turbine systems with uncertain parameters. The simulation results include the comparison with the conventional internal model control method without the model update by adaptive rule, which show that the proposed method has fast model recovery performance with notable disturbance rejection capability.

Keywords: Industrial Applications of Control, Information and Networking, Sensors and Signal Processing
Abstract: Printed Circuit Assemblies (PCA) are manufactured through a sequence of processes in a Surface Mount Technology (SMT) line, where electronic components are mounted and soldered onto Printed Circuit Boards (PCB). Product quality is typically assessed at the final stage of the line, primarily through automated optical inspection (AOI). Consequently, most existing public datasets for PCA defect analysis focus on AOI image data. Although image-centric datasets are effective for post-process defect detection, they provide limited insight into how manufacturing processes affect defects, which in turn hinders research on early-stage defect prediction and process-aware quality analysis. To fill this gap, we introduce TUK PCA, a multi-modal dataset collected from multiple stages of an SMT line. The dataset integrates manufacturing process logs from mounting equipment, thermal data from a reflow oven, and component-level region-of-interest (ROI) images and inspection results from AOI systems. Because these sources are produced by heterogeneous acquisition mechanisms, the dataset adopts a time-based synchronization strategy to align records across stages under practical constraints. We provide a detailed description of the data collection pipeline, dataset composition, and synchronization methodology, together with an analysis of the resulting dataset. This dataset aims to support research on process-aware defect prediction, multi-modal learning, and explainable quality analysis in PCA manufacturing.

Keywords: Industrial Applications of Control, Robotic Applications, Process Control Systems
Abstract: Industrial robots increasingly operate as components of larger workflows rather than as isolated machines, so their software must sustain continuous control while also serving ad-hoc requests from external systems. This paper presents FROG, a robot software platform built on a unified Cycle–Task execution model. An application is composed of Workers, each pairing one periodic Cycle with multiple request-driven Tasks. A Worker is also the unit of temporal isolation: the runtime binds one periodic and one request-driven context to every Worker, so continuous control and request handling never share a context, with no executor or thread-pool configuration by the developer. Three mechanisms distinguish FROG from comparable robot middleware: a per-Worker overrun policy selectable at run time, a shared-state store with lock-free reads and per-key atomic publication, and built-in timing telemetry queryable from a deployed robot. We validate FROG on a quadruped-based mobile manipulator performing welding, and characterize it: the mean release interval matches its nominal period to within 20 ns up to 1 kHz, a queued Task costs 2.2 µs above the protocol floor, a 1 kHz Cycle holds its rate under sustained load, and a Worker costs 37 KiB and 0.05 % of a core.

Keywords: Industrial Applications of Control, Process Control Systems, Control Theory and Applications
Abstract: The demand for precision dispensing automation of fine powders, which are widely utilized across various industries, is steadily increasing. However, the non-linear flow behavior induced by the high angle of repose (AoR) and cohesive forces of fine powders degrades control accuracy, hindering dispensing resolution. To address this challenge, this study proposes a vibration-based precision dispensing system and optimizes its operational vibration conditions. Using a robot arm integrated with a vibration module, dispensing experiments were conducted with respect to distance and PWM control signals. A multi-objective cost function was formulated to quantitatively evaluate linearity, productivity, and stability. The experimental results indicated that increasing vibration intensity collapsed the physical AoR, thereby enhancing fluidization; however, an excessive control signal conversely deteriorated flowability due to inter particle interference and bed wave deformation. Consequently, the optimal vibration condition that maximizes both system linearity and stability was derived, which will be utilized to establish precise powder dispensing control mechanisms in the future.

Keywords: Industrial Applications of Control, Control Theory and Applications, Sensors and Signal Processing
Abstract: This paper proposes an online inter-turn short-circuit (ITSC) fault diagnosis strategy for switched reluctance motors (SRMs) using an extended Kalman filter (EKF). An algebraically reduced dynamic model of SRMs with ITSC faults is established to resolve the mathematical singularities and numerical divergence issues inherent in conventional SRM models with ITSC faults. Integrated with the proposed model, an EKF-based ITSC fault diagnosis method is proposed to estimate the fault current using only phase current measurements. Simulation results verify that the proposed method successfully detects even incipient faults with superior sensitivity.

Keywords: Industrial Applications of Control, Artificial Intelligence Systems
Abstract: This paper presents a compact, low-cost robotic workcell for autonomous bottle packing, featuring a dual-mode end-effector (vacuum suction and parallel-jaw grasping), a YOLOv8 planar-homography localisation pipeline, and a ROS2 Behaviour Tree coordinator. Validated at the WRS2025 Manufacturing Challenge (Aichi, Japan), the system achieved 100% pick success in Run 2, a mean cycle time of 20.67 s/bottle, and hardware cost below 500 EUR excluding the robot arm.

Keywords: Industrial Applications of Control, Process Control Systems, Robotic Applications
Abstract: RGB-D-based vision models have been widely used as general-purpose perception methods in robotic manipulation tasks. However, in certain task environments, visual task recognition can be limited by strong illumination, dust, fumes, and other adverse conditions. This paper proposes a mesh-difference-based cutting path generation and motion planning method for robotic cutting tasks that does not rely on vision-based perception. The proposed framework extracts a cutting surface from the geometric differences between pre-cut and post-cut mesh models registered in the same coordinate system. The reference cutting path is then generated based on the robot-facing outer boundary of the extracted cutting-surface mesh. Subsequently, an executable TCP path is constructed from the reference path, enabling the robot end-effector to follow the target path. To execute the generated path, target poses are constructed by aligning the cutting-tool axis parallel to the approach direction, and cutting motion planning for the robotic manipulator is performed in the Isaac Sim environment using cuRobo-based motion generation and robot kinematics computation methods. In the simulation, the shape approximation performance of the generated cutting trajectories and the TCP position-tracking performance are evaluated using mesh models with various geometries. The results demonstrate that a reference cutting path can be generated from the geometric differences between pre-cut and post-cut meshes and converted into a TCP position path executable by the simulated robot.

Keywords: Industrial Applications of Control, Control Theory and Applications, Robot Mechanism and Control
Abstract: This paper summarizes a thruster-assisted stabilization method for floating offshore wind turbines under wind and wave disturbances. A four-degree-of-freedom nonlinear model describes platform heave, roll, pitch, and tower deflection, while a reduced three-degree-of-freedom model is used for controller design. A super-twisting sliding-mode controller is employed to regulate heave and pitch motions. Simulations show faster response attenuation and improved vibration suppression over proportional–integral control, particularly under an extreme sea state.

Keywords: Industrial Applications of Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: This paper addresses the multi-drone task scheduling problem in warehouse environments under a battery constraint. The problem is NP-hard because the limited battery couples task assignment, task ordering, and depot selection into a single decision. As a result, a feasibility check is required at every scheduling step, and the search space explodes as the number of tasks grows. To address this challenge, this paper proposes a cluster-based algorithm that groups connected tasks into cluster, each corresponding to exactly one battery trip. When a cluster is generated, the farthest depot is selected so that the cluster stays feasible under any depot. The algorithm consists of i) cluster generation, ii) greedy scheduling, and iii) depot optimization. We demonstrate the effectiveness of the proposed algorithm through a warehouse simulation.

Keywords: Industrial Applications of Control, Control Theory and Applications
Abstract: Extended-range electric vehicles (EREVs) require an efficient energy management strategy because the battery, electric motors, and engine-generator system are tightly coupled. Conventional optimization-based approaches generally optimize only the control strategy while assuming a fixed battery energy capacity, which may limit the achievable system efficiency. This paper proposes a simultaneous control co-design framework that jointly optimizes the battery energy capacity and energy management strategy. The optimization problem is formulated using a direct transcription approach, where the system states, control inputs, and battery energy capacity are optimized within a unified nonlinear programming framework. Simulation results show that the proposed framework achieves efficient power distribution while maintaining vehicle speed tracking performance and battery operating constraints. The proposed approach provides an effective framework for integrating battery sizing and energy management optimization in EREVs.

Keywords: Industrial Applications of Control, Process Control Systems, Control Theory and Applications
Abstract: Flexible in-process structures in additive manufacturing exhibit time-varying dynamics as their structural properties change with increasing build height. Such variations in dynamic characteristics can be represented as a linear parameter-varying (LPV) system using build height as a scheduling variable. However, independently bounding scheduling parameters that are physically coupled through structural height change can introduce unattainable parameter combinations and increase conservatism in robust controller synthesis. This work compares an independent box representation with a proposed parameter-varying representation that preserves physical dependence among the scheduling parameters. Robust state-feedback controllers are synthesized under the same common quadratic Lyapunov framework. Numerical results show that the proposed representation provides a larger feasible decay-rate bound and improved closed-loop stability along the parameter-varying trajectory, indicating reduced conservatism in robust control of growing structures during additive manufacturing.

Keywords: Information and Networking
Abstract: In recent information-driven society, protecting image data from eavesdropping and tampering is critical. Optical encryption, specifically Double Random Phase Encryption (DRPE), leverages the physical properties of light and the Fourier transform to secure data. However, DRPE typically suffers from high communication overhead because the security keys must be the same size as the original image. This paper proposes a method to significantly reduce data size by using a reduced key, only one-tenth of the original pixel count. By tiling this reduced key to match the original dimensions, we successfully decrease the data size by approximately 98% without compromising the quality of the decrypted image. To address security vulnerabilities—specifically periodicity introduced by tiling and the predictability of simple random generation—we integrate chaos theory via a two-dimensional logistic map with DRPE. We apply a diffusion process to the tiled phase mask using chaos parameters derived from the reduced key. This process randomly rearranges pixels, eliminating periodicity and enhancing key sensitivity. Ultimately, this proposed method establishes a robust encryption method that improves security and complexity while achieving a massive reduction in communication and management costs.

Keywords: Information and Networking, Artificial Intelligence Systems
Abstract: This paper studies a time-indexed task scheduling problem with worker-dependent processing times and subjective task utilities. Fairness is defined specifically as the balance of accumulated subjective utility among workers and does not include workload, effort, or fatigue. A Nash product based objective is approximated by a piecewise-linear representation of logarithmic utility to obtain a mixed-integer linear programming (MILP) formulation. To make the temporal role of scheduling explicit, precedence constraints and the maximum completion time C_{max} are incorporated. Experiments compare total utility, Nash welfare, and lexicographic max-min objectives, and evaluate the effects of precedence and completion-time weighting. Additional tests on four problem sizes show increasing computation time but relative optimality gaps below 10^{-4} in all tested instances. A breakpoint sensitivity test also confirms that a finer approximation substantially reduces the logarithmic approximation error.

Keywords: Information and Networking, Sensors and Signal Processing
Abstract: Enabling wireless communication for mobile robots in shipyards requires first understanding the wireless-link characteristics in environments dense with steel structures. This paper presents an on-site measurement study of sub-GHz WiFi HaLow (IEEE 802.11ah) links in two assembly shops of a commercial shipyard. Using a purpose-built HaLow measurement platform, we collected link quality—in terms of signal strength, throughput, and loss rate—versus distance under line-of-sight conditions and as a function of structural factors under steel-bulkhead conditions. We found that, under line-of-sight conditions, distance was the primary factor in link quality, whereas under steel-bulkhead conditions the presence or absence of a steel structure able to reflect the signal was associated with substantial differences in link quality even at similar received signal strength. At a position lacking a reflecting surface, where the link was formed essentially through steel penetration alone, the UDP loss rate rose to 7.82%, markedly higher than at adjacent positions of similar signal strength, while the auto-selected bandwidth stayed low. These observations indicate that link quality inside a shipyard can be associated not only with distance but also with the reflection/penetration geometry of steel structures, pointing to the need for infrastructure placement that exploits reflected paths and for link-adaptive communication.

Keywords: Information and Networking, Artificial Intelligence Systems, Sensors and Signal Processing
Abstract: Classical graph coloring models a conflict between two vertices by a fixed binary edge. This representation can be overly conservative when the actual conflict depends on multidimensional system states, interaction directions, and admissible tolerances. This paper introduces matrix-weighted graph coloring (MWGC), in which an edge-dependent positive semidefinite matrix and a pair of vertex states determine whether a potential edge becomes an active conflict. An MWGC-proper coloring is defined through a quadratic matrix detector. The MWGC chromatic number is shown to be equal to the classical chromatic number of the induced active conflict graph. The framework is then applied to a server assignment problem, where vertices represent computational tasks and colors represent servers. The equivalence result shows that the minimum number of required servers is determined by the chromatic number of the induced active conflict graph, allowing task pairs with admissible state-dependent interactions to share a server.

Keywords: Navigation, Guidance and Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: This paper presents an automatic identification system (AIS)-based vessel tracking framework designed to enhance shore-based remote situational awareness at remote operation centers in coastal environments. As maritime autonomous surface ships are increasingly expected to operate alongside conventionally crewed vessels, continuous and reliable vessel state estimation from a shore-based perspective becomes a critical operational requirement. AIS provides a scalable and widely deployed data source for this purpose; however, its practical utility is constrained by nonuniform transmission intervals and frequent signal loss, which pose significant challenges for maintaining track continuity. The proposed framework addresses these limitations through a filtering-based state estimation scheme that preserves track continuity beyond what can be achieved using direct AIS observations alone. Experimental results based on real-world AIS datasets collected from a representative coastal region demonstrate the practical feasibility of the proposed method.

Keywords: Navigation, Guidance and Control
Abstract: Safe navigation in multi-agent systems requires each agent to reach its goal while avoiding collisions with other agents and moving obstacles. Graph-CBF-based methods improve scalable safe control by combining control barrier functions with graph-based distributed policies, but they commonly construct obstacle nodes from immediate LiDAR observations. This representation can be insufficient in dynamic environments because current hit points alone do not indicate whether obstacles are approaching the agent. To address this limitation, this paper proposes a motion-aware LiDAR adapter based on an inflated unsafe invariant set for multi-agent safe navigation. The adapter uses recent LiDAR hit-point history, together with the agent and goal positions, to estimate ray-wise motion risk and bounded range inflation. The inflated hit points are then used as adapted obstacle nodes before graph construction, while the pretrained GCBF+ policy remains frozen. The proposed method is evaluated in moving-obstacle scenarios against nominal, centralized CBF, decentralized CBF, and graph-CBF baselines under Single Integrator, Double Integrator, and Dubins Car dynamics. Experimental results show improved safe rate and success rate across all dynamics, including dynamics not used during adapter training, while maintaining goal-reaching performance with a small safety–completion trade-off.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Sensors and Signal Processing
Abstract: Long-term autonomous driving requires LiDAR place recognition that remains robust under cross-day appearance changes while satisfying real-time runtime constraints. This paper presents CIRRUS-BEV (Cartesian Image Retrieval with Residual Update for Scan Localization in BEV), a lightweight Cartesian bird’s-eye-view (BEV) pipeline for cross-day LiDAR place recognition and local 2-D pose refinement. Unlike polar Scan Context-style descriptors, CIRRUS-BEV uses a unified six-channel Cartesian BEV representation for both fast global retrieval and metric residual refinement. The pipeline retrieves a database scan using a precomputed BEV descriptor, estimates an initial alignment by phase correlation, and predicts a bounded residual update from an 18-channel BEV pair tensor. Experiments on a Munji-campus cross-day protocol demonstrate that CIRRUS-BEV improves valid-positive top-1 recall at 1 m from 79.92% to 88.91% compared with Scan Context, while reducing online query runtime from 33.997 ms to 5.118 ms. The resulting 195.4 Hz online rate is approximately 6.6× faster than the measured Scan Context pipeline. These results show that a height-aware Cartesian BEV representation can support both accurate retrieval and efficient local refinement for real-time cross-day LiDAR localization.

Keywords: Navigation, Guidance and Control, Robot Vision, Sensors and Signal Processing
Abstract: Multi-camera visual-inertial odometry (VIO) improves pose estimation robustness by leveraging wide field-of-view coverage. However, most existing systems weight all cameras and observations equally during optimization. In degenerate indoor environments, geometrically unreliable views therefore retain full influence and cause drift. We propose AdaM-VIO, an adaptive two-level weighting method for robust multi-camera VIO. In the optimization back-end, AdaM-VIO rescales the information matrix of each reprojection factor using two weights: a camera-level weight derived from per-camera feature support and a depth-aware weight derived from landmark depth. Together, they reduce the influence of weakly supported views and close-range observations. We evaluate AdaM-VIO on the HILTI SLAM Challenge 2022 benchmark, where it achieves the highest total localization score among all compared visual-inertial systems.

Keywords: Navigation, Guidance and Control, Sensors and Signal Processing
Abstract: Radar-based odometry has attracted growing attention due to its robustness under adverse weather, degraded illumination, and geometrically degenerate environments. Accurate ego velocity estimation is important for radar-based odometry systems, as it provides motion constraints for state estimation. However, conventional Doppler-based methods assume that static points dominate the measured scan. This assumption may fail in highly dynamic environments, which can lead to substantial estimation degradation. To address this challenge, the temporal dynamic prior filter (TDPF), a static-conditioning framework that identifies and removes dynamic points by exploiting temporal priors before ego velocity estimation, is proposed, thereby recovering a static-dominant measurement set. Rather than classifying dynamic points independently in every frame, TDPF systematically accumulates dynamic points across frames and constructs a set of persistent spatial priors to guide dynamic removal on the current scan. Experimental results on the MSC-RAD4R Urban and NTU4DRadLM datasets show that TDPF improves ego velocity estimation accuracy and odometry precision on dynamic urban sequences.

Keywords: Navigation, Guidance and Control, Robot Vision, Robotic Applications
Abstract: Image-Based Visual Servoing (IBVS) provides an efficient vision-guided control paradigm for unmanned aerial vehicles (UAVs) by directly regulating image-space errors. However, conventional IBVS controllers are vulnerable to two critical issue: loss of closed-loop stability near the target due to input and state constraints, and control failure caused by intermittent loss of moment-based visual features under aggressive motion. To address these challenges, this paper proposes a terminal-constraint model predictive control (TC-MPC) framework for IBVS, integrating with a Kalman filter (KF)– based state-prediction mechanism. The TC-MPC explicitly incorporates terminal-state constraints and a terminal cost into the IBVS error dynamics, ensuring recursive feasibility, improved convergence behavior, and closed-loop stability under control and state constraints. In parallel, the Kalman filter predicts the temporal evolution of image moments during short-term visual degradation, enabling the controller to preserve control continuity when moment measurements are partially unavailable. The proposed approach is validated through real-time UAV visual servoing experiments.

Keywords: Navigation, Guidance and Control, Sensors and Signal Processing
Abstract: This paper presents an aerodynamic angle estimation framework for launch vehicles under weak observability conditions. Aerodynamic angles, such as the angle of attack and sideslip angle, are important quantities for aerodynamic load reduction and flight performance evaluation. However, because GPS/INS measurements provide ground velocity rather than air-relative velocity, the air-relative velocity required for aerodynamic angle information cannot be directly obtained in the presence of wind. This problem becomes more pronounced for launch vehicles, which generally operate near a nominal trajectory with limited attitude excitation over many flight segments. As a result, the ground velocity measurement alone provides insufficient information to clearly separate the air-relative velocity from the wind compo- nent, leading to weak observability in aerodynamic angle estimation. To mitigate this estimation problem, prior wind information is introduced as a nominal wind profile. The nominal wind reduces the ground velocity measurement residual by accounting for the dominant wind contribution in advance. The remaining wind mismatch is modeled as a residual wind uncertainty and treated as a nuisance state. Accordingly, a Schmidt-Kalman filter update is applied to propagate this residual wind uncertainty while limiting its direct influence on the correction of the air-relative velocity. This nuisance- state treatment improves estimation robustness under weak observability by reducing erroneous corrections that may be introduced when a weakly observable state is directly estimated. The numerical simulation results demonstrate that the proposed method improves aerodynamic angle estimation performance under weak observability conditions.

Keywords: Navigation, Guidance and Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: Side-scan sonar (SSS) has emerged as a promising sensing modality for underwater simultaneous localization and mapping (SLAM) because of its long-range sensing capability and wide-area seabed coverage. In landmark-based SSS SLAM, salient seabed features are typically represented using slant-range and azimuth observations derived from acoustic measurements. However, due to the sensing geometry of SSS, landmark elevation remains only weakly constrained by the measurements. This limitation, referred to in this work as elevation degeneracy in SSS SLAM, can introduce estimation inconsistency when conventional full 6-DOF formulations are directly adopted. This study investigates how landmark elevation uncertainty influences state estimation in landmark-based SSS SLAM under elevation degeneracy. In particular, Jacobian analysis of the slant-range and azimuth observation model is performed to examine how native SSS measurements contribute to each component of the vehicle state and influence the overall estimation behavior. Based on this analysis, an effective vehicle state representation is identified according to the observability characteristics imposed by the sensing geometry. The effectiveness of the resulting reduced 4-DOF state representation, which excludes roll and pitch, is validated through comparative physics-based underwater simulation experiments against a conventional 6-DOF formulation.

Keywords: Navigation, Guidance and Control, Robotic Applications, Sensors and Signal Processing
Abstract: Poultry farm corridors are narrow and contain repeatedly arranged feeding and watering structures, making stable path perception and localization difficult for mobile robots. In addition, GNSS reception is limited inside poultry houses, so auxiliary position information is required to compensate for the accumulated error of LiDAR-IMU-based localization. In this study, a mobile robot system for poultry farm corridor navigation was implemented. The proposed system generates a corridor reference and a driving target point using a depth camera and LiDAR point cloud, and estimates the robot trajectory using LiDAR-IMU-based localization and UWB position information together. Quantitative evaluation was performed through line-tracing output stability evaluation and localization evaluation in an outdoor terrain validation environment, and field applicability was then confirmed by operating the robot in an actual poultry farm corridor. The experimental results showed that line-tracing achieved a target generation rate of more than 96%, and the proposed algorithm reduced the localization error by approximately 60.2% in terms of ATE RMSE compared with the baseline.

Keywords: Navigation, Guidance and Control, Human-Robot Interaction, Artificial Intelligence Systems
Abstract: Autonomy is most useful when it knows its limits. As artificial intelligence (AI) systems take on more open-ended decisions, their reliability depends on knowing when to act alone and when to seek guidance. We study this question in zero-shot object-goal navigation, where vision-language frontier maps can search for unseen object categories but become unreliable when multiple frontiers receive nearly tied semantic scores. We propose an ambiguity-triggered consultation layer that preserves the autonomous navigation pipeline by default and asks a human advisor only at uncertain frontier decisions. The advisor is constrained to select from the robot’s own candidate frontiers, making consultation a local correction rather than a replacement for navigation. On Habitat-Matterport 3D ObjectNav, ambiguous frontier decisions are common and associated with unreliable autonomous choices. Our analysis shows that ambiguity-triggered consultation improves success from 55.0% to 70.0%. On the full 2,000-episode validation benchmark, the map-agnostic and map-aware advisors reach 63.4% and 65.8% success, respectively, compared with 52.5% for VLFM.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Robotic Applications
Abstract: Autonomous navigation of uncrewed aerial vehicles (UAVs) in cluttered environments remains an open challenge, particularly due to the local minima problem: concave obstacle geometries and sequential trap configurations routinely cause planning failure in reactive schemes. This paper presents a fuzzy logic (FL) supervisory layer to arbitrate in real time among three planning strategies with complementary strengths — artificial potential fields (APF), wall-following (WF), and virtual obstacles (VO) – functioning as a decision-making brain that selects the most suitable planning arm for each encountered scenario. Prior reactive planners address local minima by modulating the force vector within a single strategy; operating instead at the strategy-selection level, the proposed architecture switches among architecturally distinct planners, enabling robust escape from multiple sequential traps and preventing cyclic re-entrapment. A novel geometric switching indicator is developed that governs smooth, loop-free inter-strategy transitions. An ablation study evaluates the incremental integration of each escape mechanism arbitrated by the FL supervisor. Monte Carlo simulations across five purpose-built environments — featuring chained concave geometries of substantially greater complexity than standard benchmarks — demonstrate a significantly higher success rate without sacrificing path efficiency or safety.

Keywords: Navigation, Guidance and Control, Industrial Applications of Control, Control Theory and Applications
Abstract: This paper presents an autopilot-aware proportional navigation gain-scheduling framework for short-range missiles engaging maneuvering targets. The classical Proportional Navigation Guidance (PNG) law is re-designed under joint time- and frequency-domain constraints to enhance interception accuracy while preserving closed-loop robustness. Unlike conventional cascaded configurations, the proposed method models the guidance and autopilot subsystems as a single closed-loop structure, enabling coordinated dynamic behavior. A previously designed inner-loop Linear Quadratic Integrator (LQI) autopilot is employed as the acceleration tracking system, providing a stable and well-damped baseline for the proposed framework. The navigation constant N is reformulated as a tunable control gain and optimized under joint overshoot and stability-margin constraints derived from the closed-loop guidance–control dynamics, enabling gain scheduling that balances agility and robustness throughout the engagement. Within this unified framework, Monte Carlo simulations including aerodynamic, propulsion, sensor, and environmental uncertainties, as well as actuator constraints, are conducted to evaluate the robustness of the proposed architecture. The results show that the resulting closed-loop guidance–control structure provides improved terminal accuracy and smoother control effort during moving-target engagements compared with conventional fixed-gain PNG configurations, demonstrating the effectiveness of time–frequency constrained navigation gain scheduling for short-range missile scenarios.

Keywords: Navigation, Guidance and Control, Robotic Applications, Robot Mechanism and Control
Abstract: Dynamic Window Approach (DWA) is widely used for local path planning of mobile robots. However, conventional DWA relies on fixed velocity-space sampling and fixed parameter settings. This can cause unnecessary computation in safe velocity regions. It can also reduce candidate resolution in complex environments. In addition, fixed parameters limit adaptation to changing driving conditions. This paper proposes Adaptive Focus-Band (AFB)-DWA to mitigate these limitations. The proposed method consists of two components. First, a Focus-Band (FB) sampling scheme augments candidate velocities near collision-candidate boundaries. Second, the collision-candidate ratio quantifies the complexity of the surrounding environment. AFB-DWA then adjusts the prediction horizon, maximum considered distance, and obstacle-avoidance weight according to this ratio. Simulation and real-robot experiments show that AFB-DWA improves goal-reaching reliability, travel efficiency, and trajectory smoothness compared with the baseline DWA.

Keywords: Navigation, Guidance and Control, Robot Vision, Robotic Applications
Abstract: Object-centric 3D Gaussian Splatting (3DGS) depends strongly on the initial posed images available before optimization. This extended abstract presents a mesh-guided viewpoint initialization method for robotic photometric reconstruction when the target object and its approximate geometry are known. A multi-shell Fibonacci prior generates robot-feasible candidate viewpoints around the target mesh, and a reconstruction-oriented utility selects a compact image-pose set by balancing mesh coverage, surface frontality, viewing-direction diversity, top-view balance, and local view overlap. In ROS/Gazebo experiments on engine, cat, and cow object models, the proposed initialization improves mean SSIM and PSNR, consistently reduces LPIPS, and reduces exploration CPU time compared with a SWAP-style baseline under the same viewpoint budget.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Artificial Intelligence Systems
Abstract: This paper presents a deep reinforcement learning-based integrated guidance and control (IGC) method for missile target interception. The proposed method uses infrared (IR) image information. Conventional missile systems usually separate guidance and autopilot control. This structure is simple, but it can reduce performance when missile dynamics, target motion, and sensor uncertainty are strongly coupled. To solve this problem, the proposed method learns one end-to-end policy based on soft actor-critic (SAC). The policy uses guidance-control states and visual features from synthetic IR images. In addition, an action-smoothing term is added to the actor loss. This term reduces control chattering without using an external command filter during inference. Simulation results show that the proposed IR-enhanced SAC policy improves success rate, terminal accuracy, and control smoothness. The results also show that perception-aware DRL is feasible for integrated missile guidance and control.

Keywords: Navigation, Guidance and Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: Large-scale drone swarm systems require safe, smooth, and synchronized 3D formation transitions for thousands of unmanned aerial vehicles (UAVs). However, multiple consecutive formation transitions can result in uneven cumulative flight distances among UAVs, and storing dense trajectories for large swarms can impose substantial data overhead. This paper proposes an integrated framework that combines workload-balanced task allocation, acceleration-bounded ORCA-based trajectory generation, and compact Bézier trajectory fitting. The proposed method minimizes the maximum cumulative travel distance over consecutive formation transitions to balance workload among UAVs. Based on the balanced assignment, synchronized and collision-free trajectories are generated under velocity and acceleration constraints. The dense trajectories are then converted into compact piecewise Bézier representations while preserving trajectory accuracy, continuity, and dynamic feasibility. Simulation results with up to 10,000 drones demonstrate improved workload balance, collision-free transitions, and more than 95% reduction in trajectory data size, confirming the scalability and effectiveness of the proposed framework.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Autonomous Vehicle Systems
Abstract: Wave-induced disturbances can degrade the station-keeping performance of an unmanned surface vehicle (USV) by causing deviations from the target position. Conventional position-feedback station-keeping control generates corrective inputs from the resulting errors, making its response inherently reactive after the disturbances have appeared as vehicle motion. To overcome this limitation, this paper proposes a force-sensing-integrated hybrid feedback station-keeping control method for a USV equipped with artificial lateral line system based pressure sensor arrays. The proposed method uses hull-side pressure variations as a force-sensing feedback source related to wave-induced lateral loading. Pressure sensor arrays installed on both sides of the hull measure asymmetric pressure variations under lateral wave disturbances, and the left--right pressure imbalance is converted into a lateral compensation term in the sway control input. Tank-based experiments were conducted using an artificial wave generation system to provide repeatable lateral disturbance conditions. Compared with position-feedback control (PFC), force-sensing-integrated hybrid feedback station-keeping control (HFC) was observed to reduce the RMS position variation by 56% in the direction parallel to the applied wave disturbance and by 62% in the perpendicular direction.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Control Theory and Applications
Abstract: This paper proposes a robust Risk-aware Model Predictive Path Integral (MPPI) control framework that quantifies state-space uncertainty online using a Sparse Gaussian Process (SGP). The framework uses the SGP-predicted mean to correct nominal bicycle dynamics and utilizes the predictive variance to dynamically scale disturbance covariance for risk propagation. Validated on an F1TENTH platform in both Gazebo simulations and real-world tracks, the proposed method significantly improves safety and trajectory tracking performance over baseline MPPI techniques without compromising real-time feasibility.

Keywords: Navigation, Guidance and Control, Robotic Applications, Autonomous Vehicle Systems
Abstract: Accurate state estimation is a fundamental requirement for reliable navigation and control of unmanned surface vehicles (USVs). Since the planar motion of a USV includes both translational position and heading, representing the pose in a conventional Euclidean state space may lead to inconsistent error computation, particularly for angular states. This paper presents an SE(2)-based state estimation method for USVs using dual-antenna GNSS and AHRS. The USV pose, consisting of position and heading, is represented as an element of SE(2), while the forward speed and yaw rate are included as additional motion states. The proposed estimator propagates the pose on SE(2) through the exponential map and computes the innovation in the corresponding Lie algebra using the logarithmic map. The measurement vector is constructed from state- corresponding sensor outputs, where the dual-antenna GNSS provides position, heading, and speed information, and the AHRS provides yaw-rate information. By performing the prediction and correction processes on the manifold, the proposed method handles pose-related errors in a geometrically consistent manner. Simulation and experimental evaluations are conducted to verify the feasibility of the proposed estimator for planar USV navigation. The results show that the SE(2)-based formulation provides stable state estimation performance for position, forward speed, and heading states.

Keywords: Sensors and Signal Processing
Abstract: As the demand for large-scale training data grows to enhance computer vision models, conventional methods such as manipulation-based augmentation or generative (artificial intelligence) AI synthesis face limitations, including corrupted physical consistency and the generation of optical hallucinations. To address these challenges, this paper proposes a physics-based data augmentation technique utilizing integral imaging technology. The proposed method interprets multi-view elemental images as spatial-angular sequential data and synthesizes high-resolution images from virtual viewpoints using the Virtual Focal Point with Beams (VFPB) model. Experimental results demonstrate that the augmented images achieve high quality, with an average peak signal-to-noise ratio (PSNR) of 38 and structural similarity index measure (SSIM) of 0.98 compared to the original reference. Furthermore, by precisely adjusting optical parameters such as the virtual focal distance and beam size, the generated datasets can flexibly inherit and expand the visual characteristics of the original multi-view images without requiring additional real-world data collection or complex computational rendering. This method presents a new physics-based learning paradigm that provides high-quality customized data optimized for domain-specific AI training, which is often difficult to address with general-purpose datasets.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Robotic Applications
Abstract: This paper proposes a Double-loop Sliding Mode Controller (SMC) integrating kinematics and motor dynamics for high-precision path tracking of a differential drive mobile robot. The upper-level kinematics-based SMC computes reference velocities to minimize tracking errors, while the lower-level motor-dynamics-based SMC generates torque commands to compensate for actuator uncertainties. This hierarchical structure effectively resolves dynamic disturbances from a kinematic perspective without requiring full rigid-body modeling. The proposed controller was validated through high-fidelity 3D simulations in NVIDIA Isaac Sim. Results demonstrate that the Double-loop SMC achieves a steady-state position tracking RMSE of 0.0034 m and a motor angular velocity RMSE of 0.0045 rad/s, significantly outperforming a conventional PID controller (position RMSE: 0.0219 m). Under intentional external disturbances, the proposed SMC achieved a position tracking RMSE of 0.0004 m, compared with 0.0195 m for the PID controller.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Robotic Applications
Abstract: This study proposes a Perception-Uncertainty-Aware D* Lite path planning method that incorporates perception confidence into both the magnitude and spatial spread of a Gaussian cost function within a Multi-Layer Costmap. The static environment is represented in the L1 layer, whereas the positional uncertainty of dynamic obstacles is modeled in the L2 layer. The confidence-dependent cost is then integrated with D* Lite to enable real-time incremental replanning. The proposed method is validated in MATLAB through two scenarios. In Scenario A, the planned path varied according to the perception confidence assigned to a single obstacle, confirming that confidence affects both the magnitude and the spatial extent of the cost function. In Scenario B, with five dynamic obstacles undergoing random-walk motion, the vehicle reached the goal without collision. Compared with the conventional method, the proposed method achieved a larger average obstacle-vehicle clearance for all obstacles, while D* Lite efficiently replanned the path by recomputing only the cells affected by cost changes.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Autonomous Vehicle Systems
Abstract: Safe coordination of multiple unmanned aerial vehicles (UAVs) under communication delays is challenging because each vehicle must avoid collisions using uncertain neighbor information. This paper presents an uncertainty-aware safety-critical control framework based on a high-order control barrier function (HOCBF). Each UAV predicts neighboring states from delayed communication packets and uses a data-driven uncertainty bound to inflate the pairwise safety constraint. This yields a conservative robust safety filter that maintains collision avoidance under prediction uncertainty. To reduce unnecessary conservatism, a consensus-based safety filtering structure is introduced, where pairwise safety responsibility is distributed through continuous safety-burden variables. The allocation is updated in a distributed manner using a consensus alternating direction method of multipliers (C-ADMM) procedure. Simulation results on an 8-UAV scenario demonstrate a 23.9% faster task completion compared with the robust safety filter baseline, while maintaining all safety constraints.

Keywords: Navigation, Guidance and Control, Sensors and Signal Processing, Control Theory and Applications
Abstract: This paper proposes an angular velocity estimation method for reaction-wheel-based satellites under gyroscope fault conditions. In small satellites, high-performance gyroscopes are expensive, and redundant sensor configurations are often difficult to implement due to mass, volume, and cost constraints. These limitations motivate the development of an alternative approach that can provide reliable angular velocity information even when gyroscope measurements become unavailable. To address this issue, the proposed method utilizes Hall sensor measurements from the reaction wheels and combines them with an Unscented Kalman Filter (UKF) to estimate the satellite angular velocity. By exploiting the relationship between reaction wheel information and satellite attitude dynamics, the proposed approach can provide the state information required for attitude control without relying solely on gyroscope measurements. In this sense, the method can serve not only as a fault-tolerant solution under gyroscope failure conditions but also as a potential estimation framework for gyroless satellite systems. Simulation results demonstrate that the proposed method can effectively estimate angular velocity under gyroscope fault conditions. These results indicate that the proposed approach has the potential to improve the fault tolerance and operational reliability of reaction-wheel-based small satellites.

Keywords: Navigation, Guidance and Control, Robot Vision, Artificial Intelligence Systems
Abstract: At least 2.2 billion people worldwide live with near or distance visual impairments, motivating the development of portable assistive navigation systems that can extend sensing capability beyond the physical reach of a cane without the additional hardware, calibration, and synchronization requirements of multi-sensor platforms. This paper presents a real-time obstacle-avoidance and local path planning system for people with visual impairments that relies on a single monocular camera. Per-pixel relative depth is estimated using Depth Anything V2-Small and converted into an occupancy grid. A three-stage classifier then filters to classify obstacles, free space, and floor cells while reducing isolated false positives. The grid is projected into a bird’s-eye view, and an A* planner with a straight-ahead-priority cost computes a stable route that is conveyed through a spatial-audio interface. Each grid is computed independently per frame, without pose estimation, temporal fusion, or a global map. Thus, planning is local to the current view. On 100 manually annotated Mapillary Vistas frames, the proposed grid achieves a precision of 0.9792, a recall of 0.8436, an F2 score of 0.8676, and an IoU of 0.8287 for near-field obstacle cells. On 976 consecutive frame pairs drawn from 24 egocentric walking clips of a public first-person pedestrian video dataset, the grid achieves a mean motion-compensated temporal consistency of 0.9043. The full pipeline runs at ~30 FPS on an NVIDIA RTX 4050 Laptop GPU, with the local path planning stage capped at 10 FPS so that each ~1 Hz audio cycle latches a stable, clearly communicable route.

Keywords: Navigation, Guidance and Control, Sensors and Signal Processing, Robotic Applications
Abstract: FAST-LIO2 has been widely adopted in various 3D LiDAR-based SLAM applications owing to its high accuracy and real-time performance. However, most existing studies have considered platforms with LiDAR sensors mounted at sufficient heights to ensure a wide field of view. In low-profile mobile robots, such as robotic vacuum cleaners, LiDAR sensors are often mounted close to the ground, the observed point clouds tend to be concentrated near floor-level features, which may contribute to limited geometric diversity. As a result, FAST-LIO2 becomes more susceptible to degeneracy under such configurations. Existing degeneracy analyses have mainly relied on the eigenvalue structure of the information matrix. However, this approach reflects only geometric observability and may not fully capture the accumulated uncertainty of the filter state. To address this limitation, this paper proposes the simultaneous use of condition number and position covariance as degeneracy indicators. The condition number reflects geometric observability, whereas position covariance represents accumulated state uncertainty within the filter. Experimental results show that both indicators remain stable in normal sequences, whereas at least one changes significantly in degraded situations. This behavior confirms the effectiveness of the proposed indicators for degeneracy identification in FAST-LIO2 with low-mounted 3D LiDAR.

Keywords: Navigation, Guidance and Control, Control Theory and Applications, Autonomous Vehicle Systems
Abstract: This paper proposes a probabilistic safety filter based on control barrier functions (CBFs) for stochastic nonlinear systems under non-Gaussian disturbances. Instead of identifying the full dynamics or the disturbance distribution, the proposed method learns the one-step transition of a sampled safety coordinate defined from a physical barrier. A random Fourier feature (RFF) representation motivated by kernel mean embedding (KME) and a bilinear Koopman surrogate are used to obtain an affine-in-control safety constraint. To account for unsafe overestimation by the learned model, a one-sided residual margin is computed from independent calibration transitions and inserted into the CBF condition. The resulting residual-calibrated constraint is implemented as an online quadratic program (QP) safety filter with input bounds. Simulation results on adaptive cruise control and Dubins car obstacle avoidance under Student-t disturbances show that the proposed method reduces observed safety violations compared with Gaussian chance-constrained CBF, while maintaining empirical performance comparable to bounded-disturbance Robust CBF.

Keywords: Navigation, Guidance and Control, Robotic Applications, Robot Vision
Abstract: Survey-grade terrestrial laser scanners produce dense, colorized point clouds but require the platform to remain stationary for minutes at each scan pose and provide no real-time registration, so conventional continuous-motion frontier exploration does not apply. This paper presents an automated stop-and-scan mapping subsystem that couples a frontier-exploration layer with a stop-and-scan sequencer through a distance gate and a coverage-based scan/skip decision. A multi-objective frontier cost with a bounded-horizon visit-order optimization reduces redundant travel between costly scan stops, while a prior-scan coverage radius suppresses redundant stationary acquisitions in already-covered space. In a physics-based simulation built from a real scan of a laboratory testroom, the proposed coverage-aware trigger maps the room with three stationary scans versus twenty-one for a distance-traveled baseline—a sevenfold reduction in expensive acquisitions—while holding the free-space map IoU at 0.965 and reaching 91.0% room coverage. The system is also verified against a real BLK360 scanner over a wireless link in a hybrid configuration.

Keywords: Navigation, Guidance and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: This paper presents a feasibility study of failure-aware LLM-DWA replanning for mobile robot navigation in dynamic obstacle environments. Existing LLM-DWA navigation has shown promise in static maze-like settings, but moving obstacles can invalidate initially generated waypoints and cause local planners to oscillate, time out, or fall into local-minimum states. To address this problem, we propose an event-triggered replanning strategy that monitors progress stagnation, low-speed behavior, and repeated recovery actions, while using a replanning cooldown to avoid excessive LLM calls. New LLM-generated waypoints are requested only when persistent local navigation failure is detected. The proposed method is evaluated in a dynamic maze with three moving obstacles and compared with a classical ROS-based navigation baseline, one-shot LLM-DWA, and fixed-periodic LLM-DWA. The baselines exhibit distinct failure modes, including start-stage planning collapse, stale waypoint execution, excessive replanning, and delayed timeout behavior. In contrast, the proposed strategy achieves the highest observed success rate of 5/10. The best run reached the goal with a 71.552 m path length, 436.315 s navigation time, three LLM calls, and two replanning events. These results indicate that selective failure-aware replanning can extend LLM-DWA navigation toward dynamic obstacle scenarios, although robust success in all trials remains future work.

Keywords: Navigation, Guidance and Control
Abstract: Traditional RTK technology has inaccurate positioning accuracy in environments with poor GNSS signals, and MEMS-IMU magnetometers are susceptible to electromagnetic interference. To overcome these limitations, we developed the dual-GNSS RTK/MEMS-IMU fusion positioning algorithm. This algorithm is implemented with a 15-dimensional loosely coupled Extended Kalman Filter (EKF). The EKF models sensor biases as first-order Gauss-Markov processes and incorporates updates: magnetometer yaw every 0.5 s, GNSS-RTK position/velocity every 0.2 s, and Dual-GNSS RTK yaw every 1 s. A dedicated multi-sensor device featuring a Unicore UM982 module and an Xsens MTi-1 sensor was fabricated for validation. Experimental results from a straight driving trajectory demonstrate centimeter-level precision, yielding a horizontal Root Mean Square Error (RMSE) of 0.03 m and a vertical RMSE of 0.01 m.

Keywords: Navigation, Guidance and Control, Autonomous Vehicle Systems, Sensors and Signal Processing
Abstract: This paper presents the real-time implementation and field validation of a dead reckoning (DR) method for unmanned surface vehicles (USVs) that relies solely on a Doppler velocity log (DVL) under GNSS-denied conditions. GNSS is used only during an initial 60-s window, in which the heading offset is estimated as the circular mean of the differences between the compass heading and the GNSS course; thereafter, the position is propagated from the DVL bottom-track velocity and the corrected heading. The onboard Robot Operating System (ROS) software performs validity checking, zero-order hold, and median-absolute-deviation (MAD)-based outlier rejection in real time. Across nine field datasets (0.44–1.75 km) acquired in shallow water near Jebu Island, Korea, the filtered DR reduced the position error relative to the raw integration in every dataset, with a final error of 1.1–11.3% of the distance traveled. In the worst case the estimated trajectory was rotated as a whole, indicating that the residual error is dominated by compass heading error.

Keywords: Navigation, Guidance and Control
Abstract: Visual-inertial odometry (VIO) provides accurate localization for unmanned aerial vehicles (UAVs), but its performance can degrade significantly when visual measurements become unreliable because of camera blockage, featureless scenes, illumination changes, or motion blur. This paper presents a prediction-based VIO framework that improves UAV localization during temporary visual degradation. The proposed method integrates a motion prediction prior into the backend optimization of VINS-Fusion. A visual degradation detection module evaluates the stability of feature tracking, the number of tracked features, and variations in pixel intensity to classify the current visual condition as either severe or partial degradation. According to the detected degradation level, the proposed framework adaptively adjusts the contribution of visual and prediction constraints. The conventional VINS-Fusion optimization consists primarily of marginalization, inertial measurement unit (IMU), and visual residual terms. In the proposed method, an additional prediction residual term is introduced to support localization when visual constraints become weak or unavailable. The prediction module estimates future accelerations from recent inertial measurements and uses them to predict the corresponding position and velocity states. The proposed framework is evaluated on five EuRoC sequences under four-second visual blackout conditions and compared with ORB-SLAM3 and VINS-Fusion. Experimental results demonstrate that the proposed method reduces localization error and maintains more stable trajectory estimation during visual degradation.

Keywords: Process Control Systems, Industrial Applications of Control, Information and Networking
Abstract: This study addresses the berth allocation and crane assignment problem (BACAP) in automated container terminals. BACAP jointly determines the berthing position, berthing time, and quay crane assignment for vessels. Frequent disruptions, such as vessel delays and unscheduled arrivals, can invalidate existing schedules. A feasible revised plan must therefore be generated within minutes for replanning. To address this requirement, this study proposes a feasibility-boundary heuristic based on an optimization-to-decision transformation. The method poses a simple Yes/No question: whether a target number of vessels can be assigned within a given waiting-time bound while satisfying constraints on berthing position, berthing time, and crane resources. This structure narrows the search space to operationally feasible candidates aligned with the target assignment level. This allows the method to derive a high-quality revised schedule with reduced vessel waiting time and fewer required time of departure (RTD) violations. Experiments using real port data show that the proposed method maintains a high assignment rate, short computation time, and scalable performance, while remaining close to a Gurobi-based mixed-integer linear programming (MILP) reference on small instances. These results demonstrate that the proposed heuristic provides a fast and reliable approach to berth and crane replanning under operational disruptions.

Keywords: Process Control Systems, Industrial Applications of Control, Robotic Applications
Abstract: Precision surface measurement is a key quality control process in manufacturing, used to verify whether curved components meet their design intent. Conventional approaches, such as manual inspection and fixed metrology equipment are limited in repeatability and automation, or require dedicated jigs and setup procedures whenever the target geometry or pose changes, reducing their flexibility for small-batch, high-mix production. Robot-based automated approaches have been proposed to address these issues but typically rely on predefined inspection surfaces or CAD models, making it difficult to apply when the target pose varies or prior geometric information is insufficient. In this study, we propose a multimodal surface measurement system that combines two RGB-D cameras, a six-degree-of-freedom robot arm, and a laser displacement sensor to integrate global perception with local precision measurement. The system first reconstructs the coarse three-dimensional geometry using two RGB-D cameras, and then performs local precision measurement based on Gaussian process adaptive sampling. To evaluate the feasibility of the system, a pilot test was conducted on a dome-shaped specimen with positive double curvature. The experimental results show that the proposed system achieved a mean absolute distance of 0.173 ± 0.018 mm with respect to the raster-scanned reference surface using only 35.3 sampling points on average. In future work, RGB-D depth data will be directly incorporated into the Gaussian process model further improving measurement efficiency.

Keywords: Process Control Systems, Multimedia Systems, Robotic Applications
Abstract: 3D Gaussian Splatting(3DGS) has recently attracted significant attention in the fields of robotics and simulation due to its ability to provide high-quality real-time rendering. However, conventional 3DGS reconstructs objects only from observed regions, resulting in limitations where unobserved areas, such as the rear or bottom surfaces of objects, remain incomplete or hollow. These structural deficiencies cause missing regions to become visible when objects are moved or rotated. The purpose of this study is to express an object that is structurally complemented in a simulator environment based on 3DGS. To this end, we propose an object-centered 3DGS complementary pipeline. The proposed method first performs segmentation on objects in the scene and then inpainting to generate surface information for the non-observed area. After that, the generated Gaussian model and the object Gaussian model are matched to form a more complete three-dimensional structure. As a result of the experiment, the proposed method effectively alleviated the hole and structural defect problem in the existing 3DGS when moving objects, and confirmed that the real-time rendering performance and visual quality of the original 3DGS can also be maintained.

Keywords: Process Control Systems, Robot Mechanism and Control, Artificial Intelligence Systems
Abstract: This paper presents a multibody co-simulation-based reinforcement learning framework for vibration-aware transfer motion of a linear motor stage. A MATLAB/Simulink learning agent is coupled with a RecurDyn multibody dynamics plant through the GRDClient interface, enabling synchronized exchange of control inputs and physical states. The agent generates jerk commands that are integrated into acceleration commands and applied to the plant as a driving force. The intended motion objective is to achieve fast transfer while considering a prescribed limit on the base-frame velocity. In the present implementation, violations of this limit are handled through a fixed-weight reward penalty rather than a hard constraint mechanism, and strict constraint satisfaction is therefore not guaranteed. Preliminary simulation results show that the trained policy reaches the target position; however, the base-velocity limit is exceeded during the transient motion and the response subsequently decays within the prescribed band. Accordingly, the results are interpreted as a preliminary demonstration of the proposed co-simulation learning framework rather than as validation of a strictly constraint-satisfying controller. The framework provides a simulation-based basis for further development of explicit constraint-handling strategies and experimental validation.

Keywords: Rehabilitation Robot, Biomedical Instruments and Systems, Control Theory and Applications
Abstract: This study biomechanically analyzed gait efficiency fluctuations and long-term changes in walking functions and maximum step length (MSL) in seven community-dwelling patients with post-stroke hemiplegia. Participants received Lower Extremity Constraint-induced Movement Therapy (LE-CIMT) using a piston device for leg joint treatment. Evaluations, including Fugl-Meyer Assessment (FMA), 10-Meter Walking Test (10MWT), Time Up and Go Test (TUG), and MSL, were conducted at baseline, pre-intervention, post-intervention (2 weeks), and follow-up (16 weeks). Three-dimensional gait analysis at baseline and post-intervention calculated joint angles, moments, and strength. Results showed significant post-intervention increases in hip extension angle, ankle dorsiflexion angle, and ankle plantar flexion moment/strength on the paralyzed side, indicating improved propulsive force. Consequently, 10MWT and TUG scores improved, with effects lasting at least 16 weeks. Notably, the discrepancy between predicted and measured MSL values increased after intervention. In conclusion, LE-CIMT using a piston device contributes to long-term walking function improvement, though a delayed recognition of functional gains may occur.

Keywords: Rehabilitation Robot, Sensors and Signal Processing, Human-Robot Interaction
Abstract: With the rapid aging of society, gait assessment has become increasingly important for rehabilitation and fall prevention. Conventional gait analysis methods, such as motion capture systems and RGB-D cameras, can provide detailed motion information but often require expensive equipment and dedicated environments. To address these limitations, we have developed an Internet of Robotic Things (IoRT) care-walker equipped with load cells, inertial measurement units, and rotary encoders. The system detects gait events, including heel contact and toe-off, and calculates gait parameters such as stride length and gait cycle.　This study proposes a method for estimating gait posture using only sensor data obtained from the IoRT care-walker. A human digital model is constructed from gait events, gait parameters, upper-limb load information, and walker position data. Upper-body and foot positions are estimated from these measurements, and joint positions are calculated using a human link model with geometric constraints. The proposed framework is also integrated with an XR-based visualization system to realize a gait digital twin. This paper describes the system architecture and the method for constructing the human digital model.

Keywords: Rehabilitation Robot, Robot Mechanism and Control, Robotic Applications
Abstract: Maximizing locomotor recovery requires a strategic focus on training parameters such as the intensity and variability of stepping practice. This study investigated the regulation of stepping speeds and neuromuscular exertion through a task-specific paradigm integrating audio-visual cueing, pivoting neuromuscular control, and eccentric demands. Twenty-three participants performed stepping tasks under three distinct conditions: (1) Muscle Power (MP) mode, guided stepping with audio-visual cueing at 100% and 120% of their baseline speed; (2) Neuromuscular Control (NC) mode, which integrates transverse-plane pivoting control into the MP condition; and (3) Eccentric Control (EC) mode, involving stepping under additional biomechanical loading (e.g., squat posture). Significant negative correlations were observed between the 10-Meter Walk Test (10MWT) and stepping speeds (r = -0.65 to -0.68, p < 0.05), demonstrating task-specific associations with overground walking ability. Furthermore, the audio-visual cueing precisely modulated stepping speeds between target levels (p < 0.05). Interestingly, the EC condition elicited significant speed increases that lacked correlation with the 10MWT, appearing to reflect condition-specific speed regulation in response to increased eccentric demands rather than functional walking performance. Ultimately, these findings demonstrate that this comprehensive paradigm provides the essential intensity and task variability required to safely modulate neuromuscular exertion, highlighting its potential for task-specific robotic rehabilitation.

Keywords: Rehabilitation Robot, Sensors and Signal Processing, Artificial Intelligence Systems
Abstract: Proprioceptive Neuromuscular Facilitation (PNF) upper limb pattern is widely used to improve activities of daily living after a stroke. However, it has complex movement patterns; thereby, novice-led sessions may result in inappropriate posture control or resistance application. Existing rehabilitation monitoring systems mainly focus on detecting task completion or postural errors, providing limited clinical interpretation and insufficient corrective rationale. This study proposes a PNF upper limb D1 pattern coaching system by utilizing 3D skeleton and force data for Retrieval-Augmented Generation Vision Language Model (RAG-VLM) based therapist feedback. Based on sensor data, the system quantifies joint range of motion, trajectory, compensatory motion, and force level. These metrics and visual information are integrated with rehabilitation-domain knowledge through an RAG-VLM to generate personalized, evidence-grounded corrective feedback. A pilot investigation with non-expert participants based on a PNF upper limb D1 resistive training session showed the potential of the developed system.

Keywords: Robot Mechanism and Control
Abstract: A control-oriented conceptual design and preliminary CFD-based evaluation are presented for a bio-inspired fixed-wing unmanned aerial vehicle (UAV) with whole-wing actuation. Unlike conventional fixed-wing UAVs that rely on local hinged control surfaces, the proposed platform uses the main wings as movable aerodynamic effectors through wing angle-of-attack (AoA) and backsweep variation. The present study focuses on symmetric wing motions to examine the baseline lift-, drag-, and pitch-moment responses of the architecture. A rigid CFD analysis was conducted using Autodesk CFD at a freestream velocity of 25 m/s. Under symmetric wing AoA variation, the computed lift-direction force increased from 13.81 N at 0 deg to 54.92 N at 12 deg, exceeding the vehicle weight reference between 6 deg and 8 deg. The pitching moment also became more negative with increasing AoA, indicating an increased nose-down moment tendency under the adopted sign convention. Symmetric backsweep produced a weaker drag-direction response, with an approximately 4.0% reduction in Fx from 0 deg to 30 deg and a small non-monotonic variation near 18 deg. These results indicate that wing AoA is the primary lift- and moment-related input in the present configuration, whereas backsweep acts as a secondary aerodynamic modulation input.

Keywords: Robot Mechanism and Control, Robotic Applications
Abstract: The complete kinematic and dynamic modeling of a three-degree-of-freedom Delta robot driven by linear actuators is presented in this paper. Both inverse and forward kinematic equations are derived from the geometric constraints of this parallel mechanism. Two complementary approaches for computing the Jacobian matrices are developed: a derivative-based approach using partial differentiation of the kinematic chain and a constraint-equation-based approach adapted from existing methods for rotary-actuated Delta robots. The equations of the inverse dynamic model are formulated using the virtual work principle under standard simplifying assumptions, including lumped link masses and neglected joint friction. The complete set of derived equations - covering kinematics, velocity mappings, accelerations, and actuator forces - is validated against a high-fidelity Simscape Multibody simulation model constructed from a CAD model of a commercial Delta robot. The excellent agreement between the analytical model and the multibody simulation validates the proposed methods. These models can be directly applied in model-based control design, trajectory planning, and real-time simulation of Delta robots with linear actuators.

Keywords: Robot Mechanism and Control, Robotic Applications, Artificial Intelligence Systems
Abstract: This paper presents DGIST PRIME, a 25 kg parallel-linkage quadruped robot. PRIME stands for Parallel-Link Robot with Intelligent Mechanism-Aware Encoding, reflecting that the learning and control system explicitly encodes the robot's physical linkage mechanism into the joint-observation, control, and reward coordinates. PRIME uses parallel-linkage legs in which the knee actuator is placed near the hip side. This design reduces distal leg inertia and swing load, which is beneficial for fast swing motion, crouching, takeoff, landing, and aggressive acceleration. However, in a parallel-linkage leg, the virtual serial joint coordinate q_s used by the simulator and policy is not identical to the physical parallel actuator coordinate q_p. Because the proposed knee linkage uses a parallelogram-like link arrangement with equal link lengths, the simplified one-leg model gives the coupled knee-side coordinate as q_p2 = q_s1 + q_s2. Thus, even if q_s1 and q_s2 are individually regularized in the serial space, the physical coordinate q_p2 can repeatedly approach the parallel-linkage ROM boundary. To address this coordinate mismatch, the proposed learning framework transforms serial joint states into the parallel coordinate before constructing the joint-related observation, evaluating the low-level PD tracking error, and computing the ROM penalty. Sagittal ROM analysis shows that q_p2 can mechanically approach pi rad; therefore, a conservative upper threshold of 3.0 rad is used in simulation. Under the same command and the same ROM penalty weight, the serial-space penalty and the proposed parallel-space penalty are compared in a start-stop task. The results show that the parallel-space penalty preserves comparable velocity tracking while keeping the learned trajectory farther from the physical parallel-coordinate ROM boundary. Hardware rollout with rapid acceleration and abrupt deceleration further verifies that the learned policy remains inside the conservative parallel-space ROM threshold.

Keywords: Robot Mechanism and Control, Robotic Applications, Artificial Intelligence Systems
Abstract: Reinforcement learning-based quadrupedal locomotion controllers have recently achieved strong traversability over rough terrain by relying only on proprioception. However, without exteroception, these controllers cannot observe upcoming terrain in advance. As a result, they tend to learn conservative gaits that maintain high swing-foot clearance and large actuation margins, even when the underlying terrain-robot interaction becomes locally predictable. This conservative behavior improves robustness against unexpected contacts, but can impose an unnecessary energetic cost during less demanding locomotion phases. We present Adaptive Locomotion with Attuned Context and Energy Regulation (ALACER), a reward modulation framework for improving the energy efficiency of blind locomotion. Instead of explicitly classifying terrain types, ALACER estimates the short-window consistency of the latent context inferred from proprioceptive history. This consistency cue modulates the torque penalty and the desired swing-foot clearance. Consistent contexts promote a low-clearance, energy-efficient gait, whereas inconsistent contexts caused by terrain transitions or persistent contact disturbances relax the additional torque penalty and preserve robust traversal behavior. Simulation results show that ALACER reduces mean joint power by 23.5% on flat terrain compared with the baseline. The ablation study further confirms the benefit of the proposed modulation, with ALACER reducing the cost of transport (CoT) by 14.8% on the uphill-to-plane terrain relative to the baseline.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: Deep reinforcement learning has demonstrated remarkable success in continuous bipedal locomotion, yet formulating effective reward functions often leads to severe gradient interference. To resolve the inherent scale conflicts between dominant survival objectives and delicate kinematic tracking, we introduce a synergistic framework that combines a parameterized gait template with a Dual-Critic Actor-Critic architecture. Rather than treating these as isolated solutions, our approach leverages the gait template to provide physically feasible reference motions without motion capture , while explicitly utilizing the dual critics to decouple the competing reward signals. Validations in simulation and on the real robot demonstrate that this integrated methodology significantly improves natural gait generation and disturbance robustness compared to baseline approaches.

Keywords: Robot Mechanism and Control, Robotic Applications, Artificial Intelligence Systems
Abstract: Obstacle avoidance for robotic arms in dynamic environments relies on online motion planning, where the planner must search efficiently under moving obstacles while maintaining collision safety and continuous execution. This paper proposes HRG-RRT, a heuristic-region-guided online planner for a robotic arm. Point-cloud observations are used to predict task-space regions that are likely to support feasible connections between the current end-effector state and the goal. The predicted region is not treated as an executable trajectory; it only biases sampling in a time-aware RRT planner, while IK feasibility and whole-arm collision checking determine whether candidate motions are valid. To make the learned prior usable during replanning, HRG-RRT combines heuristic-region repair, mixed sampling, multi-subtree bridging, failed-region feedback, and execution-aware fallback. In PyBullet experiments with 50 trials per scene, HRG-RRT shows higher success rates than standard RRT in the static, dynamic, and mixed scenes. It also obtains shorter average task-space paths in all scenes, with lower execution time in the static and mixed scenes and comparable execution time in the dynamic scene.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Sensors and Signal Processing
Abstract: Action chunking has become a common strategy for vision-language-action (VLA) robot policies because it enables temporally consistent robot commands while allowing policy inference to run at a lower rate than the low-level control loop. However, the execution of action chunks introduces a tradeoff between smoothness, responsiveness, and task duration. In this paper, we present an execution-side adaptive action chunking method for closed-loop robot manipulation. First, we formulate overlapping action chunks and introduce a recursive multi-chunk smoothing rule that blends newly predicted chunks with previously accumulated command trajectories. Second, we propose a per-joint adaptive update interval based on the within-chunk action variance, allowing each joint to adopt new predictions at a different rate according to its predicted motion. The method is evaluated in an Isaac Sim–GR00T–ROS 2 closed-loop pipeline using a simulated AI Worker FFW-SG2 robot performing a pick-and-place task. Experiments with a fixed action horizon of H = 16 compare update intervals of R = 16, R = 8, and R = 2, as well as the proposed adaptive interval. The results show that smaller update intervals reduce tracking error but increase task duration, while the adaptive method provides an intermediate tradeoff without modifying the underlying VLA policy.

Keywords: Robot Mechanism and Control, Autonomous Vehicle Systems, Control Theory and Applications
Abstract: This paper proposes an implementation-oriented control framework for a winch-tethered UAV system under tether-induced disturbances. The architecture follows a cascaded hierarchy: an outer translational loop generates a robust acceleration command, a thrust/attitude reference generator maps it into a collective thrust and attitude setpoint, and a kinematic attitude-reference layer shapes the setpoint provided to the PX4 attitude/rate cascade. In parallel, a distance-based winch controller regulates tether length using only the relative UAV--UGV distance, without explicit tension measurements. Unlike a torque-level design, the proposed implementation does not assume direct access to body torques; instead, residual attitude-tracking errors of the cascade are modeled as bounded kinematic mismatches. Under bounded reference signals, bounded tether disturbances, and bounded attitude residuals, we establish uniform ultimate boundedness of the attitude-reference and translational tracking errors through a Lyapunov-based analysis. The controller is implemented in PX4 SITL and validated in Gazebo simulation.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems
Abstract: Energy efficiency is a critical challenge for bipedal humanoid robots operating under limited battery capacity in real-world environments. Conventional model-based controllers struggle to integrate energy optimization with gait generation, and existing deep reinforcement learning approaches rarely incorporate structured biomechanical knowledge into the reward design. This paper proposes a Proximal Policy Optimization (PPO)-based reinforcement learning framework with a phase-aware reward function inspired by human gait biomechanics to achieve energy-efficient forward walking of a 12-DOF bipedal humanoid robot. The proposed reward function detects the single-support phase (SSP) using a ground reaction force threshold and selectively activates five biomechanics-inspired reward terms: knee extension, hip pitch swing, step length, toe clearance, and ankle roll alignment. Training is conducted in NVIDIA Omniverse Isaac Sim with domain randomization applied to mass, posture, sensor noise, and disturbances to reduce the sensitivity of the policy to modeling error. Experiments on flat terrain at 1.0 m/s, evaluated over five independently trained policies with ten randomized 10 s trials each (50 rollouts), show that the proposed policy reduces the mean RMS joint torque by 17.4% and the absolute mechanical cost of transport by 15.5% compared to a stability-oriented baseline, while maintaining comparable command-tracking performance. These results demonstrate that structured biomechanical reward shaping is an effective strategy for energy-efficient bipedal locomotion.

Keywords: Robot Mechanism and Control, Control Theory and Applications, Robotic Applications
Abstract: Residual reinforcement learning (RL) adds learned corrections to a sampling-based motion planner, so a robot can adapt as its environment changes. These corrections, however, become unreliable in out-of-distribution (OOD) environments not seen during training. Existing MC-Dropout-based methods act conservatively whenever the action uncertainty σ2 is high. But high uncertainty does not necessarily mean danger. In our experiments, this uncertainty-only rule reaches only 23.5% Precision, so 76.5% of its interventions are unnecessary. To address this, we present CASM, which combines uncertainty (σ2) with control authority (Capacity). Capacity indicates whether a collision-avoiding action is available within the robot’s bounded action space. When Capacity is sufficient, CASM keeps the RL policy in control even under high uncertainty, and switches to an OMPL-based planner only when high uncertainty and low Capacity occur together. On a simulated UR10e manipulator, CASM raises Precision from 23.5% to 94.1% at 100% Recall and reaches a 0% collision rate in our experimental environments.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Industrial Applications of Control
Abstract: Robotic manipulators operating in shared or constrained environments must reach target poses while avoiding collisions with their own structure and surrounding obstacles. This paper presents a staged reinforcement learning framework for collision-aware goal reaching of a 6-DOF robot manipulator. The proposed approach decomposes the learning process into progressive training stages, starting from a simplified 3-DOF reaching task and extending toward full 6-DOF motion control. To support safe motion generation, the robot body is represented using discretized geometric points, enabling distance-based evaluation of self-collision and obstacle proximity during training. The reward function integrates target-reaching accuracy, collision penalties, near-collision avoidance, and motion smoothness to guide the policy toward feasible and stable behavior. In addition, a multi-agent structure is considered, where different joint groups can be trained and controlled in a coordinated manner. Preliminary results show that staged learning can improve training stability and help the robot acquire collision-aware reaching behavior in complex configurations. The developed framework provides a foundation for future physical AI systems that require autonomous, safe, and adaptive robot motion in dynamic environments.

Keywords: Robot Mechanism and Control, Robotic Applications
Abstract: This study presents an artificial arm that is based on a linkage joint, combining soft electrohydraulic actuators with a rigid linkage mechanism to substantially improve the amplification of angular motion. Soft electrohydraulic actuators exhibit a relatively limited range of linear displacement. Within this framework, the linkage joint transforms the limited linear displacement produced by the electrohydraulic actuators into a greatly amplified rotational motion. A maximum flexion angle of 103 degrees is achieved, which corresponds to an increase of roughly eight times compared to the system without a linkage mechanism. Notably, the hybrid actuator configuration markedly enhances cyclic stability while effectively reducing hysteresis behavior. A comprehensive series of experiments has been carried out to meticulously investigate the effects of various design and input parameters on motion and force performance. These findings highlight the considerable and encouraging potential that hybrid soft–rigid actuation systems possess for driving progress across a broad spectrum of robotic applications.

Keywords: Robot Mechanism and Control
Abstract: The connection mechanisms between neighboring modules are crucial subsystems in modular robots, and magnetic connectors have been proposed as an effective solution. They are widely utilized due to their ease of connection and mechanical simplicity. However, they have the disadvantage of a lower maximum tensile strength compared to connectors using rigid docking mechanisms. This limitation is critical, as modular robots are frequently subjected to impact forces during locomotion, reconfiguration, docking, and undocking. To address this issue, this study proposes a silicone air chamber structure designed to enhance the impact resistance of magnetic connectors while mitigating the degradation of the undocking capability caused by this increased resistance. Furthermore, we introduce an experimental evaluation method inspired by the Charpy impact test. Finally, because the proposed structure is too complex to evaluate through analytical methods alone, the Taguchi method is employed to optimize the four design parameters, ultimately yielding the optimal design values.

Keywords: Robot Mechanism and Control, Robotic Applications
Abstract: This paper examines how foot morphology changes the gait learned by an end-to-end reinforcement-learning controller for humanoid walking. Flat, curved, and active-toe feet were modeled on the Unitree G1 platform and trained with the same learning framework. The flat and curved models share the same action and observation dimensions, whereas the active-toe model includes two additional toe joints and only minimal toe-related regularization. To make the effect of foot design visible during training, we introduced a Froude-number-based step-length reward that scales the target step length with the commanded speed. The trained policies were evaluated in MuJoCo sim-to-sim at several forward walking speeds. The results show that each foot design produces a different step-length pattern: the flat foot has large variability at low speeds, the curved foot becomes less repeatable at high speeds, and the active-toe foot maintains a comparatively narrow step-length distribution over most tested speeds. An additional active-toe policy trained without toe-specific rewards produced smaller steps while keeping low variance, suggesting that the toe morphology itself contributes to repeatable stepping, whereas toe-related rewards help the policy use the toe joint to generate longer steps.

Keywords: Robot Mechanism and Control, Robotic Applications
Abstract: Stable walking pattern generation for humanoid robots in real time remains a key challenge in robotics. The Linear Inverted Pendulum Model (LIPM) has been widely used for this purpose due to its computational efficiency. However, conventional analytical methods assume a constant Zero Moment Point (ZMP) velocity during the support phase, which restricts the ZMP trajectory to a simplified form and leads to increased Center of Mass (CoM) velocity fluctuation. To address this limitation, this paper proposes a Model Predictive Control (MPC)-based walking pattern generation method that subdivides the single support phase (SSP) into multiple subintervals, each assigned an optimized ZMP velocity, thereby generating a piecewise linear moving ZMP trajectory. The optimization problem is formulated for both the SSP and double support phase (DSP), with the ZMP velocity of each subinterval as decision variables. An analytical solution to the LIPM is derived to keep the computational cost low, making the method suitable for real-time control. Simulation results using the Unitree G1 humanoid robot confirm that the proposed method successfully generates moving ZMP trajectories based on walking velocity and effectively reduces CoM velocity fluctuation compared to conventional methods.

Keywords: Robot Mechanism and Control, Robotic Applications
Abstract: This paper presents a GPU-parallelized Model Predictive Path Integral (MPPI) framework for real-time bimanual object manipulation. The framework independently samples contact-wrench rates and joint jerks and couples them through an acceleration-level kinematic consistency cost. Wrench sampling directly propagates object dynamics through the Newton--Euler equations without repeatedly reconstructing object motion from end-effector poses, while jerk sampling explicitly evolves joint-space states for joint-space objectives without online inverse kinematics. Object orientation is propagated on S^3 using an exact discrete quaternion update under piecewise-constant angular velocity, avoiding Euler-angle representation singularities. MuJoCo simulations with two Franka Emika FR3 manipulators evaluate 4096 trajectories in parallel at 300~Hz and demonstrate stable 6-DoF tracking, including a rapid 90^{circ} pitch maneuver for which an Euler-angle-based linear MPC baseline exhibits gimbal-lock-related degradation.

Keywords: Robot Mechanism and Control, Control Theory and Applications, Artificial Intelligence Systems
Abstract: Existing reinforcement learning (RL)-based locomotion policies are highly dependent on the low-level PD gains used during training, often requiring repeated controller tuning and policy retraining for sim-to-real deployment. In this paper, we propose an action conversion framework that enables locomotion policies to be transferred across different low-level controller settings without retraining. By converting policy actions to compensate for changes in controller gains while preserving the intended closed-loop behavior, the proposed method decouples hardware motor tuning from policy learning and improves policy reusability. The framework was validated through simulation and hardware experiments on the Unitree G1 humanoid robot using multiple controller configurations. Experimental results demonstrate that the proposed method maintains the intended control behavior and enables stable locomotion across different low-level controller settings. These results highlight the potential of the proposed framework to simplify controller adaptation and accelerate the sim-to-real transfer process for RL-based locomotion policies.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: This paper presents a self-learning approach to address current difficulties that arise when developing flight control systems using reinforcement learning (RL). We implemented a self-learning system that allows fully automated learning of an actual micro-aerial vehicle (MAV) in consideration of safety. This system utilizes an emulated free flight test environment based on magnetic levitation with a state-of-the-art RL algorithm directly to train the control policy of a physical MAV, thereby narrowing the reality gap associated with numerical modeling in simulation-based learning. Moreover, the trial-and-error-based learning procedure can be conducted during ground-emulated flight tests without serious safety concerns because the magnetic forces acting on the MAV are dynamically adjusted. The feasibility of the proposed approach was demonstrated experimentally; a MAV deliberately configured to have unstable yaw dynamics was challenged to use a self-control strategy to regulate its yaw attitude through piezoelectric air brakes without any prior knowledge. These results offer an alternative means of learning by demonstrating the successful solution of a demanding physical control problem.

Keywords: Robot Mechanism and Control, Rehabilitation Robot, Biomedical Instruments and Systems
Abstract: This study presents the Cable-actuated Robotic Prosthesis with Parallel mechanism (CROPP), a transradial prosthetic system designed to reduce distal mass while providing fully actuated three-degree-of-freedom (3-DoF) wrist motion. Conventional multi-axis prosthetic wrists often require multiple actuators and transmission components near the distal joint, increasing weight, mechanical complexity, and user burden. To address the trade-off between dexterity and wearability, CROPP integrates a compact 3RRR parallel mechanism for flexion-extension (FE) and radial-ulnar deviation (RUD) with an independent central-shaft mechanism for pronation-supination (PS). The actuators are relocated proximally, and motion is transmitted through Bowden cables, reducing the mechanical load concentrated at the distal joint. The wearable prototype has a total mass of 380 g and provides angular ranges of ±60° for FE and RUD and ±90° for PS. Motion-capture-based experiments evaluated independent trajectory tracking under 0.5 Hz square, triangular, and sinusoidal reference inputs. Across all tested conditions, the mean absolute error ranged from 2.26° to 3.99°. The PS DoF exhibited the lowest overall tracking error, attributed to its direct 1:1 pulley transmission, simplified transmission path, and reduced cable slack. The FE and RUD responses showed slight phase delay and waveform deformation, mainly due to Bowden-cable friction, elastic deformation, and tension variation. These results demonstrate the feasibility of a lightweight, remotely actuated multi-DoF wrist architecture and indicate its potential for transradial prosthetic applications involving activities of daily living.

Keywords: Robot Mechanism and Control, Control Theory and Applications, Robotic Applications
Abstract: This paper presents a risk-adaptive preferred-clearance extension of nonlinear model predictive cooperative control (NMPCC) for dual-arm manipulators. A preferred clearance is introduced and adjusted according to the obstacle risk level. As a result, the controller generates more conservative trajectories near high-risk obstacles such as humans, while allowing more efficient paths near lower-risk objects when the minimum safety boundary is still satisfied. Simulation results validate that, compared with the baseline NMPCC, the proposed method eliminated safe-zone violations in all scenarios and increased the minimum obstacle distance while maintaining target convergence.

Keywords: Robot Mechanism and Control, Robotic Applications, Human-Robot Interaction
Abstract: In teleoperated robotic systems, accurate perception of physical interaction forces is required to prevent grasping failures and ensure manipulation stability. Current wearable haptic interfaces, nevertheless, exhibit an inherent structural dichotomy. Specifically, ungrounded cutaneous devices fail to provide joint-level kinesthetic cues, whereas conventional kinesthetic exoskeletons restrict natural dexterity due to their rigid kinematics. To address these limitations, an integrated hybrid haptic glove is developed based on a spatially decoupled sensory routing mechanism. At the distal phalanx, a lightweight tendon-driven module embedded with a force-sensing resistor (FSR) provides closed-loop cutaneous feedback, actively compensating for the inherent friction of the tendon routing. Concurrently, a compact solenoid-spring array mounted on the metacarpophalangeal (MCP) joint mechanically renders seven discrete levels of equivalent stiffness to simulate object compliance. Experimental evaluations validate the proposed architecture, demonstrating accurate closed-loop force tracking at the fingertip and linear force-displacement profiles across the discrete stiffness states. Consequently, the proposed interface provides a robust, ungrounded framework for multi-modal compliance perception, enabling stable force rendering in telemanipulation tasks.

Keywords: Robot Mechanism and Control, Human-Robot Interaction, Robotic Applications
Abstract: This paper presents a multimodal teleoperation framework integrating vision-based manipulation, pedal-based navigation, and an LLM-based voice interface for a dual-arm mobile manipulator. Human hand motions provide intuitive manipulation control, while voice commands enable robot interaction through natural language. Experimental results in simulation and real-world environments demonstrate effective multimodal teleoperation and validate the benefits of voice-assisted interaction for improving manipulation efficiency and precision while preserving intuitive control.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: Human legs exhibit high efficiency, agility, and impact absorption, with toes playing a crucial role in these capabilities. While many attempts have been made to implement human-like toes in robots, they have not fully replicated human characteristics nor rigorously validated their benefits. We propose a 14-DOF bipedal robot emulating human toes’ lightweight, high-torque, robust nature. To quantitatively analyze the effectiveness of the active toes in terms of agility, efficiency, and impact absorption, we developed an actuator- and transmission-aware simulation training environment that reflects actual actuators with coupled transmissions and motor-side power consumption. To ensure a fair comparison between configurations with and without active toes, we applied an identical training procedure to both. The simulation results indicate that, at 1.33 m/s walking, the toe-equipped model reduced the cost of transport (CoT) by 17.5% and heel-strike ground reaction force (GRF) by 5.0% compared with the toe-ablation configuration. On the agility test, the toe-equipped model was 4.4% slower on average but reduced mean path deviation by 25.0%, completing the course in the same time to within 0.1%.

Keywords: Robot Mechanism and Control, Industrial Applications of Control, Control Devices and Instruments
Abstract: This paper proposes a dual-viewpoint vision system that uses a single 2D camera mounted on a 6-degree-offreedom robotic arm. The proposed system moves the camera up and down to perform triangulation based on Oriented FAST and Rotated BRIEF feature points. To handle stacked environments, we propose an object-separation method that uses the bottom object to compute XY-plane coordinates and the top object to estimate height, independently of each other. We also apply a top-band masking technique that extracts only the topmost region of the object, which reduces feature-matching errors and greatly improves the accuracy of height estimation for the top surface of the target object. Finally, based on the estimated height, 2D pixels are back-projected into 3D space to dynamically calculate the physical grasp width for adaptive gripper control. By applying a Median Absolute Deviation-based time-series stability filter, the proposed system achieved a 97.5% grasp success rate across various single-object and stacked-object scenarios, demonstrating highly robust manipulation performance without any additional 3D sensor.

Keywords: Robot Mechanism and Control, Robotic Applications, Artificial Intelligence Systems
Abstract: Color-conditioned object sorting is difficult for text-only Vision-Language-Action (VLA) policies because they must simultaneously interpret color-related language, localize the target object, infer the desired destination, and generate low-level robot actions. This paper proposes AutoGround-VLA, an autonomous visually grounded VLA frame- work that separates target reasoning from action generation. A Qwen3-VL 30B Vision-Language Model (VLM) pre- dicts the target bounding box from the current robot image and a target command, while a GR00T N1.7-based VLA policy executes manipulation using the marked image, original multi-view observations, robot state, and a simplified destination command. The system also uses an inference trig signal to invoke VLM grounding only when new target reasoning is required. Experiments on the DARU humanoid gripper platform show that the proposed method im- proves in-distribution sorting, color-destination compositional generalization, and unseen-color extrapolation compared with text-only baselines.

Keywords: Robot Mechanism and Control, Exoskeleton Robot, Control Devices and Instruments
Abstract: Knee osteoarthritis is a prevalent musculoskeletal disorder, yet the clinical efficacy of existing active lower-limb exoskeletons is often limited by poor kinematic alignment and high metabolic penalties. Conventional single-axis and four-bar joints fail to accurately track the human knee's migrating instantaneous centre of rotation (ICR), inducing parasitic shear forces, while heavy joint-mounted actuators disrupt natural gait. This paper presents the design and optimisation of a lightweight, active knee exoskeleton featuring a geared five-bar polycentric linkage. Optimised via a genetic algorithm, the mechanism eliminates kinematic singularities within the functional range of motion and achieves an ICR-tracking root mean square error of 0.47 mm over a 0◦– 60◦ range, satisfying the 0.5 mm design target. A decentralised actuation architecture routes power from a waist-mounted motor via bidirectional Bowden cables to reduce added mass on the leg. A quasi-static transmission analysis indicates a motor torque of 8.1 Nm is required to deliver the 35 Nm target assistive torque during stance; a 15 Nm continuous-torque actuator was selected. An iterative learning controller is proposed to compensate for Bowden cable friction and compliance hysteresis. The fabricated prototype demonstrates the physical feasibility and wearable form factor of the design.

Keywords: Robot Mechanism and Control, Artificial Intelligence Systems, Robotic Applications
Abstract: This paper compares a fixed-toe Flat-Foot model and a movable-toe Toe-Foot model in a reinforcement- learning-based one-leg jumping task. Both models are trained with proximal policy optimization under the same command distribution, reward structure, network configuration, and transition budget. The command vector specifies planar velocity, yaw rate, and base height, and all three objectives are explicitly represented in the reward. The effect of toe mobility is evaluated from the angular velocities of the six actuated leg joints: hip yaw, hip roll, hip pitch, knee pitch, ankle pitch, and ankle roll. Root-mean-square (RMS) and signed peak angular velocities are compared for each joint, together with the reduction rate from Flat-Foot to Toe-Foot. One independently trained policy per model is evaluated over ten consecutive jumping cycles under identical command conditions. Overall, the Toe-Foot model reduces angular-velocity RMS by 27.9% and peak magnitude by 34.75%.

Keywords: Robot Mechanism and Control, Robotic Applications, Sensors and Signal Processing
Abstract: This paper presents a spring-preloaded dual-worm anti-backlash mechanism for a two-axis robotic fire monitor intended for firefighting robot applications. Worm-gear transmissions are suitable for robotic fire monitors because of their compact structure, high reduction ratio, and load-holding capability under water-jet reaction forces. However, backlash in worm-gear transmissions can cause lost motion, hysteresis, and direction-dependent aiming error during motion reversal. To reduce backlash-induced angular error, the proposed mechanism adds a synchronized passive worm gear to the driving worm gear and applies axial spring preload to maintain tooth contact with the output gear. The mechanism was implemented in a two-axis fire monitor prototype, and its initial experimental evaluation was conducted on the azimuth axis. Backlash was evaluated by comparing the motor-side ideal output angle, calculated from motor rotation and gear ratio, with the output-side angle measured by an external encoder. Dry and water-discharge experiments were conducted to investigate the effects of spring preload and approach direction on angular error and motor current. The experimental results show that the proposed mechanism reduces backlash-induced angular error under both dry and water-discharge conditions without a noticeable increase in motor current.

Keywords: Sensors and Signal Processing, Artificial Intelligence Systems
Abstract: This paper presents a framework for joint estimation of golf club motion and full-body golfer motion using a club-mounted inertial measurement unit (IMU) and an external RGB-D camera. An indirect Kalman filter fuses IMU measurements with visual observations to estimate golf club motion while compensating for sensor drift, calibration errors,and mounting uncertainties. The estimated club motion is then integrated into an optimization-based SMPL-X framework through 3D joint constraints and hand–club interaction constraints to reconstruct full-body motion. Experimental results demonstrate that the proposed method significantly improves clubhead trajectory estimation compared with an IMU-only approach, achieving a mean position error of 2.8 cm while producing physically consistent golfer motion. The proposed system provides an effective solution for comprehensive golf swing analysis.

Keywords: Sensors and Signal Processing
Abstract: This paper proposes a photon counting stabilized Retinex method for low-light image enhancement in extremely dark environments. The proposed method first interprets the luminance component of the input image as a photon count representation and applies a Poisson-aware stabilization process to reduce shot noise caused by insufficient photon arrivals. The stabilized luminance is then processed using multi-scale Retinex (MSR) to generate a Retinex prior that compensates for illumination variations and enhances structural information. The photon stabilized luminance and Retinex prior are fused through a maximum a posteriori (MAP)-based weighted fusion scheme to restore brightness and fine details in dark regions. To further suppress noise amplified during enhancement, a mask-based post-processing stage for preserving the brightness is applied in the YCrCb color space, selectively reducing luminance and chrominance noise while preserving edges and structural details. Experiments using the See-in-the-Dark (SID) dataset show that the proposed method improves dark-region visibility and reduced color and background noise compared with conventional Retinex-based methods, and the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) results prove the effectiveness of the proposed method.