Keywords: Dynamics and Control, Aerial & Field Robotics, Mechatronic systems
Abstract: Fully actuated tilt-rotor unmanned aerial vehicles, designed to enable complex spatial exploration, have demonstrated significant potential in recent years. Nevertheless, the strong coupling among control objects has motivated increasing studies to adopt single-controller approaches, which compute actuator inputs directly from system attitudes rather than relying on pseudo-inverse matrix operations to circumvent the associated computational complexity and numerical instability. In this paper, a nonlinear model predictive control-based strategy is proposed to address the control allocation problem in a tilt-rotor quadcopter prototype, while computational efficiency is enhanced by reducing system order through decomposition of control allocation from kinematic control. The study presents a custom-built prototype alongside detailed aerodynamic modelling, and introduces a low-level model-based allocator that integrates aerodynamic and electrical characteristics to optimally distribute force and torque across thrust and tilt actuators. In the comparative simulation studies, trajectories have been designed to demonstrate the efficiency of the proposed allocator.

Keywords: Aerial & Field Robotics, Multi-Robot Systems
Abstract: Drones (Uncrewed Aerial Vehicles, UAVs) are increasingly used for data collection in domains such as nature conservation, agriculture, and disaster response. Commercial off-the-shelf platforms, particularly DJI drones, dominate the market due to their reliability and affordability. However, low-end models often lack developer-friendly interfaces and onboard Software Development Kit (SDK) support, limiting their applicability in research contexts that require real-time control, multi-drone coordination, or vision-based autonomy. We introduce WildBridge, an open-source Android application that extends the DJI Mobile SDK to provide telemetry, video, and low-level control capabilities. WildBridge runs on the mobile device connected to the DJI remote controller. It exposes network interfaces such as HTTP for telemetry and control, and RTSP for video streaming over a local area network, enabling interaction with ground stations and external programs. In feasibility experiments, round-trip telemetry latency at the ground station remained below 113 ms on average (90th percentile under 290 ms) for request rates up to 32 Hz across ten concurrent drones, while video latency stayed within 1.9 s for up to six simultaneous streams. By simplifying interaction with DJI drones, WildBridge lowers the entry barrier for research applications, supports reproducible experiments, and enables integration with common robotics frameworks such as ROS 2.

Keywords: Aerial & Field Robotics, Dynamics and Control
Abstract: Autonomous recovery of Vertical Take-Off and Landing aircraft on moving ships remains a critical challenge due to deck oscillations, particularly in high sea states. Poorly timed landings under these conditions can increase the risk of rotorcraft rollover and structural damage. To address this issue, a Visual Servoing Nonlinear Model Predictive Control (VS-NMPC) framework is proposed, which integrates visual servoing kinematics with short-term vessel state forecasts. The target states are incorporated into the cost function through a barrier term, allowing the controller to mitigate unsafe landings. The approach was validated through indoor flight tests using a scaled-down platform that replicates the motion of a heavy-class destroyer under Sea State 6, including roll oscillations and forward motion. Onboard processing tests demonstrated the controller’s performance under real-time constraints. The study introduces a statistical assessment of touchdown performance considering both precision and target roll limits. The results show the potential of VS-NMPC as a robust solution for autonomous recovery operations in challenging maritime conditions.

Keywords: Aerial & Field Robotics
Abstract: This paper proposes a unified framework to assess the reliability of multirotors with rotor failures under diverse internal and external operating conditions. Specifically, we address the following question: Given the H/W and S/W assets equipped in the multirotor, what would be the reliable flight regime in which the multirotor can tolerate the rotor failures while resisting the wind? To this end, we develop a fully automated simulation testbed whereby the behavior of the faulty multirotor is examined with respect to different operational scenarios composed of various combinations of rotor failures, H/W and S/W assets, and wind conditions. The simulation process is streamlined by selectively executing a representative subset of scenarios and inferring the results of those that are not executed. This strategy facilitates the efficient assessment of multirotor reliability with fewer simulations while maintaining accuracy comparable to complete enumeration of all possible combinations. Furthermore, we introduce effective visualization techniques that project the reliable flight envelope to 3-D heatmap, which will be particularly instrumental in developing fault-tolerant strategies from a holistic perspective.

Keywords: AI & ML & Deep RL, Aerial & Field Robotics, Dynamics and Control
Abstract: This paper proposes an anomaly detection framework for multicopter systems that combines a one-dimensional convolutional autoencoder with long short-term memory (LSTM) networks. Traditional approaches to anomaly detection often rely on predefined features or require large amounts of labeled fault data, making them less effective in complex flight environments where fault data is scarce. To address this limitation, the proposed method integrates a 1D convolutional neural network (1D-CNN) to extract local temporal features and an LSTM network to capture long-term dependencies within time-series flight data. The autoencoder is trained on normal operating conditions, and anomalies are detected when the reconstruction error between input and output exceeds a predefined threshold. The framework is validated through simulation experiments on hexacopter systems, demonstrating its capability to achieve accurate and efficient fault detection even in the absence of sufficient fault data. These results confirm the potential of the proposed model to enhance the safety and reliability of multicopter operations.

Keywords: Dynamics and Control, Aerial & Field Robotics
Abstract: This paper presents a robust path-following control strategy for a tilt-wing unmanned aerial vehicle (UAV), a hybrid platform that merges the vertical take-off and landing (VTOL) capabilities of a multirotor with the efficient forward flight of a conventional (CTOL) aircraft. Addressing the platform's inherent nonlinear dynamics, we propose a novel cascaded control architecture. Distinct from prior works that have generally restricted Incremental Nonlinear Dynamic Inversion (INDI) to inner-loop attitude stabilization, our framework applies this powerful technique to both a high-bandwidth inner-loop and an outer-loop for precise path following. This fully incremental approach minimizes reliance on an accurate model, compensating for uncertainties across the flight envelope. The controller's ability to outperform both a single-loop INDI controller and a PID controller in both nominal and disturbed conditions is validated through high-fidelity simulations. Robustness to parametric uncertainties is statistically demonstrated via Monte Carlo simulations, while resilience against external disturbances is confirmed using a standard wind turbulence model. Results show the cascaded INDI controller provides superior tracking accuracy and robustness, enabling reliable autonomous flight for the complex tilt-wing platform.

Keywords: Multi-Robot Systems
Abstract: Solving multi‑robot task allocation (MRTA) in a distributed manner remains challenging because distributed consensus ADMM (D‑CADMM) often converges slowly in large networks. We adopt the Hybrid Consensus ADMM (H‑CADMM) framework, which introduces local fusion centers (LFCs) to accelerate information flow. Because performance hinges on LFC placement, we analyze the limitations of degree‑based greedy selection in homogeneous networks and propose a minimum dominating set (MDS)–based strategy. Monte Carlo simulations show that our MDS‑based selection consistently reduces communication volume relative to greedy selection by yielding a more efficient LFC topology. Overall, the hybrid framework significantly outperforms the fully distributed approach in scalability and stability, offering a practical, communication‑efficient solution for real‑world multi‑robot systems.

Keywords: Multi-Robot Systems, Aerial & Field Robotics, AI & ML & Deep RL
Abstract: This paper presents a novel approach for collective navigation of unmanned aerial vehicle (UAV) swarms without information exchange, using onboard LiDAR sensing within an implicit leader-follower framework. Our system detects UAVs equipped with reflective tape via high-intensity LiDAR point clustering and uses an extended Kalman filter for stable neighbor tracking. The resulting perception is robust across 360 degrees and various lighting conditions and not dependent on the external positioning system. Our approach achieves high recognition reliability while avoiding the hardware complexity of multi-camera systems typically required for omnidirectional perception. The core of our approach is a deep reinforcement learning controller based on proximal policy optimization, trained in a GPU-accelerated Nvidia Isaac Sim environment. This controller enables follower UAVs to learn a balance between flocking and obstacle avoidance, resulting in an emergent behavior of implicitly following a leader while robustly addressing perceptual challenges such as occlusion and limited field-of-view. The robustness of our approach is confirmed through extensive simulations and challenging real-world experiments with a swarm of five UAVs, which successfully demonstrated collective navigation across diverse indoor and outdoor environments under complete communication denial.

Keywords: Multi-Robot Systems, AI & ML & Deep RL
Abstract: We propose a Graph Neural Ordinary Differential Equation (Graph Neural ODE) framework for modeling and predicting trajectories of heterogeneous agents in warehouse environments. Unlike conventional graph neural networks that perform discrete message passing, our approach treats multi-agent interactions as a continuous-time dynamical system evolving over a spatio-temporal graph. The ODE function is parameterized by a graph neural network that jointly captures spatial relations among agents and their continuous temporal evolution, enabling smooth and physically consistent motion prediction while alleviating discretization artifacts and over-smoothing effects often observed in recurrent or step-based models. Experiments are conducted across five simulated warehouse configurations ranging from small to large-scale systems. The proposed model consistently outperforms recurrent baselines, achieving substantially lower position and collision errors while maintaining robust performance across all environment scales. The results show that Graph Neural ODE maintains stable performance as the system grows in size and complexity, producing coherent and collision-free trajectories even in highly congested environments. Beyond individual trajectory accuracy, the model captures realistic collective behaviors including synchronized task arrivals and adaptive avoidance maneuvers, demonstrating its capability to generalize from local motion to emergent system-level coordination. By providing continuous, interpretable representations of agent dynamics, this work establishes a principled graph-based framework for multi-agent trajectory prediction and a foundation for future extensions in coordinated robotic systems.

Keywords: Multi-Robot Systems, Safe Decision Making under Uncertainty, Risk-Aware Autonomy
Abstract: In this study, we introduce a decentralized algorithm for deadlock resolution and collision avoidance in multi-agent systems, enhancing efficiency in dense, complex environments. To this end, the proposed algorithm categorizes each agent into primary and secondary roles through communication and score comparisons among neighbors. Primary agents navigate using an initial path generated by the probabilistic roadmap (PRM), while secondary agents prioritize yielding space to primary agents. The optimal reciprocal collision avoidance (ORCA) algorithm is utilized for collision avoidance. Through benchmarking against the standard ORCA in diverse obstacle-rich maps, we observed notable reductions in the makespan and increased success rates, highlighting the efficacy of our approach.

Keywords: Dynamics and Control, Multi-Robot Systems, Navigation, Perception & SLAM
Abstract: As a key component of autonomous navigation, trajectory planning critically influences system performance but remains limited when applied to complex multi-agent systems in collaborative obstacle avoidance and trajectory optimization. This study proposes a topologybased trajectory planning method for Aerial–Ground Cooperative Cable-Driven Parallel Robot Systems, enabling efficient, robust movement. A verification platform was developed and validated through experiments, demonstrating that the proposed algorithms outperform traditional methods in disturbed environments with reduced errors. These findings provide both theoretical value and practical guidance for trajectory planning in Aerial–Ground Cooperative Cable-Driven Parallel Robot Systems.

Keywords: AI & ML & Deep RL, Multi-Robot Systems, Risk-Aware Autonomy
Abstract: This paper presents a novel reinforcement learning framework for coordinated multi-agent navigation, integrating potential field-based rewards to enhance efficiency. The decentralized multi-agent reinforcement learning approach employs a hybrid reward function that dynamically balances goal attraction, obstacle repulsion, and inter-agent spacing, enabling coordinated motion and collective goal-reaching. Utilizing LiDAR-based observations and relative position vectors, each agent optimizes its policy using the multi-agent proximal policy optimization algorithm to learn safe and efficient navigation in a continuous state-action space. The potential field-based reward formulation provides continuous feedback by modeling attractive forces toward the goal and repulsive forces from obstacles and other agents, ensuring smooth and adaptive motion. Simulations validate the framework's effectiveness in achieving real-time coordinated motion.

Keywords: Dynamics and Control, Multi-Robot Systems, AI & ML & Deep RL
Abstract: Satellite formation control has become essential as space missions grow increasingly complex. Beyond individual satellite control, successful formation flying requires coordinated navigation that accounts for all group satellites to prevent collisions. However, achieving perfect satellite formation control where multiple satellites reach desired positions while avoiding inter-satellite collisions presents significant challenges. This paper addresses these challenges by proposing a learning-based controller for safe collision-aware navigation. The approach employs neural ordinary differential equations, naturally suited for optimal control problems, combined with a novel loss function that integrates both collision avoidance and formation control objectives. The proposed method efficiently trains a lightweight neural controller with minimal parameters, and numerical experiments demonstrate the controller's effectiveness.

Keywords: Aerial & Field Robotics, AI & ML & Deep RL, Safe Decision Making under Uncertainty
Abstract: This paper addresses vision-based autonomous landing of quadrotor drones on moving platforms with high speed (up to 8~m/s) and uncertain dynamics. It is a capability vital for maritime operations, logistics, and emergency missions, thereby enhancing overall operational flexibility. Landing on high-speed platforms significantly expands mission scope but introduces challenges such as rapid relative motion change between the drone and the platform, the platform’s unpredictable dynamics, and frequent occlusions caused by the limited camera's field of view (FOV). Traditional visual servoing methods, including position-based and image-based schemes, are sensitive to estimation errors and require precise parameter tuning, whereas deep reinforcement learning (DRL) can alleviate these issues but often struggles under fast motion and visual interruptions. To overcome these limitations, this paper proposes an image-based DRL framework that integrates a pretrained keypoint encoder with an LSTM module, enabling reliable state estimation even when the target becomes occluded or moves outside the FOV. Moreover, an active perception reward is proposed to encourage actions that improve keypoint visibility and estimation reliability. Extensive simulations provide quantitative comparisons against visual servoing and DRL baselines, while real-world tests confirm the practical feasibility and robustness of the proposed framework.

Keywords: Dynamics and Control, Multi-Robot Systems
Abstract: This paper proposes a cooperative and robust multi-agent collision avoidance algorithm for satellite reconfiguration. Local estimation of other agents' positions and the augmented Lagrangian method are employed to modify the constraints and ensure the validity of the estimation, respectively. The modified problem for each agent is solved sequentially via sequential convex programming while achieving consensus. Optimality and runtime are compared with the existing decoupled approach, showing improved optimality and robustness while maintaining computational efficiency. An illustrative example is provided to investigate the characteristics of the proposed algorithm, demonstrating how the responsibility for collision avoidance is cooperatively shared among agents.

Keywords: Formal Verification in Robotics, Dynamics and Control, Aerial & Field Robotics
Abstract: This study proposes a real-time Vehicle-in-the-Loop (VIL) framework for reliable verification of quadcopter flight control systems. Unlike conventional VIL environments that suffer from limited dynamic fidelity and cannot effectively validate position controllers, the proposed framework integrates a gimbal-type test bed, a real-time computer, and a GPS simulator, enabling realistic dynamic feedback and indoor verification of both attitude and position controllers. Experimental results demonstrate the importance of dynamic fidelity compared to conventional setups and confirm that the framework enables position controller verification previously considered infeasible. Moreover, the observed consistency between SIL and VIL results highlights the effectiveness of the framework in enhancing controller reliability and applicability.

Keywords: Dynamics and Control, AI & ML & Deep RL, Aerial & Field Robotics
Abstract: This paper proposes a sensor fault-tolerant system that employs a virtual sensor based on data-driven dynamics modeling. Existing approaches to sensor fault tolerance typically rely on fault estimation using Kalman lters (KF) or observers. However, KF-based methods require fault models and assume slowly varying fault dynamics. While some observer-based methods, such as sliding mode observers, can mitigate this limitation, they require measurements from redundant sensors. This requirement restricts their applicability to hardware-redundant systems. To address these challenges, we design a virtual sensor that can replace a physical sensor by leveraging data-driven dynamics modeling. An accurate model for the uncertain dynamics can be obtained through supervised learning. In this study, physics-informed neural ordinary differential equations (PI-NODE), a type of residual learning, are applied to construct a dynamic model. Based on this model, a virtual sensor is designed to predict future states. Considering real-world implementation, flight data are used for training and validation. Performance analysis demonstrates that the trained PI-NODE model is feasible for application as a virtual sensor in fault-tolerant systems.

Keywords: Safe Decision Making under Uncertainty, Risk-Aware Autonomy
Abstract: Path planning in uncertain environments is a core challenge for autonomous systems, where guaranteeing satisfying safety is as critical as computational performance. Existing methods exhibit a clear trade-off between safety, path quality, and computation time. This paper proposes a novel 3-stage hierarchical pipeline that effectively resolves these trade-offs. First, the framework transforms the probabilistic problem into a deterministic one by constructing a risk-aware Configuration Space (C-Space) using Chance Constraints. Second, a fast topological search using Delaunay Triangulation and A* rapidly finds a globally feasible corridor. Finally, this corridor is refined into a smooth path using GPU-based optimization with ADMM and path smoothing. Simulations demonstrate that the proposed method achieves a superior trade-off between path quality and computation time compared to benchmarks. The framework consistently generates safe, suboptimal paths while maintaining an efficient computation time. Overall, the framework provides a practical approach for generating safe and efficient paths under uncertainty.

Keywords: Multi-Robot Systems, Dynamics and Control
Abstract: The proposed control framework introduces a dissensus-based strategy for multi-vehicle coordination under antagonistic interactions. Antagonistic interactions, modeled as negative weights in the communication graph, represent opposing influences where vehicles repel each other instead of aligning their states. While common in social networks, such interactions are here leveraged for safety in robotic systems, ensuring that vehicles deliberately diverge when approaching too closely. This mechanism guarantees collision avoidance without sacrificing global convergence. To analyze stability, Lyapunov exponents are employed as a measure of local divergence in the dynamics. Positive local Lyapunov exponents,induced through the symmetric part of the Jacobian of the avoidance dynamics,quantify the repulsive effect and enforce safe inter-vehicle distances. Meanwhile, the global system is stabilized by a Lyapunov function whose negative derivative guarantees convergence of all vehicles toward their designated goals. This dual mechanism of global convergence with local divergence ensures that multiple intelligent vehicles can safely navigate dense or unregulated environments. The framework is validated through simulations where six vehicles traverse an intersection while avoiding collisions and achieving their targets. Overall,the strategy establishes a dissensus-driven control paradigm that combines safety with guaranteed goal-reaching in multi-vehicle motion planning.

Keywords: Dynamics and Control, Formal Verification in Robotics, AI & ML & Deep RL
Abstract: The estimation of the Region of Attraction (ROA) is fundamental to the formal verification of nonlinear flight control systems. This paper presents a comparative study of Sum-of-Squares (SOS) optimization and a neural network-based Lyapunov approach for ROA analysis on an F-8 aircraft benchmark. We evaluate both frameworks on key metrics, including the size of the estimated ROA and computational cost. Our findings reveal a critical trade-off: while the neural network learns a Lyapunov candidate to the closed-loop dynamics, the portion of this region that can be formally verified is significantly smaller than the certified ROA from the SOS method. The SOS approach produced a certified ROA containing 84.4 percent of the sampled stable points in approximately 57 minutes. In contrast, the neural method's verifiable ROA contained only 3 percent of the same points after 12.5 hours of computation. This work highlights the practical challenges of formal verification for neural Lyapunov functions, providing critical insights for their application in aerospace engineering.

Keywords: Formal Verification in Robotics, Risk-Aware Autonomy
Abstract: This paper proposes a sampling-based safety filter algorithm that guarantees the safety of robotic systems. For discrete-time systems with known dynamics, future control input sequences are sampled based on the current input, and rollouts are performed for each sample to evaluate safety. If all rollouts are determined to be unsafe, the safety filter intervenes by replacing the current input with a safe alternative. Safety evaluation is formulated using the constraint of the Control Barrier Function (CBF) method. Simulation results validate that the proposed safety filter ensures safety while preserving the autonomy of robotic systems. Furthermore, the experimental study demonstrates the effect of the sampling horizon, the CBF parameter, as well as the sampling distribution and the number of samples on the performance of the safety filter.

Keywords: Navigation, Perception & SLAM, Risk-Aware Autonomy
Abstract: Recent studies have explored path planning methods that achieve collision avoidance in environments represented by 3D Gaussian Splatting (3DGS) maps. However, these approaches focus solely on physical collision avoidance while neglecting semantic aspects of safety. To address semantic risks that cannot be described by geometric volumes, we propose a cost model that evaluates risks from visual information and incorporate this cost model into the state expansion stage of the RRT*-based path planning method. For each candidate state, novel-view images are generated from 3DGS and provided to the cost model, which estimates the semantic risk. Only states whose risks fall below a predefined threshold are retained, ensuring that the resulting path avoids both physical and semantic hazards. We validate the proposed method in a simulation environment containing both physical pillar obstacles and semantic hazards without geometric volume. Experimental results demonstrate that it can effectively avoid semantic risks.

Keywords: AI & ML & Deep RL
Abstract: This paper presents an AI–driven scenario generation framework for safety evaluation of autonomous driving systems, designed to efficiently identify safety-critical scenarios by integrating domain knowledge, optimization, and AI. The framework is structured into three main phases: i) driving dataset generation, ii) conditional variational auto-encoder (CVAE) training, and iii) safety-critical scenario generation and evaluation. In the first phase, dynamic driving scenarios are automatically produced using ontological domain knowledge and a genetic algorithm (GA) to construct a dataset with potentially hazardous cases. The second phase involves training a CVAE network on this dataset, enabling the generation of diverse yet realistic variations of driving scenarios. In the final phase, the trained CVAE is used to adversarially perturb scenarios to uncover safety-critical conditions. Experimental results demonstrate that the proposed framework effectively discovers a wide range of safety-critical situations across different driving environments, highlighting its utility for evaluating the safety of autonomous driving systems.

Keywords: Multi-Robot Systems, Risk-Aware Autonomy, Robot Manipulation
Abstract: With the increasing frequency of satellite launches, a growing accumulation of space debris, including defunct satellites and spent rocket bodies, has become a critical issue. This trend has highlighted the urgent need for effective removal strategies. Previous studies have predominantly focused on single-satellite operations or cooperative removal of a single target by multiple satellites. In contrast, this study addresses a newly defined problem in which heterogeneous satellites with varying capabilities and constraints individually remove multiple debris. To address this, we propose a tree-based optimization algorithm named TRIMOS (Tree-based Integrated Multi-Objective Optimization for Multi-Satellite Debris Removal). TRIMOS prioritizes debris targets using hazard scores derived from orbital crowding levels and plans removal sequences through a four-phase orbital transfer model that enables precise path optimization. Simulation results demonstrate that the proposed algorithm satisfies various operational constraints and consistently outperforms greedy methods once a sufficient number of iterations is reached. This research suggests a practical framework for space debris removal that considers both the scalability of dynamic orbital environments and operational efficiency, indicating strong potential for integration into future multi-satellite management systems.

Keywords: Navigation, Perception & SLAM, AI & ML & Deep RL
Abstract: Accurate and stable perception is essential for autonomous driving systems and represents the initial step toward ensuring safe autonomous navigation. However, the majority of research in the field of autonomous driving perception has predominantly concentrated on optimal weather and lighting conditions. Consequently, achieving stable perception for autonomous driving during nighttime remains a significant challenge. This study presents the development of an advanced YOLO network, referred to as Enhanced Night-Time Perception Yolo Network(ENTP-YOLO), aimed at enhancing the safety and robustness of nighttime autonomous driving perception systems. Furthermore, extensive experiments were conducted utilizing publicly available datasets, demonstrating superior performance compared to the incorporation of additional attention mechanisms, such as the Convolutional Block Attention Module (CBAM).

Keywords: Human Robot Interaction, Multi-Robot Systems, Mechatronic systems
Abstract: This paper presents a real-time publish/subscribe protocol tailored for distributed robotic control with hot-plug-and-play capabilities. Publish/subscribe (Pub/Sub) architectures are increasingly relevant in distributed systems, especially in do-mains where real-time performance and modular scalability are crucial, such as the Internet of Things (IoT) and mobile robotics. Existing lightweight protocols like MQTT provide asynchronous communication between decoupled compo-nents but face limitations when stringent latency and determinism are required. To address this gap, we introduce a protocol designed specifically for robotic applications that integrates real-time guarantees with the flexibility of Pub/Sub communication. The proposed system architecture supports dynamic integration of robotic agents, allowing devices to seamlessly join or leave the network without disrupt-ing ongoing operations. Core features include deterministic message delivery, low-latency event propagation, and support for heterogeneous hardware plat-forms. Extensive evaluations in physical robotic environments demonstrate the protocol’s efficiency in scenarios involving collaborative robotic tasks, distrib-uted sensing, and adaptive control loops. Experimental results confirm that the protocol outperforms conventional solutions in terms fault tolerance, and scala-bility, while maintaining compliance with established real-time constraints. This work contributes to the field of distributed robotics by bridging the gap between flexibility and hard real-time requirements. Beyond robotics, the proto-col can be extended to safety-critical domains such as industrial automation and autonomous vehicles, where reliable, low-latency communication is indispensa-ble.

Keywords: Navigation, Perception & SLAM, Dynamics and Control, Mechatronic systems
Abstract: This paper investigates the integration of path planning and trajectory generation for mobile robots and autonomous vehicles, with a focus on achieving both geo-metric feasibility and compliance with dynamic constraints. The work emphasiz-es the widely used two-layer decomposition: a path planner that computes colli-sion-free, smooth reference curves, and a trajectory planner that time-parameterizes these paths according to velocity, acceleration, jerk, and curvature bounds. Special attention is given to clothoid-based curve design, which offers continuous curvature transitions and improved comfort compared to traditional polynomial or spline methods. The proposed approach not only improves feasi-bility under realistic vehicle dynamics but also ensures computational efficiency suitable for real-time robotic applications. The paper discusses the trade-offs between optimality, robustness, and computational effort, while highlighting the relevance for industrial and autonomous driving contexts. Experimental results demonstrate the effectiveness of the method in producing dynamically feasible and smooth motion profiles, even in constrained environments. Overall, this re-search contributes to the ongoing effort to bridge the gap between geometric planning and dynamic feasibility in robotics, enabling safer, more reliable, and more adaptable autonomous systems.

Keywords: Navigation, Perception & SLAM, Language Models for Robotics, Human Robot Interaction
Abstract: Conventional costmap generation for mobile robots relies primarily on geometric information to classify objects as obstacles or traversable areas. This approach lacks semantic flexibility and fails in scenarios where context-specific actions are required—such as traversing curtains or avoiding puddles. Additionally, traditional methods offer only binary representations of landmarks, limiting their expressiveness.

This study proposes a costmap generation framework guided by natural language instructions. Our system leverages a large language model to classify instructions, detects landmark regions using vision-language models, and assigns navigation costs based on four distinct Action Attribute Class. This enables robots to perform context-aware actions such as approaching, avoiding, or passing through landmarks, with support for both persistent and temporary entities. Experimental results demonstrate that the generated costmaps successfully guide robot paths in accordance with human instructions, offering enhanced flexibility and semantic alignment in autonomous navigation.

Keywords: Dynamics and Control
Abstract: This paper addresses a numerical design of the robust unknown input estimator (UIE) using the artificial bee colony (ABC) algorithm. In the previous study, a simple objective function was proposed and utilized in the ABC algorithm. One of the issues with this objective function is that the number of individuals achieving robust stability decreases during parameter search. A modified and more elaborate objective function is therefore proposed. The effectiveness of the proposed objective function is demonstrated using a numerical example. The result shows that the proposed objective function is effective in alleviating the aforementioned issue.

Keywords: AI & ML & Deep RL, Aerial & Field Robotics, Human Robot Interaction
Abstract: This paper provides a systematic review of current trends and literature gaps around explainable artificial intelligence in flight path management. The Preferred Reporting Items for Systematic reviews and Meta-Analyses framework is used to identify relevant literature, and results are summarised against: the use of artificial intelligence in flight path management, explainable artificial intelligence integration, and metrics to measure explainability. Future research directions are suggested around applicability, trust, and human-machine interaction.

Keywords: Robot Manipulation, Human Robot Interaction
Abstract: Electro-adhesive grippers offer an innovative approach in soft robotics by uti-lizing electrostatic forces to enable the safe and adaptable manipulation of delicate objects. Compared to traditional gripping mechanisms, electro-adhesive grippers operate with minimal noise and low energy consumption, making them suitable for energy-efficient and noise-sensitive robotic appli-cations. This paper provides a review of the working principles, design meth-odologies, material selection, and manufacturing techniques of electro-adhesive grippers. It discusses the influence of key parameters such as dielec-tric properties, electrode patterns, and surface roughness on adhesion per-formance. Furthermore, recent advancements, including multilayer dielectric structures hybrid grasping approaches and the importance of the grasping angle in controlled adhesion and release, are examined to enhance gripping strength and adaptability. Challenges such as high voltage requirements and reliability issues associated with electro-adhesive grasping are also highlight-ed. By synthesizing existing studies, this study aims to serve as a valuable resource for those seeking to develop next generation electroadhesive grippers for various robotic applications.

Keywords: AI & ML & Deep RL, Navigation, Perception & SLAM, Aerial & Field Robotics
Abstract: Robots that navigate walkways, such as delivery robots, are seeing increasing demand in various fields including last-mile delivery and mobility support for the elderly and disabled. However, many conventional robots require manual setup of drivable areas and routes in advance, making flexible response to environmental changes and unknown routes difficult. Furthermore, walkway edge detection using range sensors and drivable area extraction based on semantic segmentation have challenges in accuracy and stability, and deep reinforcement learning methods also face difficulties in stable goal achievement for long-term and complex tasks. This study aims to acquire behavioral policies through deep reinforcement learning that enable flexible navigation while staying within drivable areas such as walkways by giving robots only rough target positions. In particular, we constructed a multimodal world model that integrates camera images and LiDAR information, and achieved versatile behavioral strategies applicable to different starting points and destinations by learning in complex walkway environments including intersections and corners.

Keywords: Advances in Sensor Technology, AI & ML & Deep RL, Navigation, Perception & SLAM
Abstract: Robotics and automation have revolutionized aircraft Maintenance, Repair, and Overhaul (MRO) operations by streamlining tasks, reducing costs, and enhancing overall accuracy. Despite these advancements, effective localiza-tion and navigation of mobile robots in hangar environments remain chal-lenging due to complex environmental structures. Traditional approaches to mobile robot localization and navigation primarily rely on onboard sensor systems, including inertial measurement units (IMUs), wheel odometry, and LiDAR sensors. While these systems have proven effective in many applica-tions, they face significant limitations particularly in challenging industrial environments such as aircraft hangars that is characterized by sparse fea-tures, reflective surfaces, and dynamic obstacles. The proposed system integrates a custom-trained You Only Look Once (YOLOv8) deep learning model for real-time visual detection combined with pose estimation algorithms leveraging external camera networks to provide real-time, high-precision localization for a Panther mobile robot. The exper-imental results show that our deep learning-based visual tracking approach provides consistent and accurate localization data, with detection confidence scores exceeding 85% in complex environments and pose estimation errors remaining below 5cm in most operational conditions. The proposed method-ology offers a scalable solution for enhancing robot autonomy in GPS-denied environments and provides a foundation for future developments in intelli-gent industrial automation systems.

Keywords: Language Models for Robotics, AI & ML & Deep RL
Abstract: As robotic systems become increasingly connected and intelligent, their exposure to cyber threats likewise grows. Traditional signature-based anomaly detection often fails to identify emerging or previously unseen attacks. We present Vision–Language Anomaly Detection for Robots (VLADRo), a ROS2-based framework that leverages fine-tuned vision–language models (VLMs). VLADRo aggregates multi-source—network traffic, CPU utilization, and ROS2 topic subscription in-tervals—and renders it as standardized visual graphs for VLM inference. Unlike our earlier version, we rebuild the dataset in a realistic environment and redesign the prompt-learning pipeline: visualizations adopt fixed axes and reference lines with uniform styling, and prompts follow a constrained, structured format to im-prove interpretability and time localization. Using the new dataset and prompt tuning, a multi-image VLM (LLaVA-interleave-qwen-7b) attains 94.2% anomaly-detection accuracy and 70.8% explanation accuracy (onset-time estimation and attack classification) across seven attack scenarios, outperforming our previous configuration. These results underscore the potential of VLM-based, interpreta-ble, data-driven anomaly detection for strengthening the cybersecurity of ROS2 robotic systems.

Keywords: AI & ML & Deep RL, Multi-Robot Systems
Abstract: Robotic teams, such as fleets of autonomous vehicles and swarms of drones, often operate in highly dynamic and uncertain environments. Unexpected events, such as robot failures or the addition of new robots during a mission, require rapid adaptation, making intelligent and efficient task reallocation particularly challenging. This paper addresses this problem, specifically addressing supply-side changes (e.g., robot failures or new robots joining the team) during operation. Unlike most existing research that focuses on demand-side dynamics, we explicitly model supply-side events by formulating the Vehicle Routing Problem with Dynamic Fleet Changes (VRP-DFC). We formulate this problem as a Markov decision process and develop an adaptive deep reinforcement learning (DRL) framework based on the Transformer architecture. Our approach encodes fleet and task states, enabling real-time adaptation to fleet composition changes and ensuring efficient task reallocation. Experimental results demonstrate that our approach outperforms state-of-the-art heuristics and DRL baselines in terms of travel distance, task completion rate, and computational efficiency, particularly in large-scale and highly dynamic scenarios. This work provides a robust and scalable solution for intelligent task management and real-time decision-making in real-world robotic systems operating under uncertainty.

Keywords: AI & ML & Deep RL, Multi-Robot Systems
Abstract: The increasing deployment of robots and unmanned systems in areas such as intelligent warehousing, logistics, environmental monitoring, and smart agriculture has created new challenges in efficiently assigning tasks and planning routes for diverse fleets. In practice, commercial organizations often assemble temporary fleets by renting heterogeneous robotic devices to minimize both fixed hiring costs and variable operational expenses, rather than investing in permanent ownership. This scenario is modeled as the Fleet Size and Mix Vehicle Routing Problem (FSMVRP), which requires determining the optimal combination of fleet composition and routing to achieve cost-effective operations. To deal with the computational complexity of FSMVRP, especially in large-scale instances, we present a deep reinforcement learning (DRL)-based approach that formulates the problem as a Markov Decision Process and leverages an encoder-decoder policy network with specialized vehicle embeddings. Our approach integrates fleet construction into the routing process and introduces a left graph embedding to guide efficient vehicle selection. Extensive experiments demonstrate that the proposed approach consistently delivers high-quality solutions within seconds, outperforms existing algorithms in efficiency and scalability, and generalizes well to complex, cost-sensitive robotic deployment scenarios.

Keywords: AI & ML & Deep RL, Risk-Aware Autonomy, Safe Decision Making under Uncertainty
Abstract: Despite the remarkable success of deep reinforcement learning (RL) across various domains, its deployment in the real world remains limited due to safety concerns. To address this challenge, constrained RL has been extensively studied as an approach for learning safe policies while maintaining performance. However, since constrained RL enforces constraints in the form of cumulative costs, it cannot guarantee state-wise safety. In this paper, we extend the Lagrangian based approach, a representative method in constrained RL, by introducing state-dependent Lagrange multipliers so that the policy is trained to account for state-wise safety. Our results show that the proposed method enables more fine-grained specification of the constraints and allows the policy to satisfy them more effectively by employing state-dependent Lagrange multipliers instead of a single scalar multiplier.

Keywords: Navigation, Perception & SLAM, Aerial & Field Robotics, Advances in Sensor Technology
Abstract: This paper proposes Poly-Pole, a polytunnel pole association-based SLAM algorithm that provides a cost-effective solution for enhancing odometry accuracy in polytunnel environments. The most reliable and distinguishable features can be consistently detected in polytunnels are the metal supporting poles. However, their visually identical appearance and highly regular spatial arrangement make data association extremely challenging. In the proposed approach, we assume that the operator has a coarse prior knowledge of the tunnel’s structural pattern, such as the number of rows and columns and the type of distinct gaps between poles. Building upon this prior, a Expectation–Maximisation (EM) framework is employed to iteratively perform robust pole data association and refinement of the pattern parameters. The refined associations are subsequently incorporated into a factor graph, along with pole measurements, odometry, and pattern-generated landmark priors, to enable global-level optimisation. The effectiveness of the proposed algorithm was validated through extensive real-world experiments conducted in commercial polytunnel environments. For this purpose, an Antobot robot platform equipped with an Ouster OS0 LiDAR, a MicroStrain 3DM-GX5 IMU, wheel odometry, and RTK-GNSS was deployed. The robot trajectories were designed to emulate typical farming operations, where the robot traversed each row sequentially before moving to the next—minimising path overlap while maximising the distance covered per row. Experimental results demonstrate that Poly-Pole SLAM effectively mitigates odometry drift and enhances trajectory consistency, achieving notable improvements in localisation accuracy with minimal computational overhead.

Keywords: Dynamics and Control, Aerial & Field Robotics
Abstract: Although many drone attitude control algorithms have been presented, few of them ensure consistent performance when maneuvers require simultaneous satisfaction of both state and actuator limitations.While exceeding actuator torque constraints might result in instability, violating orientation safety restrictions can result in the loss of mission goals. In aerial robots, it is still very difficult to guarantee that both constraints are satisfied in real time without compromising tracking accuracy. In this paper a novel control framework is presented that integrates a Barrier Lyapunov function for continuous enforcement of quaternion-based orientation constraints with a dynamic reference modification mechanism for handling actuator saturation. The barrier formulation ensures that the attitude error always stays within a safe set while reference modification reshapes the prescribed angular velocity whenever torque demands approach their limitations. A detailed stability and error convergence study is provided along with the derivations of the suggested control law. Furthermore, numerical simulations demonstrate that the controller achieves high tracking accuracy while consistently adhering to input and state constraints.

Keywords: Dynamics and Control, AI & ML & Deep RL
Abstract: Approximate model-predictive control (AMPC) aims to imitate an MPC’s behavior with a neural network, removing the need to solve an expensive optimization problem at runtime. However, during deployment, the parameters of the underlying MPC must usually be fine-tuned. This often renders AMPC impractical as it requires repeatedly generating a new dataset and retraining the neural network. Recent work addresses this problem by adapting AMPC without retraining using approximated sensitivities of the MPC’s optimization problem. Currently, this adaption must be done by hand, which is labor-intensive and can be unintuitive for high-dimensional systems. To solve this issue, we propose using Bayesian optimization to tune the parameters of AMPC policies based on experimental data. By combining model-based control with direct and local learning, our approach achieves superior performance to nominal AMPC on hardware, with minimal experimentation. This allows automatic and data-efficient adaptation of AMPC to new system instances and fine-tuning to cost functions that are difficult to directly implement in MPC. We demonstrate the proposed method in hardware experiments for the swing-up maneuver on an inverted cartpole and yaw control of an under-actuated balancing unicycle robot, a challenging control problem.

Keywords: Dynamics and Control, AI & ML & Deep RL
Abstract: There exist numerous ways of representing 3D orientations. Each representation has both limitations and unique features. Choosing the best representation for one task is often a difficult chore, and there exist conflicting opinions on which representation is better suited for a set of family of tasks. Even worse, when dealing with scenarios where we need to learn or optimize functions with orientations as inputs and/or outputs, the set of possibilities (representations, loss functions, etc.) is even larger and it is not easy to decide what is best for each scenario. In this paper, we attempt to a) present clearly, concisely and with unified notation all available representations, and "tricks" related to 3D orientations (including Lie Group algebra), and b) benchmark them in representative scenarios. The first part feels like it is missing from the robotics literature as one has to read many different textbooks and papers in order have a concise and clear understanding of all possibilities, while the benchmark is necessary in order to come up with recommendations based on empirical evidence. More precisely, we experiment with the following settings that attempt to cover most widely used scenarios in robotics: 1) direct optimization, 2) imitation/supervised learning with a neural network controller, 3) reinforcement learning, and 4) trajectory optimization using differential dynamic programming. We finally provide guidelines depending on the scenario, and make available a reference implementation.

Keywords: AI & ML & Deep RL, Dynamics and Control
Abstract: The fidelity of the aerodynamic model is a decisive factor in the control performance of a missile. This study proposes a framework to precisely correct inaccurate missile aerodynamic models by maximizing the utilization of information from limited flight-test data. The proposed method employs a Dynamics-Embedded Neural Ordinary Differential Equation to decouple the known rigid-body dynamics from the unknown aerodynamic coefficients. By embedding known physics, the proposed method avoids the black-box nature of purely data-driven methods and ensures greater physical consistency. These coefficients are then effectively identified using a noise-robust integral loss function combined with a two-stage training strategy involving pre-training and fine-tuning. Numerical simulations demonstrate that the proposed framework success fully corrects the initial model’s inaccuracies and estimates the true aerodynamic force coefficients with high precision, using limited flight data. The results indicate that the proposed framework can be a promising tool for improving the fidelity of aerodynamic models in missile development, thus improving overall system performance.

Keywords: Dynamics and Control
Abstract: This study explores the applicability of Automatic Differentiation (AD) to missile guidance at multivariable sensitivity analysis and real-time navigation constant estimation. In three-dimensional environment, first, parameter sensitivities of terminal miss distance are evaluated using AD. Compared to numerical differentiation, AD-based gradients are faster and more accurate, enabling identification of relative parameter influence. This demonstrates the feasibility of quantifying parameter impacts on terminal performance in high-dimensional systems. Second, an online estimation framework for guidance parameter is implemented using AD gradient-based optimization. In scenarios with stationary and constant velocity targets, guidance parameter is updated in real-time to minimize cost function. Simulations confirmed that miss distance is reduced through adaptive updates, while computation times remained feasible for real time. Overall, the results verify that AD can be applied to missile guidance, providing sensitivity evaluation and enabling online parameter estimation. This study represents an initial step toward broader AD use in guidance and control applications.

Keywords: Human Robot Interaction, Healthcare & Assistive Robotics
Abstract: Many children experience challenges in emotional regulation and social interaction, which can limit their participation in everyday activities and therapeutic programs. For socially assistive robots to be effective in this context, it is essential that children remain consistently and meaningfully engaged. We explore engagement strategies for a tactile robot designed to support children suffering from anxiety through daily interactions. The robot delivers either synthetic emotional feedback or point rewards to encourage user participation. We evaluated these strategies through two studies: a preference assessment with 16 school children aged 6–8 years, and a behavioral study with 14 university students aged 20–27 years in naturalistic environments. The study with school children indicated a preference for emotional engagement over points-based approaches. The follow up study with university students across a full day of interactions revealed contrasting results: points-based systems produced significantly higher task accuracy (p < 0.05) and sustained performance over time. Findings from different user groups suggest that stated preferences and behavioral outcomes can diverge depending on engagement context, highlighting the importance of validating design assumptions through observed interaction. This work contributes insights into age-related differences in engagement strategy effectiveness in human-robot interaction design.

Keywords: Human Robot Interaction, Robot Manipulation
Abstract: Human-robot collaboration in manufacturing settings requires robots to exhibit seamless human-like actions. To achieve this ultimate goal, robots need to be able to imitate human movements as well as generalise learned behaviours to novel scenarios. In this work, we propose a layered cognitive architecture to orchestrate the process of collaborative robots (cobots) learning from human guidance, reproducing the manipulation skills with the visual feedback in assembly tasks. This architecture is composed of somatic, reactive, adaptive, and contextual layers, enabling the robot to acquire skills through demonstration and adaptively generate new motion trajectories. In the learning phase, the force-torque sensor captures the human's kinesthetic guidance, which the admittance controller transforms into the robot's movements. The demonstrated movements are generalised into a motion model using Gaussian Mixture Models (GMMs). In the reproduction phase, ``unseen'' trajectories are generated by recalling the GMM associated with the target point. This approach is validated in a pilot human-robot collaborative assembly task using the UR3 robot. The results demonstrate the system’s capability to generate unseen trajectories, adapt to varying object locations and different action sequences. The performance of the admittance controller during the learning phase is assessed, and the trajectory reproduced in the execution phase is presented. This work demonstrates the potential of the proposed framework for human-robot collaboration tasks.

Keywords: Human Robot Interaction, Robot Manipulation, Navigation, Perception & SLAM
Abstract: High-mix low-volume manufacturing poses significant challenges for automation, as frequent product changes and short production runs make traditional robot programming inefficient and costly. Small and medium-sized enterprises are particularly affected, since existing solutions often lack the flexibility to handle complex machine tending tasks. This paper introduces TendAssist, a prototype system that enables the rapid configuration of robotic tending processes within minutes. The approach combines a user-friendly configuration wizard, machine learning-based object detection, and intuitive robot interaction, allowing operators without robot programming experience to adapt workflows quickly. Industrial use cases, such as handling aluminum and steel profiles or reflective metal plates, demonstrate the system’s ability to manage diverse parts and difficult stacking patterns. Initial experiments show that, while first-time use may take longer, with minimal training operators can consistently achieve the targeted setup times of under five minutes, thereby lowering barriers to automation and enabling more flexible, efficient production. The most significant constraint is the occasional need to adapt the object detection model to previously unseen objects, a process that involves the time-consuming collection and labeling of training images.

Keywords: Navigation, Perception & SLAM, AI & ML & Deep RL, Human Robot Interaction
Abstract: This study addresses autonomous navigation of mobile robots in crowded environments with pedestrian flows. Conventional collision avoidance methods, such as VO, and RVO, assume perfect knowledge of surrounding agents and focus solely on pairwise avoidance, making flow-consistent and socially harmonious navigation difficult in real-world settings. Deep reinforcement learning (DRL) enables the learning of complex interactions; however, few approaches explicitly consider pedestrian flows, and many still rely on precise pedestrian observations. In this work, we propose a DRL-based model that directly uses 2D LiDAR distance measurements without requiring explicit pedestrian observations and incorporates a flow-consistency reward based on the cosine similarity between the robot's and pedestrian's goal directions to achieve socially harmonious navigation. Simulation results demonstrate that the proposed method enables the robot to avoid collisions while moving along pedestrian flows, and that the inclusion of the flow-consistency reward significantly improves alignment with pedestrian movement.

Keywords: Language Models for Robotics, Human Robot Interaction, AI & ML & Deep RL
Abstract: Human motivation refers to the internal drive that compels an individual to act and engage in goal-directed behavior aimed at achieving specific objectives or fulfilling needs. Motivation is closely linked to the learning process, as learning itself can be a goal in which an individual seeks to develop new skills, gain knowledge, or master new skills. Even when learning is not the primary goal, goal-directed behavior can still lead to the acquisition of new information and capabilities. In this work, we investigate how motivation can support and facilitate language learning, using a computational language robot model. This developmental model of early language acquisition is inspired by how human infants learn language and includes a motivational module, a perception module, and a communication/action module. The robot learns actively by relying on its perception, social interaction with a caregiver, and online learning to acquire motivation-grounded language. We implemented this architecture in a humanoid robot to study the development of its communicative skills. We aim to follow major milestones in language learning, such as babbling and lexical development (learning nouns and verbs). The results indicate that the robot successfully learned basic nouns and verbs and can also distinguish between them, while acquiring the pragmatic aspects of the language. The findings further show that the nature of the caregiver’s responsiveness has a significant influence on the language the robot acquires.