Keywords: Artificial Intelligence Systems, Robotic Applications, Biomedical Instruments and Systems
Abstract: Robotic grippers relying on fixed or PID-based control strategies fail to accommodate variations in object stiffness, leading to premature slip or crush damage. Inspired by force-dependent molecular clutch dynamics in biological cell adhesion, we developed a physics-based simulation framework integrating Bell-model clutch kinetics, stochastic Monte Carlo state transitions, and a spring-in-series gripper model. Mechanical features extracted from simulated interactions with objects drawn from the Yale-CMU-Berkeley (YCB) benchmark were partitioned into three mechanically distinct clusters using K-means, enabling cluster-specific optimal grip-force assignment without labelled object models. Across 20,000 Monte Carlo trials, the adaptive controller achieved a significantly longer mean time-to-slip compared to fixed-force control (Δ = 0.073 s; 95% bootstrap CI: [0.048, 0.099] s; permutation p = 5×10⁻⁵) and outperformed PID control (Δ = 0.031 s; CI: [0.001, 0.061] s; p = 0.046). These results demonstrate that bio-inspired stochastic mechanics paired with unsupervised classification can improve grip stability under object uncertainty, providing a principled framework for extending adaptive control to broader robotic manipulation tasks.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems, Industrial Applications of Control
Abstract: Recent advances in Vision-Language Models (VLMs) are opening new possibilities in the field of Driver Monitoring Systems (DMS). Conventional DMS approaches combine individual modules such as facial landmark detection, head pose estimation, and gaze tracking in a pipeline architecture, making them vulnerable to environmental variations like occlusion and backlighting, while also suffering from inter-module error propagation. To overcome these limitations, this paper proposes a real-time driver monitoring system for embedded environments that combines LoRA (Low-Rank Adaptation) fine-tuning and INT8 symmetric quantization based on Qwen3-VL-2B, a lightweight VLM. The proposed system performs LoRA fine-tuning on a dataset specialized for infrared camera imagery to adapt the model to the in-vehicle domain, while significantly reducing model size and computational cost through INT8 quantization. Experimental results on the NVIDIA Jetson AGX Orin platform demonstrate that the proposed model achieves approximately a 2.3× inference speedup (11.5 FPS) and a 44.2% reduction in GPU memory usage compared to the original FP16 model, while maintaining practical recognition performance with an average accuracy loss of less than 1.3% achieving 94.2% for drowsiness detection, 91.8% for gaze deviation detection, 95.5% for phone usage detection, and 91.0% for eating/drinking detection.

Keywords: Artificial Intelligence Systems, Navigation, Guidance and Control
Abstract: Robots exploring indoor environments often receive only partial observations of the surrounding scene, leav- ing unobserved objects and spatial relations unknown. We propose LatentSGC, a latent flow matching framework for uncertainty-aware 3D scene graph completion from partial robot observations. LatentSGC first uses a Graph Autoencoder to encode scene graphs into per-node latent slots, and then applies Conditional Flow Matching to complete the unobserved slots by integrating a learned ODE conditioned on the observed subgraph. Since each inference run starts from indepen- dent Gaussian noise, multiple samples produce diverse yet plausible completions, providing a probabilistic belief over hidden room structures rather than a single deterministic prediction. On a benchmark derived from HM3D-Semantics scenes at 30%, 50%, and 70% observation ratios, LatentSGC achieves Node F1 of 0.432 and Edge F1 of 0.255, outper- forming deterministic baselines by 0.187 in Node F1 and 0.166 in Edge F1. We further demonstrate the integration of LatentSGC with an RGB-D perception pipeline in Habitat Simulator on rooms unseen during training, where it produces plausible completions of unobserved room structures.

Keywords: Artificial Intelligence Systems, Sensors and Signal Processing
Abstract: Scattering media such as fog and smoke introduces atmospheric degradation that significantly impairs visual contrast and spatial details in outdoor image. Although convolutional neural networks (CNNs), such as DehazeNet, effectively estimate localized transmission maps via physical atmospheric scattering models, they frequently fail to restore consistent structural data under dense and non-uniform haze conditions. To address these challenges, this paper proposes a novel end-to-end hybrid architecture that concurrently leverages local and global features. The proposed method applies to a symmetrical U-Net pathway to preserve sharp object contours and fine-grained textures alongside a parallel vision transformer (ViT) to model long-range environmental dependencies across the entire scene. These multi-scale representations are adaptively unified through a Cross-Attention mechanism. Furthermore, an uncertainty map is systematically generated to dynamically guide iterative training phases. Through quantitative and qualitative evaluations conducted on the Dense-haze dataset, we demonstrate that our algorithm substantially outperforms conventional benchmarks across gradient magnitude similarity deviation (GMSD), learned perceptual image patch similarity (LPIPS), and CIEDE2000 metrics.

Keywords: Artificial Intelligence Systems, Control Theory and Applications, Robot Mechanism and Control
Abstract: Although large language models (LLMs) provide a promising interface for generating executable reward code, directly applying them to high-dimensional humanoid locomotion can be fragile, as a candidate that improves a scalar score may still induce undesirable behaviors such as excessive action oscillation or unstable foot placement. To address this issue, we propose Evidence-Gated Pareto Reflection (EGPR), a generation-based LLM-assisted reward evolution framework for humanoid locomotion. EGPR treats the LLM as a bounded reward-hypothesis generator and evaluates each candidate through structured training evidence, deterministic validation, and multi-objective Pareto reflection. Candidate rewards are materialized as separate local modules without modifying the original baseline reward implementation, and a deterministic validation gate filters out invalid or unsafe candidates before full training. After each generation, EGPR preserves Pareto-front candidates, metric-specific strengths, and diagnostic failure modes as structured evidence to guide the next generation of reward hypotheses. We validate EGPR on a 29-degree-of-freedom (DoF) Unitree G1 humanoid robot simulation environment. Experimental results demonstrate that EGPR successfully discovers balanced and transferable reward functions, significantly improving performance over existing approaches by reducing velocity tracking errors, improving energy efficiency (cost of transport), and enhancing control smoothness.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Process Control Systems
Abstract: We present a hybrid vision inspection framework that couples YOLO-Seg instance segmentation with PatchCore anomaly detection for industrial quality control under spatial variability of components. Pixel-level segmentation masks define an inspection region that reduces contamination from the variable component, while a hierarchical decision rule simultaneously enforces component presence and surface integrity. We evaluate the system through (i) a seven-configuration ablation that probes the contributions of YOLO, the PatchCore memory bank, the bounding-box and segmentation-mask regions of interest (ROIs), and the hierarchical gate; (ii) a direct comparison of cosine and L2 distances under an identical memory bank with a scale-invariance analysis; (iii) standard binary-classification metrics (AUROC, precision, recall, and F1) alongside the normal-score range; and (iv) a backbone comparison that replaces ResNet-18 with the DINOv2-S vision transformer. On a 27-image exploratory rice-cooker evaluation corpus, the proposed cosine configuration achieves an accuracy of 0.741, F1 of 0.741, and AUROC of 0.929. It detects 10/17 FAIL samples while accepting all 10 PASS samples. Cosine and L2 yield identical rankings within matched ROIs, whereas native-scale calibration produces different thresholded operating points. Pipeline inference latency is 22.1 ms per image on an RTX 3090 GPU (45.3 FPS) and 341.4 ms on a single CPU thread. These observations require independent validation before autonomous release.

Keywords: Artificial Intelligence Systems, Multimedia Systems, Robot Vision
Abstract: Recent advancements in generative modeling have seen a paradigm shift from convolutional U-Net architectures to transformer-based backbones, most notably with the introduction of the Multi-modal Diffusion Transformer (MMDiT). While these models demonstrate unprecedented scale and high-fidelity synthesis, their capacity for precise, localized, and context-aware text-guided image editing remains underexplored. In this paper, we present a rigorous analysis of the MMDiT architecture, specifically investigating how its symmetric, dual-stream attention mechanism facilitates superior controllability compared to traditional cross-attention-based diffusion models. By leveraging the bidirectional information flow between text and visual latents, we demonstrate that MMDiT provides a more granular control surface for localized object manipulation and attribute binding. Furthermore, we conduct a comprehensive layer-wise ablation study on the FLUX.1 architecture using the PIE-Bench++ dataset to identify critical "bottleneck" blocks that govern semantic alignment. Our quantitative findings reveal a non-uniform distribution of structural importance, where specific mid-network blocks act as primary anchors for text-image grounding. Building on these insights, we propose a set of architectural refinements that enhance editing precision without requiring extensive retraining. Experimental results indicate that our approach significantly outperforms baseline diffusion-transformer models in both semantic consistency and structural preservation, providing a robust framework for high-precision, text-driven image manipulation.

Keywords: Artificial Intelligence Systems, Robot Vision, Robotic Applications
Abstract: Feed-forward 3D Gaussian Splatting has recently emerged as an efficient alternative to optimization-based 3DGS by reconstructing Gaussian scene representations directly from input images. While these methods substantially reduce inference cost, their predicted appearance attributes can remain imperfect, resulting in localized rendering errors even when the underlying scene geometry is reasonably accurate. In this work, we propose a lightweight refinement adaptor that improves the appearance of feed-forward Gaussian representations without altering their geometry. Starting from a frozen Gaussian predictor, our adaptor combines image features with rendered normal cues and camera-ray information to estimate residual updates for opacity and spherical harmonic coefficients. To avoid indiscriminate modification of all Gaussians, the adaptor additionally predicts a confidence value that controls the magnitude of each residual update. This allows the model to focus refinement on uncertain or poorly rendered regions while preserving reliable predictions. The adaptor is optimized using only image reconstruction objectives, requiring no additional supervision for geometry, camera parameters, or Gaussian regularization. Experiments on DL3DV under multiple input-view and camera calibration settings demonstrate consistent improvements over the original feed-forward model as well as ungated residual refinement. The results suggest that selectively correcting appearance attributes is an effective and computationally lightweight strategy for enhancing feed-forward 3DGS.

Keywords: Artificial Intelligence Systems, Robot Vision, Sensors and Signal Processing
Abstract: Robust feature matching is the bottleneck for sonar-based underwater localization and mapping, a core ca- pability for inspecting submerged structures, where strong speckle noise and the scarcity of labeled real-world data limit learning-based methods. We present a contrastively learned (InfoNCE) patch descriptor for forward-looking sonar that is trained entirely in simulation, combined with a saliency-guided keypoint detector that exploits structure-tensor saliency to suppress speckle alongside a CFAR-style blob detector. Matched correspondences are verified with a Cartesian (fan) rigid model and passed to acoustic structure-from-motion (ASFM) for trajectory estimation. Using no real data and no fine-tuning, our descriptor transfers zero-shot to two real sonars—an open-water river survey and a water-tank experiment. On identical keypoints it substantially outperforms the hand-crafted AKAZE and ORB and the learning-based SONIC in simulation, remains on par with or better than these baselines on the real data, and, in trajectory estimation, achieves the lowest ATE on both sequences and the lowest RPE on the most challenging one.

Keywords: Artificial Intelligence Systems, Robot Vision
Abstract: Multi-task dense prediction commonly assumes that all task annotations are jointly available for the same images. In practice, however, labels for semantic segmentation, depth estimation, and surface normal prediction are often collected asynchronously because they require different sensors, annotation pipelines, and costs. This motivates a task- incremental heterogeneous dense prediction setting, where a model sequentially acquires dense prediction tasks while retaining previously learned capabilities. In this paper, we study this setting under a parameter-efficient constraint: a frozen vision transformer backbone is equipped with a single task-shared LoRA module, while task-specific heads handle heterogeneous output spaces. Although this fixed-capacity shared adapter avoids task-wise parameter growth, standard LoRA provides only a linear low-rank residual, which can limit its expressiveness when adapting to heterogeneous dense prediction tasks. To improve the expressive capacity of the shared adapter without introducing task-specific adapter parameters, we propose Polynomially Calibrated LoRA, which augments the LoRA bottleneck with global polynomial activation calibration. Experiments on NYUv2 show that polynomial calibration provides a competitive nonlinear shared- adapter baseline for task-incremental dense prediction.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems
Abstract: Autonomous locomotive operation in industrial rail logistics faces significant safety challenges, particularly in safety blind spots and during pushing operations where onboard sensors have limited visibility. To address these issues, this paper proposes a collaborative perception framework using vision AI-based smart CCTV infrastructure in an integrated steel plant. The proposed framework integrates infrastructure-level perception with a centralized control system to support safe autonomous operation of molten iron transport locomotives between blast furnace and steelmaking plants. Multiple vision AI modules monitor the operational environment, including obstacle detection on railway tracks, signal recognition, and torpedo ladle car (TLC) status monitoring. The perception system combines deep learning-based object detection models with conventional computer vision algorithms to achieve reliable monitoring in complex industrial environments. To coordinate multiple locomotives, the centralized control system communicates with distributed vision AI servers through a TCP socket–based network. Experimental results using CCTV datasets collected from blast furnace and steelmaking plants demonstrate high recognition performance of the proposed modules. These results confirm that the proposed framework effectively reduces safety blind spots and enables safe autonomous locomotive operation in steel plant environments.

Keywords: Artificial Intelligence Systems, Sensors and Signal Processing, Control Devices and Instruments
Abstract: Vehicle weight is a key hidden state in heavy-duty trucks because it affects longitudinal response, braking behavior, energy consumption, and vehicle-control performance. This study investigates a CAN-only weight estimation framework using real-vehicle driving data collected across different payload and speed conditions. The framework is designed as a dual-phase analysis, an AI phase and a vehicle-dynamics phase. The AI phase estimates vehicle weight from CAN time-series signals, while the vehicle-dynamics phase examines load-dependent vehicle responses from a physical perspective. Both phases use the same preprocessed CAN signal database, but they construct different input representations. In the AI phase, asynchronous CAN signals are synchronized at 50 Hz and converted into short-term multivariate windows. A lightweight MLP-Mixer is then used to learn temporal patterns and feature-wise relationships among selected CAN features. The dynamics analysis further supports that the selected CAN signals are physically meaningful and can complement the AI-based estimation phase. The results show that CAN time-series signals contain useful information for onboard weight estimation without additional load sensors or roadside weighing infrastructure.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Robotic Applications
Abstract: Parts kitting --- arranging components in a tray for downstream assembly --- in manufacturing is a difficult task to automate due to the requirement that parts arrive in arbitrary poses and must be placed precisely into designated tray. We propose a closed-loop kitting system that performs manipulation with a vision-language-action (VLA) policy and integrates a vision-language model (VLM)-based task planner and manager as a supervisory upper layer. The task planner decomposes a high-level kitting command into a sequence of subtasks executable by the VLA policy, while the task manager monitors the environment during execution, detects failures, and triggers re-planning by the planner. The two components share a single VLM, conforming to the resource constraints of the edge node. In addition, we compare candidate action/state representations for the manipulation policy under identical conditions and confirm that the 6D rotation representation is most effective for motions involving large wrist rotation behavior. In real-robot evaluation, the proposed VLM-supervised closed-loop system improves the overall kitting success rate from 60% to 85% over a standalone VLA policy, demonstrating that VLM supervision can effectively complement the silent failures of a single VLA policy.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems, Sensors and Signal Processing
Abstract: 3D reconstruction of non-cooperative space targets is a key technology for on-orbit servicing, where reconstruction methods must remain reliable under substantial viewpoint and illumination changes. This study comparatively evaluates Instant-NGP, 3DGS, 2DGS, GOF, and NVDiffRec using a synthetic multi-view benchmark comprising three controlled observation trajectories: horizontal circular observation around the target (Case 1), top-view circular observation (Case 2), and an approach-type spiral trajectory (Case 3). Four illumination conditions were applied to each trajectory. Performance was assessed using PSNR, SSIM, LPIPS, and rendering speed. 3DGS achieved the highest overall reconstruction fidelity, with an average PSNR of 53.71 dB, SSIM of 0.9975, and LPIPS of 0.0052. Its PSNR values were 52.83, 52.15, and 56.15 dB for Cases 1–3, respectively. The maximum PSNR difference was limited to 4.00 dB, indicating comparatively stable reconstruction quality across trajectory changes. However, its rendering speed decreased from 21.83 FPS in Case 1 to 13.33 FPS in Case 3, a reduction of approximately 38.9%, revealing a trade-off between reconstruction fidelity and computational efficiency. Excluding 3DGS, NVDiffRec showed the most consistent performance across trajectories. Its PSNR ranged from 40.79 to 45.09 dB, with a maximum difference of 4.30 dB, while SSIM remained within 0.9899–0.9944 and rendering speed within 22.57–23.53 FPS. Its maximum FPS variation was only approximately 4.1%, indicating stable computational performance across trajectory changes. In contrast, 2DGS and GOF showed PSNR reductions of 14.01 dB and 14.73 dB, respectively, in Case 3 relative to their best-performing cases, demonstrating greater sensitivity to the approach-type spiral trajectory. These results indicate that reconstruction quality and computational stability depend strongly on both the reconstruction method and observation geometry.

Keywords: Artificial Intelligence Systems, Robot Vision, Robotic Applications
Abstract: Visuomotor policies have become a prominent approach for robot manipulation, mapping observations directly to control actions. However, their performance degrades substantially under out-of-distribution conditions, such as novel object positions or changes in camera viewpoint. While prior works address this by scaling up demonstration datasets, such approaches compromise data efficiency. This paper presents a data-efficient visuomotor policy framework that leverages active vision. The proposed method employs an open-vocabulary object detector and represents the robot state and actions in the camera coordinate frame. By actively repositioning the camera to align the perceived object position with the training distribution, the policy can manipulate objects at previously unseen locations without additional data collection. Experiments on a humanoid robot and a dual-manipulator system, in both simulation and the real world, show that the proposed framework substantially outperforms state-of-the-art baselines, raising the Pick-and-Place task success rate in out-of-distribution region from 56% to 88%.

Keywords: Artificial Intelligence Systems, Multimedia Systems, Robot Vision
Abstract: Image retrieval has largely benefited from compact global descriptors and learned reranking, yet distinguishing hard candidates remains challenging when images share repeated structures, similar layouts, or partial visual overlap. In this work, we revisit cross-view attention from pretrained cross-view completion models as a source of pairwise evidence for image retrieval. While recent studies show that such attention maps contain rich correspondence cues, they are not directly optimized for retrieval scoring and often include noisy or ambiguous activations. We propose CVAR, a Cross-View Attention-based Reranker that transforms raw cross-attention maps into discriminative query-candidate similarity scores. CVAR extracts cross-attention maps from a cross-view completion backbone and refines them using ARM, a 4D transformer-based Attention Map Refinement Module that preserves the joint spatial structure of the query and candidate token grids. ARM denoises layer-wise attention responses, aggregates correspondence patterns across views, and outputs an image-level score for reranking. We train CVAR with symmetric pairwise supervision and hard-negative sampling. Experiments on ROxford5k and RParis6k show that CVAR improves both local-feature and global-feature rerankers, especially under Hard and 1M-distractor settings.

Keywords: Artificial Intelligence Systems, Civil and Urban Control Systems
Abstract: Large language models (LLMs) are increasingly considered as high-level supervisory modules for cyber physical and energy systems. However, semantic alignment with a user prompt does not necessarily imply safe closed loop behavior when LLM outputs affect control-relevant decisions. This paper studies this issue in a multi-energy system (MES) supervisory control setting, where an LLM selects blending weights over economic, emission-oriented, and comfort-oriented model predictive control (MPC) experts. We compare prompt sanitization, runtime safe-floor correction, and their combined guardrail variant. Prompt sanitization rewrites risky or contradictory requests, whereas safe-floor correction directly enforces a minimum comfort-controller weight under safety-sensitive conditions. Experiments are conducted using five policies, twelve prompts, six scenarios, and three repeated trials, resulting in 1080 closed-loop runs. Results show that the rule-based supervisor achieves the lowest cost and CO2 ratios but exhibits larger comfort error, hard thermal violation hours, and safety-filter intervention rate. Among the LLM-family policies, LLM+Safe-floor achieves the strongest comfort-related safety behavior. In contrast, prompt sanitization improves alignment but does not consistently improve runtime safety.

Keywords: Artificial Intelligence Systems, Robot Vision, Sensors and Signal Processing
Abstract: The rotating brushes of an automatic vehicle wash can damage protruding exterior parts, so such attachments must be detected before each cleaning cycle. Detectors based on deep learning, however, perform poorly on attachments that are uncommon in real wash data, producing missed and false detections because the scarce examples are insufficient to capture their visual variation. We note that vehicle components keep a consistent spatial arrangement when viewed from the entry camera, since each body region reliably hosts the same parts across vehicles. Motivated by this regularity, we couple a YOLOv11 backbone with hypergraph-based relational reasoning to counter the data scarcity. The proposed viewpoint-prior hypergraph transformer module models semantic relations among attachments under each camera viewpoint, letting the detector rely on learned co-occurrence when a part's visual signal is weak. Tested on images gathered at an operating wash facility, the method reaches 71.3% mAP50--95 and exceeds D-FINE and a recent graph-based detector by 3.8 and 1.4 points.These findings suggest that reasoning over viewpoint-aware hypergraphs yields dependable detection when training data are limited, contributing to safer unattended vehicle washing.

Keywords: Artificial Intelligence Systems, Robot Vision
Abstract: Accurate crowd people counting in physical space is essential for practical crowd safety management systems. Most of crowd counting methods estimate density maps on an image plane that projects a 3D space onto 2D plane, so crowd counting or density estimates are distorted. In this paper, considering statistical characteristics in 3D space is proposed. The proposed crowd people counting consists of two stages: a crowd people counting stage that incorporates depth and zero-inflated Poisson distribution, and a stage that transforms into a bird’s-eye-view 3D space through homography transformation. In the first stage, a crowd density map is estimated using EBC-ZIP model with depth-based correction applied. In the second stage, the crowd density map is mapped to the bird’s-eye view space through homography transformation. Camera parameters are estimated through Canny edge detection, Hough transformation, and vanishing point detection using RANSAC. The physical distance of the ROI is restored using the depth map estimated in Step 1, and a homography matrix is derived. Using a discrete block projection method, the center of each block is mapped to the bird’s-eye view plane using the homography matrix. Counting performance is improved by suppressing density diffusion caused by conventional pixel-level warping. In this paper, simulation results on the ShanghaiTech part A and RGBD datasets demonstrate the effectiveness of the proposed method, and bird’s-eye-view projection evaluation on real-world crowd scenario videos confirms geometrically consistent crowd localization.

Keywords: Artificial Intelligence Systems, Robot Vision, Robotic Applications
Abstract: Dynamic objects significantly degrade both tracking and mapping in online 3D Gaussian splatting-based SLAM. Existing uncertainty-aware methods improve robustness, but impose substantial computational overhead through repeated uncertainty inference and frequent global optimization. This paper reformulates dynamic-scene 3DGS-SLAM as a computation allocation problem and presents an event-driven uncertainty-aware online SLAM framework. The proposed method employs a lightweight affine uncertainty model, removes hard online depth filtering through uncertainty-weighted bundle adjustment, and adaptively schedules global optimization using drift and uncertainty-conflict observers. Experiments on dynamic benchmarks demonstrate that the proposed approach improves the dataset-average online trajectory accuracy while achieving approximately 13x higher online throughput and reducing tracker keyframe usage by 48--57% compared with the existing method.

Keywords: Artificial Intelligence Systems, Biomedical Instruments and Systems
Abstract: Lysine crotonylation (Kcr) is an important post-translational modification that plays critical roles in transcriptional regulation, chromatin remodeling, cellular differentiation, and disease progression. Although numerous computational predictors have been developed for Kcr site identification, existing methods often rely on handcrafted sequence descriptors or computationally expensive deep learning architectures, limiting their predictive performance and practical applicability. In this study, we propose Pred-KCr, a hybrid framework that integrates a lightweight protein language model with ensemble machine learning models for accurate and efficient Kcr site prediction. Specifically, we employ the ESM2-35M protein language model and adopt a partial fine-tuning strategy in which the first nine transformer layers are frozen while the last three layers are fine-tuned on Kcr-specific data. The resulting task-adapted embeddings are subsequently used to train multiple machine learning classifiers. Comprehensive experiments conducted on the independent dataset demonstrate that partial fine-tuning substantially enhances the discriminative capability of the learned representations compared with pre-trained embeddings alone. Among the evaluated classifiers, CatBoost achieved the best overall performance and was selected as the final predictor. On the independent test set, Pred-KCr attained an accuracy of 0.879, sensitivity of 0.888, specificity of 0.870, Matthews correlation coefficient of 0.758, and an area under the receiver operating characteristic curve of 0.943. Furthermore, bootstrap analysis confirmed the robustness and stability of the proposed framework. Comparative evaluation against existing state-of-the-art Kcr predictors showed that Pred-KCr consistently outperformed competing methods across all major performance metrics. These results demonstrate that combining partial fine-tuning of lightweight protein language models with ensemble learning provides an effective and computationally efficient strategy for lysine crotonylation site prediction.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems
Abstract: Distinguishing traffic controllers (TCs) from regular pedestrians is critical for autonomous driving, yet conventional detectors suffer from severe semantic confusion due to shared structural similarities. To resolve this feature entanglement, we propose Spatial-Prior Supervised Contrastive Learning (SP-SCL). By injecting a human visual prior, SP-SCL introduces a spatial constraint that explicitly guides the model to focus on discriminative upper-body attributes (e.g., helmets, vests) while ignoring identical lower-body features. Experiments on YOLOv11 demonstrate that SP-SCL successfully disentangles the feature space, eliminating critical misclassifications and achieving a 4.5% absolute improvement in overall mAP.

Keywords: Artificial Intelligence Systems
Abstract: Accurate recognition of crowd activity patterns and the generation of natural-language descriptions of those patterns are complementary tasks with practical value in public-safety monitoring, event management, and urban planning. We present a lightweight multi-task framework that jointly addresses both objectives by fusing pre-computed Farneback optical-flow embeddings with frozen CLIP image and text features through a cross-modal attention module. A single mean-pool encoder maps 100 per-frame flow vectors into a 512-dimensional motion representation aligned with the CLIP embedding space. The fused representation is routed to three output heads: an 11-class focal-loss classifier for crowd activity type, a Zero-Inflated Poisson count head for pedestrian density estimation, and a fine-tuned distilGPT-2 decoder for natural-language caption generation. Evaluated on the Crowd-11 benchmark (910 test samples, 11 classes), our method achieves 80.4% accuracy and 0.784 macro-F1, surpassing VideoMAE (78.0%) and TimeSformer (66.7%) while requiring only pre-computed flow features at inference time. Caption quality (BLEU-4 = 4.64, ROUGE-L = 26.60, CLIPScore = 0.621) is competitive with CLIPCap, a dedicated captioning baseline, demonstrating that strong captioning performance can be obtained from the same lightweight representation used for classification.

Keywords: Artificial Intelligence Systems, Robot Mechanism and Control, Robotic Applications
Abstract: This paper presents a hybrid control architecture for dynamic robotic picking tasks. The framework combines a Deep Reinforcement Learning policy for high-level interception with a dedicated Inverse Kinematics controller for precise terminal grasping, while mitigating precision limitations of monolithic learning-based approaches. The framework utilizes a Proximal Policy Optimization agent to approach moving targets, seamlessly transitioning to an Inverse Kinematics solver that reduces terminal orientational and positional errors while minimizing cumulative control effort. To facilitate deployment on physical hardware, a robust sim-to-real pipeline incorporating system identification, domain randomization, and latency injection is employed. Experimental results on a physical Franka Emika Panda manipulator validate this hybrid architecture. The system achieves an 80 % success rate in pick-and-place tasks, compared to 60.8 % for unadapted baselines, with no safety-critical violations such as joint limit breaches or collisions observed during testing.

Keywords: Artificial Intelligence Systems, Sensors and Signal Processing
Abstract: Understanding videos of work processes is essential for robot automation and has traditionally relied on manual analysis and annotation. Recently, multimodal large language models (MLLMs) have enabled the semantic interpretation of videos without human supervision. However, unedited videos recorded at work sites in real-world contexts tend to be long. Directly applying MLLMs to long video sequences leads to high computational costs and degraded performance. In this paper, we propose an adaptive temporal compression (ATC) method for video understanding. The proposed method reduces redundancy in less relevant segments based on frame-to-frame analysis and preserves high temporal resolution in informative segments. We evaluate the proposed method using MLLMs on videos of disassembly tasks recorded at an actual work site. Experimental results show that the proposed method improved the F1 score compared with both the uncompressed baseline and uniform temporal sampling, while reducing average runtime by 88%.

Keywords: Artificial Intelligence Systems, Process Control Systems, Industrial Applications of Control
Abstract: In Resistance Spot Welding processes, various process variables such as current, voltage, press force, and time interact non-linearly, making it difficult to determine optimal process conditions. Consequently, industrial sites rely heavily on the expertise of skilled workers and face challenges in automating the process. To address this issue, this study proposes a synthetic data-based framework. The proposed framework generates synthetic data that reflects non-linear process responses, interactions between variables, noise, and physical characteristics. It then sequentially per- forms a comparison of 10 regression models, Bayesian optimization-based hyperparameter tuning, SHAP-based variable importance analysis, and the identification of promising regions. Experimental results showed that the optimal model demonstrated excellent predictive performance with an R² value of approximately 0.97, enabling a quantitative explana- tion of the impact of welding variables on quality prediction. These results demonstrate the proof-of-concept feasibility of the proposed framework under synthetic-data conditions. In the future, we plan to verify its industrial applicability by applying it to actual production line data to perform multi-objective optimization that optimizes quality, productivity, and energy efficiency.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control
Abstract: Food manufacturing requires reliable inspection systems to detect foreign material contamination and maintain product safety. Sub-THz transmission imaging provides material-dependent attenuation characteristics that are useful for detecting low-density contaminants in food products. However, existing unsupervised anomaly detection methods mainly rely on RGB-pretrained visual representations, which may not adequately capture the transmission behavior of Sub-THz images. This paper proposes a Beer-Lambert guided representation learning framework for unsupervised anomaly detection in Sub-THz food inspection images. The proposed method introduces an attenuation decomposition module as an auxiliary regularization module that constrains student representations through attenuation reconstruction during training. In addition to the conventional one-class setting, we introduce a Leave-One-Food-Out protocol to evaluate generalization capability under unseen food categories. Experimental results on the Inline-Food-Inspection-THz dataset show that the proposed method improves overall anomaly detection performance over the baseline method.

Keywords: Artificial Intelligence Systems, Human-Robot Interaction, Robotic Applications
Abstract: This paper presents a speech-driven beat gesture generation system for the EveR-6 social robot, which has an 18 degrees-of-freedom (DOF) upper-body. Generating natural co-speech gestures is a key challenge in social robotics: gestures must be temporally synchronized with speech prosody while reflecting semantic content. We propose a Transformer-based architecture that takes prosodic features (fundamental frequency and root-mean-square energy) along with beat onset labels as inputs, and conditions gesture generation on window-level text embeddings via Feature-wise Linear Modulation (FiLM). Unlike sentence-level conditioning, the proposed window-level scheme assigns a distinct text embedding to each 4-second generation window based on the words spoken in that interval, reducing intra-batch embed- ding cosine similarity from ∼1.0 to 0.30. The model is trained on the BEAT2 dataset with joint angles mapped to the EveR-6 kinematic structure via SMPL-X body pose conversion. Experimental results demonstrate a Beat Consistency Score (BCS) of 0.707 and a mean absolute joint error (MAE) of 12.9 degrees, representing a 66.5% reduction in MAE compared to a prosody-only baseline.

Keywords: Artificial Intelligence Systems, Navigation, Guidance and Control, Robot Vision
Abstract: Unmanned Surface Vehicles (USVs) require pixel-level perception of navigable water, together with reasoning about contextual hazards such as nearby vessels, port infrastructure, and adverse weather. Recent vision-language models (VLMs) can supply this reasoning, but hallucination and inference latency rule them out as the direct controller in safety-critical maritime operations. We therefore propose a grounded reasoning framework in which Vertical and Detail Attention (VaDA), a maritime segmentation model, runs in a real-time main loop alongside an asynchronously invoked frozen VLM advisor. The VaDA masks ground the VLM, and a mask-based consistency check verifies each VLM proposal against this evidence before execution; rejected proposals fall back to a classical reactive controller. We evaluate the framework on four open-weight VLMs spanning two architecture generations: LLaVA-1.5-7B (2023) and the more recent InternVL3-1B, Qwen2.5-VL-3B, and Qwen3-VL-2B (all 2025). Our five-decision navigation protocol extends OASIs (Ocean AI Segmentation Initiatives) and is annotated by three independent annotators. On OASIs, grounding consistently improves both direction and mean decision accuracy across the three modern VLMs (2025), and the consistency check raises final decision accuracy further. A case study on LLaVA-1.5-7B (the oldest of the four) reveals an asymmetric response, suggesting that the benefit of grounding depends on VLM generation rather than parameter count. Among nine mask providers, VaDA achieves the best deployability–accuracy trade-off on our real-time reference platform.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems, Navigation, Guidance and Control
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 encounter 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
Abstract: Long-term prediction data in videos is inherently expensive to construct, as it requires high-level temporal reasoning over causal relationships, scene transitions and complex event dynamics. Existing data construction approaches primarily rely on manual annotation or automatic generation using a single Vision-Language Model (VLM), often resulting in semantic hallucinations, temporal inconsistencies and redundant samples. Furthermore, collecting and refining large-scale datasets under user-specified conditions remains a significant challenge. In this paper, we propose an agentic data construction pipeline for automatically building high-quality long-term prediction datasets tailored to user-defined collection environments. The proposed framework adopts a collaborative multi-agent architecture in which multiple Vision-Language Model agents jointly perform video retrieval, semantic filtering, temporal event extraction, data verification and prediction target generation. The proposed system employs cross-agent verification and temporal consistency filtering to reduce errors commonly produced by single-VLM-based automatic generation pipelines while improving semantic consistency and redundancy reduction. Experimental results demonstrate that the proposed multi-agent data construction framework consistently outperforms single-VLM-based pipelines across multiple dataset quality metrics, including annotation accuracy, redundancy reduction, temporal consistency and overall data quality.

Keywords: Artificial Intelligence Systems
Abstract: Generating interactive motions for two characters remains a challenge due to the spatial and temporal coupling between the participants. Recent models can generate natural responses to an opponent’s movement, but they lack a way for users to control what specific action the character should perform. Meanwhile, existing models that support text control are limited to offline generation and cannot operate in real-time. To solve this problem, we propose a text-guided online reactive motion generation system for two-character boxing. Our model integrates text conditioning into R2R by replacing its standard transformer blocks with Adaptive Layer Normalization (AdaLN) modulated MARTransBlocks. By using classifier-free guidance during inference, our system provides clear direction to the model while maintaining real-time streaming at 91.5 ms per step. When tested on the DuoBox dataset, our model outperforms the previous baseline on all FID metrics and achieves an Action Accuracy of 86.1%, demonstrating reliable text adherence with realistic motion dynamics.

Keywords: Artificial Intelligence Systems, Sensors and Signal Processing, Industrial Applications of Control
Abstract: The shipbuilding industry, which is highly labor-intensive, is facing a decline in competitiveness due to a shortage of skilled welders and a decreasing workforce. To address this issue, welding robots have been deployed in shipyards. However, harsh working environments still restrict their productivity. In particular, welding robots require a touch-sensing process, in which the robot physically contacts the start and endpoints of the weld. As welding robots typically process hundreds of weld lines per day, touch-sensing becomes a major factor reducing productivity. To overcome this limitation, we proposes an automatic endpoint detection method based on the Chi-square statistic, which learns current data measured in real time during welding. Experimental results obtained from actual welding tests demonstrate that the proposed method can detect the endpoint within 3 mm. These results indicate that the proposed method has the potential to replace the conventional touch-sensing process and significantly improve productivity in welding processes.

Keywords: Artificial Intelligence Systems, Multimedia Systems, Information and Networking
Abstract: Fine-tuning text-to-image diffusion models on small custom datasets enables personalization but amplifies training-data privacy risks, as fine-tuned models preserve their training images more strongly than unseen ones. We study this risk through membership inference, which determines whether a target image was used during fine-tuning. Existing diffusion-model attacks largely depend on model-internal signals such as losses, predicted noises, and intermediate denoising states, which are unavailable in deployed image-to-image services. We therefore propose a black-box attack based on the Conditional Reconstruction Gap (CRG) that uses only final image-to-image outputs. CRG contrasts reconstruction fidelity under the ground-truth caption against an empty caption, based on the observation that the matching caption preserves a member image markedly better and thus yields a measurable gap. On a fine-tuned Stable Diffusion model, this conditional signal outperforms both gray-box and black-box baselines in the evaluated setting.

Keywords: Artificial Intelligence Systems
Abstract: Sim-to-real research on reinforcement learning (RL) has focused on matching dynamics parameters, while the execution time of policy inference has received little attention as a source of performance degradation. We experimentally show that an RL policy trained on physical hardware with on-MCU inference becomes implicitly calibrated to the training-time inference execution time τ. A Discrete SAC stepper-motor policy trained at τ ≈ 266 μs (forloop-based LibTorch) degrades in chirp-trajectory (0.1–1.0 Hz, ±π/2 rad, 30 s) tracking MAE by 24.5×, 31.1×, and 183.8× when deployed without retraining at τ ≈ 185 μs (CMSIS-DSP), 161 μs (Eigen), and 473 μs (sequential), respectively. A parallel inference comparison confirms that all implementations produce identical actions from identical states, and padding the Eigen path with a 105 μs delay restores MAE to 0.187°, 100% of baseline, isolating inference execution time as the sole cause. These results indicate that faster inference does not necessarily improve real-time control performance; this effect is difficult to capture in simulation and warrants explicit attention when deploying RL policies on embedded hardware.

Keywords: Artificial Intelligence Systems
Abstract: This study proposes OISNet, a model optimized for low-power edge platforms for optical image stabilization (OIS) marker-tracking inspection. OISNet combines CoordConv-based spatial-information injection, limited spatial downsampling, a lightweight residual block, channel compression, and a compact fully connected coordinate-regression head, with Wing Loss for small-error optimization. A synthetic-data generator provides precisely known ground-truth coordinates without manual labeling error while incorporating Gaussian noise, motion blur, and metallic-reflectivity distortions. The dataset contains 10,000 training, 2,000 validation, and 1,000 test images. TensorRT-accelerated inference was evaluated on an NVIDIA Jetson AGX Orin operated in its lowest-power 15 W mode. OISNet achieved a mean L2 error of 0.00465, latency of 1.85 ms, jitter of 0.398 ms, and throughput of 540.1 FPS, demonstrating a practical precision-latency balance for embedded OIS inspection.

Keywords: Artificial Intelligence Systems, Control Theory and Applications, Industrial Applications of Control
Abstract: Combining physics-based and data-driven models has shown promise for capturing the behaviour of complex real-world mechanical systems, but the designer must decide how strongly the data-driven component is constrained by the assumed physics, and the appropriate strength depends on the data. We propose a controlled family of four architectures that share the same learned component and differ only in constraint strength --- from an unconstrained Transformer to a strictly constrained Physics-Informed Transformer (PIT), the fourth being a Gated-PIT whose learned, state-conditioned gate decides how much to trust the learned component. We evaluate them on a slider-crank mechanism whose contact force is uncertain and exhibits time-varying dynamics, using synthetic datasets with injected memory effects of increasing duration as well as measurements from an experimental test rig. We find that the preferred architecture depends strongly on the degree of regime mismatch and the strength of the memory effects, and that the identified physical parameters, though highly repeatable, are systematically biased whenever memory is present --- revealing a trade-off between prediction accuracy and parameter identification. We show that the learned gate improves this trade-off, and that reformulating the training cost to score the equation of motion directly reduces the parameter bias substantially at almost no accuracy cost. Finally, we demonstrate that synthetic-to-real pretraining benefits the unconstrained model most. Together, these results give practical guidance on choosing the architecture, constraint strength, and training cost of physics-informed sequence models.

Keywords: Artificial Intelligence Systems, Control Theory and Applications, Sensors and Signal Processing
Abstract: This paper presents the implementation of a wireless neural network controller based on a reference compensation scheme, known as the Reference Compensation Technique (RCT), which is a real-time neural network control method. A dedicated neural network controller is developed and employed to compensate for uncertainties in gimbal systems, thereby improving angle-tracking performance. Experimental studies are conducted to demonstrate the feasibility and effectiveness of the RCT-based neural network controller for online control of nonlinear systems.

Keywords: Artificial Intelligence Systems, Robotic Applications, Robot Vision
Abstract: Task-level pick-and-place controllers are typically forward scripts; a fixed sequence of perceive, grasp, transport, and place that halts on the first deviation. In multi-object operation deviations are routine: empty grasps, mid- transport slips, and stale planner attachment state. We reformulate the controller around a continuously checked object-secured invariant, realized as a behavior tree whose memoryless root re-derives the invariant on every tick (10 Hz) and re-enters the pick phase whenever it fails; the pick branch is memory-enabled so committed picks run atomically. Recovery is therefore not a scripted handler but a structural property of the tree. Two complementary proprioceptive checks, a stall-at-close flag and a transport position-jump, both route to a single re-pick branch, tolerating either check’s failure mode. We validate on an Indy7 V2 + RG6 cell in NVIDIA Isaac Sim and on hardware, with RT-DETR to FoundationPose perception and cuMotion planning. The controller recovers from all injected empty grasps, recovers from injected slips when the displaced object remains perceivable, adds ≈ 8% throughput overhead on the failure-free path relative to the stock NVIDIA BT, and places 72/84 objects across 12 physical whole-scene runs with no slip and no bad pick.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Robot Mechanism and Control
Abstract: Action Chunking with Transformers (ACT) enables robots to acquire structured motion policies for contact-rich manipulation from limited demonstrations without large-scale pre-training. A prior efficiency-oriented study found that, on a structured surface-cleaning task, a small batch size (8) matched or exceeded the task completion of a larger batch size (32) while requiring less GPU memory and wall-clock training time. We examine whether the smaller-batch policy carries internal representational deficits that would undermine this efficiency advantage. Through offline analysis of frozen checkpoints, each replayed on its own recorded deployment episode, applying Principal Component Analysis (PCA) to encoder outputs and inspecting encoder and decoder attention patterns, we find that the larger-batch checkpoint exhibits higher effective representation dimensionality and distinct attention structure, although the attention measures do not vary monotonically with batch size. Yet the dimensionality ordering is not reflected in the prior study's deployment results, where the larger-batch policy recorded the lowest task completion. We treat this as an observed mismatch between internal representation metrics and single-trial deployment outcomes, rather than as evidence of a general batch-size-dependent performance effect. The lower-dimensional representations of the smaller-batch policy, though inferior by these internal measures, carry no deployment penalty in the case examined. We therefore caution that internal learning-quality metrics are not reliable proxies for deployment readiness in this regime, and present the result as a representation-level characterization and offline analysis method rather than a causal account, discussing competing explanations for the gap.

Keywords: Artificial Intelligence Systems, Autonomous Vehicle Systems, Sensors and Signal Processing
Abstract: This paper presents identification of road surface conditions by neural network models. Driving experiments were conducted to collect data from various road surface conditions, and input signals to neural networks were reconstructed to effectively represent vehicle dynamics. Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM) networks, and hybrid CNN–LSTM architectures were evaluated. The trained models were implemented on an actual vehicle to validate their feasibility in real-time driving environments. The GUI has been designed to monitor the road surface conditions.

Keywords: Artificial Intelligence Systems, Process Control Systems, Sensors and Signal Processing
Abstract: Fast charging of lithium-ion batteries requires charging-current decisions that reduce charging time while satisfying prescribed constraints on terminal voltage and cell temperature. Reinforcement learning (RL)-based charging policies can adaptively determine the current from the battery state. However, during exploration, the actor may propose currents that would violate the voltage or temperature limits. To address this issue, this paper introduces a physics-informed action shielding method for SAC-based fast charging. A safety threshold model is first pretrained using PyBaMM-based electrochemical-thermal simulations to estimate a state-dependent current threshold. When the SAC actor generates a current above this threshold, the shield corrects it before execution. For shielded actions, a one-step counterfactual simulation assesses whether the raw current would violate the prescribed margin limits. The resulting label is then used for online threshold adaptation. Actor safety regularization is also introduced to discourage repeated over-threshold actions. Simulation results show that the shielding framework reduces hard-limit violations compared with no-shield SAC. It also achieves a lower shield activation rate and a higher average return than naive-shield SAC. These results suggest that physics-informed action shielding can improve the balance between constraint satisfaction and charging progress under the evaluated simulation conditions.

Keywords: Artificial Intelligence Systems, Biomedical Instruments and Systems, Information and Networking
Abstract: Speech-language assessments play a critical role in the diagnosis, prognosis, and treatment planning of individuals with speech disorders. However, conventional assessment methods often rely on manual analysis and subjective judgments, which may limit efficiency and consistency. This study aimed to develop an explainable artificial intelligence (XAI)-based speech disorder diagnostic program capable of automating speech-language assessments and providing rapid diagnostic feedback. A total of 1,000 speech recordings were retrospectively collected from publicly available AI-Hub datasets, including Voice Data for Diagnosis of Parkinson’s Disease and Related Disorders and Speech Command Data from Individuals with Dysarthria. Acoustic features such as jitter, shimmer, harmonics-to-noise ratio (HNR), formants, fundamental frequency (F0), Mel-frequency cepstral coefficients (MFCCs), alternating motion rate (AMR), pitch standard deviation, and nasalance were extracted. The collected speech signals underwent preprocessing procedures including noise reduction, speech alignment, volume normalization, and sampling-rate standardization. The extracted features were analyzed using an XAI-based diagnostic module designed to classify speech disorders, predict severity levels, and identify error types. Diagnostic outputs were summarized through interpretable reports presenting abnormal acoustic indicators, recommended speech training tasks, and graphical comparisons of speech performance. The proposed framework supports objective assessment and individualized rehabilitation planning while enabling continuous model improvement through user feedback and retraining mechanisms. Currently, the XAI-based speech disorder diagnostic and rehabilitation program is under development and validation.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Sensors and Signal Processing
Abstract: This paper presents a fully data-driven sensorless control scheme for Permanent Magnet Synchronous Motors (PMSMs) utilizing a Koopman Linear State Observer based on time-delay coordinate embedding. Conventional sensorless control methods rely on complex nonlinear observers that require precise mathematical system models and heavy computational resources. To overcome these limitations, the proposed method lifts measurable nonlinear electrical states (voltages and currents) into a finite-dimensional Koopman-invariant linear subspace via Singular Value Decomposition (SVD). Subsequently, a linear regression model is applied to the projected states to directly estimate the unmeasured mechanical rotor speed. The primary contribution of this work lies in integrating the spatial projection and linear estimation processes into a single static gain matrix. This proposed architecture bypasses the need to solve differential equations online, thereby significantly reducing the computational burden and facilitating real-time execution on standard microcontrollers. Simulation results demonstrate that the proposed data-driven observer successfully achieves highly accurate speed estimation while maintaining significant computational efficiency and robustness against system nonlinearities.

Keywords: Artificial Intelligence Systems, Industrial Applications of Control, Autonomous Vehicle Systems
Abstract: Permanent magnet synchronous motors (PMSMs) are widely used in electric vehicles, robotics, and other high-performance drive applications. However, parameter uncertainties, unmodeled nonlinearities, and operating-condition-dependent variations cause discrepancies between the baseline model and the actual PMSM plant, thereby degrading current tracking performance. To address this problem, this paper proposes a hybrid PMSM control method based on Koopman operator-based residual compensation. The key idea is to define the mismatch between the baseline model and the actual plant as residual dynamics, approximate them linearly in a Koopman lifting space, and convert them into a compensation voltage input. By preserving the model-based control framework, the proposed method maintains the stability-oriented structure and physical interpretability of conventional model-based control, while mitigating performance degradation caused by model mismatch arising from parameter uncertainties and unmodeled nonlinear effects through data-driven residual compensation. MATLAB/Simulink-based simulation results demonstrate that the proposed compensator enables the actual PMSM plant to accurately track the corresponding baseline-model currents in both the d-and q-axes under model mismatch, thereby validating the effectiveness of the proposed control strategy.

Keywords: Artificial Intelligence Systems, Robot Vision, Robotic Applications
Abstract: This paper presents a resource-efficient modular framework for humanoid obstacle avoidance. Recent humanoid parkour and character-control systems show that depth-conditioned RL, motion matching, and learned motion priors can produce highly dynamic skills, but these pipelines often train perception and whole-body control together or require large motion-generation infrastructure. We instead learn a narrow vision-to-reference interface: depth and proprioceptive state are converted into the full reference goal that an existing locomotion policy can already track. A motion-reference variational autoencoder compresses future motion windows into a latent variable and reconstructs the next-cycle policy reference. A depth-state encoder is then distilled to predict the same latent using only causal observations. At deployment, the encoder and decoder generate reference goals at every control step, while the robust student policy performs whole-body tracking. This separation reuses a high-quality locomotion controller, reduces vision-training cost, and avoids direct mapping from depth to torque-level actions. Closed-loop evaluations on the humanoid model in MJX over over 900 episodes show depth-conditioned yaw responses and obstacle-avoidance behavior, while revealing remaining limitations in collision-free clearance under randomized obstacle geometry.

Keywords: Artificial Intelligence Systems, Robot Vision
Abstract: Selecting a practical 3D whole-body pose estimator is difficult because body, hand, and face regions show different trade-offs in accuracy, temporal stability, part coverage, and inference speed. This paper evaluates representative whole-body, body, hand, and face estimators on RGB human-motion videos with part-wise 3D ground truth. We first analyze each anatomical part separately using PA-MPJPE, temporal errors, and FPS. We then evaluate controlled fusion settings that combine selected body, hand, and face sources to compare accuracy-focused, balanced, and speed-focused whole-body pose estimation pipelines. Under the evaluated PA-MPJPE protocol, SAM 3D Body provides the strongest body and hand accuracy, while PEAR provides the best face accuracy and the fastest unified extraction setting. Fusion experiments further show that part-wise combinations can improve accuracy, whereas a single unified model is preferable when processing speed is the primary constraint.

Keywords: Artificial Intelligence Systems, Robot Mechanism and Control, Robotic Applications
Abstract: Action Chunking Transformer (ACT) has shown potential for low-cost bimanual manipulation, but real-world deployment is affected by dataset quality, command timing, temporal ensembling, and motor velocity/acceleration limits. This study presents a practical deployment procedure for applying ACT to a Dynamixel-based bimanual leader-follower robot in an automotive part grasp-handover-placement task. Approximately 900 demonstrations were collected and about 400 episodes were curated for training. A baseline evaluation with object and holder pose variations recorded 39/56 task completions (69.6%). Approximately 90 inference logs were then used as diagnostic examples to compare command publish rate and motor limit settings. The analysis focuses on temporal ensemble source chunks, weights, source time span, weighted RMS spread, and robot-command tracking error. The observations indicate that different runtime conditions reflected in the logged metadata can be associated with different command-level behavior, even when the policy checkpoint is unchanged. Therefore, this study emphasizes recording and managing dataset curation, measured execution frequency, temporal ensemble metadata, and motor limit settings when deploying ACT-based policies on low-cost robots.

Keywords: Artificial Intelligence Systems, Biomedical Instruments and Systems
Abstract: This study was conducted to verify the validity of a new AI articulation evaluation program based on U-TAP by comparing the children's PCC (Percentage of Consonants Correct), assessed using the program, with the auditory perceptual evaluations of speech therapists. During the first visit (Visit 1), the subjects underwent the U-TAP assessment using an AI articulation assessment program inside a soundproof booth where external noise was blocked. The test voice was recorded using a separate microphone. Examiner 1 (speech therapist) and Examiner 2 (speech therapist) listened to the subjects' recorded voices, completed the U-TAP assessment result log, and verified the PCC results analyzed using the AI articulation assessment program. During the second visit (Visit 2), which took place within 7 days (±5 days) after Visit 1, the subjects underwent the U-TAP assessment using the AI articulation assessment program inside a soundproof booth where external noise was blocked. The test voice was recorded using a separate microphone during this visit. Examiner 1 (Speech Therapist) and Examiner 2 (Speech Therapist) listened to the recorded speech of the subject, completed the U-TAP evaluation result log, and verified the PCC results analyzed using an AI articulation evaluation program. The results of repeated measurements of word- and sentence-level consonant accuracy on three subjects during Visit 1 and Visit 2 are as follows. The evaluation results of Examiner 1 and Examiner 2 were generally within a similar range, but differences between the examiners were observed depending on the subject. The Revocal evaluation results were close to the examiner's values under some conditions, but generally tended to be lower than those of Examiner 1 and Examiner 2. In particular, the difference between the Revocal and examiner evaluations was greater in sentence-level consonant accuracy than in the word-level. This research was supported by the Ministry of Science and ICT (MSIT), Republic of Korea, through the National IT Industry Promotion Agency (NIPA). (Grant No. PJT-26-040359)