Skip to content
#Research
Oct 27Tue
  1. arXiv cs.RO60

    PS-DLO: A Property- and Spatially Aware Deformable Linear Object Manipulation Policy

    The authors propose PS-DLO, a deformable linear object (DLO) manipulation policy that combines object properties with spatial awareness. Through contrastive image-text learning, it maps initial observations to categorical descriptions of DLO properties, and uses SAM3 to extract mask-guided features from future observations, which are distilled into an action expert. On nine real-world tasks across single-arm and bimanual platforms, PS-DLO achieves an average success rate of 68.9%, 20.0 percentage points higher than the strongest baseline; in six of seven visual-shift conditions for butterfly-knot untying, its success rate exceeds the baseline.

  2. arXiv cs.RO55

    Semantic-Coupled Perception Framework for Autonomous Laparoscope Control

    This paper proposes a closed-loop autonomous laparoscope control framework that couples real-time surgical phase recognition with camera-motion-compensated visual perception, using phase dependencies among surgical entities to improve target state estimation under occlusion and entity overlap. Surgical phase recognition and target tracking were evaluated on benchmark datasets, and phantom experiments showed improved target visibility and task execution performance compared with fixed-field-of-view control. The paper was submitted to IEEE Robotics and Automation Letters (RA-L).

  3. arXiv cs.RO68

    HumanoidNDM: Closing the training-deployment gap in humanoid motion tracking with a neural dynamics model

    The study proposes HumanoidNDM, an online fine-tuning framework for humanoid motion tracking driven by a neural dynamics model trained on real robots, aimed at closing the dynamics gap between training and deployment. The method includes low-frequency reference perturbation data collection with an adaptive intensity mechanism, training a dynamics model based on proprioceptive history in Euclidean space combined with differentiable forward kinematics, and directly fine-tuning the policy through backpropagated model gradients.

  5. arXiv cs.RO72

    X-World: Transferring video priors to variable-view robot control

    X-World proposes a framework that transfers video priors to a shared action expert through fixed-length token sequences, and uses residual cross-view attention to extend a pretrained video backbone to one, two, or three views. The policy reaches 98.5% mean success on LIBERO, and its LIBERO-trained checkpoint obtains 76.5% on LIBERO-Plus without additional adaptation; across six dual-arm real-world tasks, it achieves a mean success rate of 81.7% over 20 trials per task.

  6. arXiv cs.RO64

    Millimeter-Wave Radar Uses Doppler Information to Track Occluded Targets' Position and Velocity

    The authors propose a method that uses millimeter-wave radar Doppler information to track non-line-of-sight occluded targets, achieving 8.95 cm position error and 1.75 cm/s velocity error on held-out trajectories. The study also releases the first indoor non-line-of-sight radar dataset, containing 576 sequences totaling 720,856 frames and covering 12 trajectories with different speeds, starting positions, and material configurations. The method models non-line-of-sight tracking as continuous position-velocity estimation, uses a shared 3D encoder to process the full range-azimuth-Doppler representation, and estimates position and velocity separately through two recurrent branches before fusing them into motion-consistent trajectories.

  7. arXiv cs.RO60

    DSGN: A Dual-Stream Gated Network for Scalable Multi-Robot Surveillance and Containment

    The authors propose the Dual-Stream Gated Network (DSGN), which separates surveillance and containment into independent representation streams and coordinates them through a context-aware gating mechanism to alleviate task-switching difficulties in large-scale robot networks. The method also introduces a probabilistic prior to regularize the gating process, reducing ambiguous intermediate decisions. Under constant-density cross-scale settings, DSGN maintains task effectiveness and scalability; ablation and staged experiments show that dynamic gating and the probabilistic prior support task arbitration and mode switching. The paper has been accepted to RA-L October 2026, and the code is publicly available.

  8. arXiv cs.RO59

    LiDAR-based GNN framework learns interaction-aware path corrections from unlabeled trajectories for construction robots

    The study proposes a real-time, LiDAR-only framework that learns interaction-aware path corrections from unlabeled demonstrations: constant-velocity extrapolation generates residual learning targets, a GNN models relationships among the robot, pedestrians, and goal context, and at deployment an extended Kalman filter tracks LiDAR detections to build dynamic graphs, with the GNN offsetting planner-generated waypoints through a planner-agnostic interface.

  9. arXiv cs.RO49

    Lightweight Change Detection for Adaptive Collective Perception in Dynamic Environments

    For dynamic environments where tile proportions and the spatial correlation of surface patterns change over time, this work integrates a lightweight change detection mechanism into a Bayesian collective decision-making framework, resetting robot beliefs when distribution shifts are detected. The detector requires only constant memory and constant per-sample computation, unlike two recent methods that must store the full sample stream, making it suitable for memory-constrained robots. Webots simulations show competitive change detection latency, more consistent accuracy across different communication settings, and improved decision accuracy when spatial correlation changes.

  10. arXiv cs.RO55

    Extended LASMP: A Language-Aided Subset Sampling Motion Planner for Drone Navigation

    The paper proposes Extended LASMP, which incorporates navigation commands directly into the sampling distribution to improve the exploration efficiency of sampling-based motion planning for quadcopters in three-dimensional environments. The method generalizes subset sampling to command-aligned regions and uses a prompt-based large language model to parse natural-language instructions into yaw-conditioned and altitude-aware geometric constraints. Simulations in complex environments show that the method generates collision-free, consistent trajectories while significantly reducing sampling effort compared with uniform RRT-based planners.

  12. arXiv cs.RO58

    T-ConSG: A LiDAR-Based Real-Time Temporal Scene Graph for Dynamic Construction Environments

    The paper proposes T-ConSG, a LiDAR-based real-time temporal construction scene graph that uses spatial-relation edges and event-participation edges to jointly represent free space, static objects, dynamic objects, and environmental events. The framework incrementally builds the graph from LiDAR geometric evidence and explicitly represents changes in accumulated spatial evidence as installation and removal events. The authors validated the representation in three studies: dynamic-aware navigation on a bricklaying testbed, construction progress monitoring on the ConSLAM dataset, and online change detection in a simulated environment.

  13. arXiv cs.RO60

    Cut a Wall, See a Wall: Identity Substitution in Occluded Point-Cloud Perception for Construction Robots

    The paper finds that when entire regions of a scene are occluded at test time, point-cloud segmentation networks relabel the hidden elements into a single sink class; under fixed layouts this class absorbs up to 70% of test points, and baseline mIoU drops to 0.15-0.27. The authors append 14 hand-computed geometric features per point (GPE), restoring mIoU to 0.54-0.92 on PointNet++ and a PTv3-style transformer and removing 64-100% of whole-object losses; on real S3DIS scans, whole-wall losses drop from 10 to 0.

  14. arXiv cs.RO58

    A Motion-Induced WiFi-CSI Sensing Simulation Framework for Multi-Robot Systems

    This study proposes an open-source simulation framework that integrates the WiFi Sensor for Robotics (WSR) sensing pipeline into Gazebo. It can generate synthetic WiFi-CSI measurements with realistic physical signal characteristics and supports processing both simulated and real WiFi-CSI data for direct comparison and validation. The authors used TurtleBot3 robots to compare simulated and real experiments, demonstrating the framework's use in bearing estimation and robot localization. The code and experimental details have been made public.

  15. arXiv cs.RO58

    ElevateHead: When Action Heads Need One-Time Decisions in Multimodal Imitation Learning

    ElevateHead proposes an action head that selects the task execution mode once at a learned decision point and keeps it fixed for the entire episode; users can also set this choice as an input. On held-out Franka Kitchen tasks, 45.2% of its episodes completed all four subtasks, versus 29.3% for BeT (+0.160, 95% CI [+0.129, +0.188]); on held-out block stacking tasks, version 4 was comparable to BeT but did not surpass it.

  16. arXiv cs.RO60

    Interactive Information-Gathering TAMP Framework for Long-Horizon Manipulation

    The authors propose a task and motion planning (TAMP) framework that alternates task-directed interactive information gathering with planning and execution. It plans optimistically for interactive actions based on PDDLStream, verifies outcomes during execution, retains actions that remain applicable, and replans only the subsequent connections. The framework was tested in simulation and on a real robot on tasks such as determining whether an object is movable, whether a pen has ink, whether a drawer can be opened, and whether a container is empty. The results show that interaction-driven state updates and recovery replanning can support continued execution under uncertain outcomes and external disturbances.

  17. arXiv cs.RO55

    Learning Functional Specialization for Dual-Arm Robot Control

    The study compares, in a simulated musculoskeletal system with dual communication-loop controllers, the effects of differentiated versus shared controller objectives and fixed versus balanced hand-task assignments on dual-arm control using a 2×2 design. Functional specialization was strongest when differentiated objectives were combined with a fixed assignment consistent with contralateral control routing, achieving a dual-arm success rate of 81.33% and the highest control efficiency. This advantage widened as motor noise increased and narrowed under stronger energy penalties. Under differentiated-fixed training, swapping the roles of the two hands reduced the success rate by 80.46 percentage points, whereas under shared-balanced training it fell by only 1.51 percentage points.

  19. arXiv cs.RO58

    Sybil-Resilient Multi-Drone Exploration of Indoor and Confined Spaces

    The paper studies the problem of Sybil attacks during multi-drone collaborative exploration of buildings: a compromised drone forges identities, claims frontiers, and sends fabricated scans, causing real drones to miss areas and generating nonexistent walls in the map. The authors propose five defenses that rely only on the drones' own sensors and maps, including auditing map data, checking for a body's presence at an identity's claimed location, trusting only identities whose bodies have been observed, quarantining identities unobserved by any drone, and binding each observed body to a single identity.

  20. arXiv cs.RO64

    RoboLongR: Causally consistent rewards guide robots to progressively master long-horizon manipulation

    RoboLongR proposes an offline-to-online reward learning framework that uses causal consistency to align reward progress with the prerequisites supported by execution history, guiding policies from behavior cloning to online reinforcement learning. Pretrained on 88 RLBench tasks and Open-X data, the method achieves an average success rate of 31.8% on 18 long-horizon tasks excluded from reward pretraining, 9.2 percentage points above the strongest comparison baseline; on four real-robot tasks it achieves a 57.5% success rate, 32.5 percentage points above the comparison method.

  21. arXiv cs.RO72

    AdaDexTrack: Dynamic Modulation for Adaptive and Generalizable Dexterous Manipulation Tracking

    AdaDexTrack proposes a modulator-in-the-loop framework for language-conditioned dexterous manipulation tracking, using three types of feedback correction—reference modulation, object-latent modulation, and positional-target modulation—to mitigate synthetic reference noise, embodiment bias from human-to-robot retargeting, and error accumulation in fixed-reference tracking.

  22. arXiv cs.RO71

    RoboWorM: A Benchmark for Evaluating Motion Realism and Acceleration Effects in Embodied Video World Models

    The authors propose RoboWorM, a deployment-aware benchmark for robotic manipulation video world models that examines structural and spatiotemporal fidelity using hardware priors such as robot kinematic trees and fixed camera configurations. The benchmark collects 1,605 episodes across single-arm, bimanual, and humanoid morphologies, covering Markovian and long-horizon non-Markovian tasks under single-view and multi-view setups, and proposes two deterministic metrics that do not rely on vision-language models: KIVA and RIGIS.

  23. arXiv cs.RO75

    Scan2Instruction and PRPT-KS: Demonstration-free adaptation for language-guided navigation from a single scan

    The study proposes Scan2Instruction and PRPT-KS, which use a single iPhone scan to build a training environment and adapt a language-guided navigation policy without human demonstrations. Across five real environments, RealNav achieved an 85% real-robot success rate, compared with 49% for a baseline post-trained on human demonstrations; in unscanned areas of the same buildings, it still reached 82%, versus 41% for the baseline.

  24. arXiv cs.RO60

    RoboReason: Improving robotics-relevant VLM reasoning with scalable data generation

    RoboReason proposes constructing spatial and temporal ordering tasks from real robot stereo camera and wrist camera images to evaluate the robotics-relevant reasoning capabilities of vision-language models; experiments show that several advanced VLMs perform poorly on the ordering tasks. The authors accordingly propose a scalable autonomous training data generation pipeline, and supervised fine-tuning and reinforcement learning on this data both significantly improve model performance on spatiotemporal ordering and downstream robotics and visual reasoning tasks. The work was accepted to the EMNLP 2026 main conference.

  25. arXiv cs.RO60

    POD-RoPE: Physical Observation Distribution Rotary Position Encoding for Embodied AI

    The authors propose POD-RoPE, which represents a token's observation footprint as a distribution over camera centers, viewing directions, and sampling times, uses expected rotations to encode relative relationships between distributions, and reduces to pointwise RoPE when each distribution collapses to a single point.

  26. arXiv cs.RO52

    Ground Friction Assessment Based on Robot Tire-Ground Interaction

    The study drove a skid-steer robot at set speed profiles over five surfaces with different friction characteristics, identifying surface type by the deviation between inertially estimated velocity and wheel odometry velocity. Response delay measurements showed a consistent surface-dependent trend, with the longest delay on soap and the shortest on carpet; spectral features achieved 97.1% balanced classification accuracy under record-level cross-validation. The authors conclude that both the magnitude of velocity mismatch and frequency-domain features contain tire-ground interaction information, which can be used for surface identification and for slip-aware wheel odometry weighting in state estimation.

  27. arXiv cs.RO36

    Quantum Closed-Loop Inverse Kinematics via Hilbert-Space Encoding of Robot Kinematics

    The paper proposes a structure-preserving quantum representation that maps rigid-body transformations and serial-manipulator kinematics to parameterized quantum circuits, encoding spatial directions in single-qubit states and reconstructing end-effector poses from Pauli expectation values. Using the parameter-shift rule, the manipulator geometric Jacobian can be reconstructed from derivatives of quantum observables, analytically equivalent to the classical result under ideal readout. On this basis, the authors propose a hybrid closed-loop method, Quantum Closed Loop Inverse Kinematics, which uses the quantum-derived pose and Jacobian information to solve inverse kinematics.

  28. arXiv cs.RO64

    Open-World Law-to-Action: Using vision-language priors and touch to infer the motion laws of parts with hidden physics

    The authors propose the Open-World Law-to-Action (L2A) framework, in which a vision-language model proposes, for each movable part, its motion type, the terms of its law of motion, and possible locking factors; the part's law is then identified online via Bayesian linear regression with closed-form evidence, and touch can overturn the priors.

  29. arXiv cs.RO62

    EXE: Egocentric Cross-Embodiment Pretraining with Factorized Latent Actions for VLA Models

    EXE is an egocentric cross-embodiment pretraining framework that learns factorized latent action representations from human manipulation videos, separating shared manipulation priors from embodiment-specific signals. The authors state that this shared action space lets robot policies acquire generalizable manipulation semantics from diverse human demonstrations, and that the same pretraining can be used for both parallel grippers and dexterous hands. On RoboTwin 2.0, RoboCasa GR1, and real-world manipulation tasks, EXE achieves consistent improvements over baselines without pretraining.

  30. arXiv cs.RO53

    SHAPE: An Energy-Based Fault Detection and Isolation Framework for Flat Port-Hamiltonian Mechanical Systems

    The paper proposes an energy-based fault diagnosis framework, SHAPE, which combines port-Hamiltonian variables and differential flatness to generate nominal power references from planned motion, constructing mechanical port-power residuals without a dedicated dynamic observer. In simulations of a six-degree-of-freedom Kinova manipulator under model predictive control, SHAPE-CWT achieved an average accuracy of 84.69% on 3000 held-out tests, compared with 81.67% for torque residuals and 97.17% for a generalized momentum observer. The authors note that severity-excitation profiles and offline velocity perturbations limit the method's diagnostic capability.

  32. arXiv cs.RO62

    HA-DiPO: A Diffusion Policy Optimization Framework for Heterogeneous Robot Cooperative Transport

    The authors propose HA-DiPO, a hierarchical multi-agent reinforcement learning framework operating under centralized training with decentralized execution (CTDE) for heterogeneous legged robots cooperatively transporting an unfixed shared payload. At the high level, a centralized critic jointly trains each robot's independent coordination policy, and diffusion policies represent multimodal coordination responses; at the low level, embodiment-specific locomotion policies track the coordination commands. The policies are learned from scratch through reward-driven simulation interaction, and at execution they use only local observations with no inter-robot communication.

  33. arXiv cs.RO62

    Keyframe Management for Dense Maps on Memory-Constrained Robots

    The authors propose a keyframe management method for dense maps on memory-constrained robots that directly limits the memory occupied by the resident map and admits and releases keyframes by semantic spatial units such as building rooms and corridors. Across every run on two NVIDIA Jetson boards, the memory budget was never exceeded and model throughput loss was at most 0.8%, whereas without the method the operating system terminates the model or the map. Under that budget, admission by semantic zone localized 6-9 percentage points more query frames than geometric partitioning by the number of matched regions, roughly twice that of RTAB-Map retrieval, with lower ATE.

  34. arXiv cs.RO62

    WorldFailBench: Can World Models Understand Policy Failures?

    The authors propose WorldFailBench, which uses three tasks—failure reason, failure stage, and failure boundary—to test whether world models faithfully simulate policy failures. The benchmark covers manipulation tasks in the LIBERO and RoboTwin simulation environments, as well as real-world manipulation tasks with the G1 humanoid robot. Experiments on seven world models show clear gaps in their understanding of policy failures.

  36. arXiv cs.RO71

    MimicAll: A Structured Embodiment Conditioning Framework for Cross-Humanoid Whole-Body Motion Tracking

    MimicAll is a reinforcement learning framework that encodes global and joint-alignment conditions using structured embodiment descriptors, enabling heterogeneous humanoid robots to share a whole-body motion tracking policy. Training covers 4,096 variants across 16 humanoid robot families, and a benchmark of 464 bodies across five distribution strata was built. With the same amount of training transitions, shared training achieved a 99.4% 30-second success rate on randomized G1 bodies in 100STYLE, versus 42.9% for single-body specialized policies.

  37. arXiv cs.RO62

    ReSight: Task-Driven Self-Improving Memory for Embodied Agents

    ReSight is a task-driven self-improving memory system for embodied agents that retains keyframes and their spatiotemporal associations, and separates observation evidence from extensible derived fields so that historical observations can be reprocessed for new needs. Through Explore-Distill-Mount (EDM), the system identifies capability gaps from execution and exploration records and converts exploration-derived solutions into reusable memory builders, query tools, and skills.

  38. arXiv cs.RO67

    Survey: Dexterous hands, simulators, training, and evaluation in 113 learning-based manipulation methods

    A survey reviews 113 learning-based dexterous manipulation methods, covering the dexterous hands used, the training simulators, and how results are reported. The authors note that nearly half of the works use the same recipe: reinforcement learning on privileged state in a GPU simulator, then distillation into a vision-only student model transferred to hardware, and nearly all of them concentrate on the same simulator.

  39. arXiv cs.RO58

    MDPO: A Multi-Constraint Policy Optimization Reinforcement Learning Method for Dexterous Manipulation

    The authors propose Multi-constraint Dexterous Policy Optimization (MDPO), which formulates dexterous manipulation as a multi-constraint policy optimization problem involving both cumulative cost constraints and state-wise cost constraints, and integrates multi-constraint projection into the policy optimization process. In experiments on ReDMan for safe joint control, finger safety, and grasping tasks, MDPO achieved better constraint satisfaction than baselines while maintaining competitive task performance. The code is publicly available.

  40. arXiv cs.RO56

    SCET-CBF: A Speed-Calibrated Soft Guidance Method for One-Step Dynamic Navigation

    Xiaoyu Han of University College London and Dongkun Han of the Chinese University of Hong Kong proposed SCET-CBF, a speed-calibrated soft guidance law for one-step dynamic navigation. It uses a short lateral escape maneuver to provide an allowable approach speed, and lets a quadratic program weigh the two effects of braking when choosing acceleration and steering.