Skip to content
#Robot models and learning
Oct 27Tue
  1. arXiv cs.RO62

    PushAround: Physics-Informed Hybrid Search for Multi-Robot Collaborative Path Clearing

    PushAround proposes a physics-informed hybrid search method that combines geometric connectivity reasoning with short-horizon simulation to generate executable clearing plans for mobile robot teams. The method uses repeated W-clearance connectivity graph (WCCG) queries to identify frontier gaps that need widening and ranks them by estimated downstream cost, then generates direct or recursive pushing candidates, selecting obstacles, directions, contacts, and forces.

  2. arXiv cs.RO55

    Factor-Graph-Based 3D Force Estimation for Dual-Arm Manipulation of Flexible Cables

    This study proposes using factor graphs to probabilistically model deformable linear objects (DLOs), estimating the magnitude, direction, and location of 3D forces acting on them while simultaneously reconstructing the cable's shape. Because cables are thin and unactuated, strain gauges and actuation readings are generally impractical; the method uses only one type of sensing information and performs force estimation in simulation tests that add measurement noise and have the cable move along a prescribed path. The authors state that this general method can apply to a broader range of tasks.

  3. arXiv cs.RO60

    VISTA: An Active Exploration Framework for Self-Improvement in Embodied Agents

    VISTA is a meta-agent framework that replaces post-hoc replay reflection with active self-exploration to improve the harness of embodied agents. It maintains a harness graph that records the lineage of harness versions, enabling the meta-agent to retrieve past hypotheses, reuse effective mechanisms, and avoid repeating failed directions. Each round starts from a retained version and its execution evidence, autonomously generates new rollouts, probes failure and recovery behavior, and on that basis iteratively revises the harness and re-evaluates it.

  5. arXiv cs.RO62

    LECO: Learning Direct Control from Optimization for Dynamic Target Tracking

    LECO converts online trajectory optimization for dynamic target tracking into offline learning of a direct-control policy, using a differentiable kinematic model to let trajectory-level objectives directly guide control learning. Across three simulated obstacle densities, its root-mean-square linear jerk is 14.1%–31.2% lower than the best comparison method on this metric, it achieves the highest tracking success rate, and its per-cycle computation time is 4.25–5.63 ms.

  6. arXiv cs.RO58

    MMGNet: A Multimodal 6-DoF Grasp Synthesis Network for Cluttered Scenes

    MMGNet proposes a multimodal grasp synthesis network that uses hierarchical image-to-3D feature injection and rotational graspness modeling to integrate multi-scale image features into 3D representations and jointly predict grasp approach direction and in-plane rotation. On GraspNet-1Billion with RealSense data, it achieves a mean AP of 60.34, surpassing all compared methods; in real-world experiments with adversarial grasp objects, its grasp success rate is 8.46 percentage points higher than AnyGrasp. Ablation studies show that the two modules provide complementary gains. Code, models, and data will be released after the paper is published.

  7. arXiv cs.RO60

    Disturbance-separated model predictive control for airborne electro-optical gimbals

    The paper proposes a disturbance-separated model predictive control (MPC) scheme for constrained pointing control of airborne electro-optical gimbals under carrier vibration, input delay, and actuator constraints. The scheme uses an adaptive periodic observer combined with linear-regression frequency estimation to predict multi-harmonic vibrations, and a linear observer plus a bounded LSTM network to estimate aperiodic disturbances and their changes over the input delay; periodic predictions are embedded in the MPC model, while aperiodic compensation is applied outside the optimizer.

  8. arXiv cs.RO59

    Study: Behavioral quality of human behavior models does not equal robot learning utility

    The paper evaluates multiple types of human behavior models using common generative model metrics in two scenarios, high-speed driving and object handover, and compares them with the downstream performance of reinforcement learning policies trained using these models. The results show that existing behavioral evaluation metrics cannot reliably predict downstream policy performance, and that evaluation results are sensitive to the trajectory representation; even when the true test behavior distribution is known, training with a different distribution may yield better downstream performance.

  9. arXiv cs.RO67

    SceneAgent: An agentic pipeline from 3D and 4D captures to simulation environments

    SceneAgent is an agentic pipeline that converts 3D Gaussian Splatting, photogrammetry, LiDAR captures, AI-generated 3DGS scenes, and 4D dynamic videos into interactive simulation environments through simulator-based visual inspection and repair. The pipeline covers object segmentation, physical parameter estimation, part decomposition, articulation, and deformable simulation.

  11. arXiv cs.RO60

    OSMA-Bench: Evaluating Safety Under Obstacle Conditions in Household Robot Navigation

    The authors propose OSMA-Bench, a benchmark that uses spatially matched scenario sets differing only in obstacle type to test whether robot navigation adjusts clearance and speed according to obstacle semantics, and they propose the Semantic Sensitivity Index (SSI) to measure this adaptation. Evaluating VLM and vision-language-action (VLA) policies on the benchmark, they find that even with explicit safety prompts, neither consistently adjusts its motion according to obstacle safety levels.

  12. arXiv cs.RO60

    Dream2Assist: Intent-Conditioned World Models for Shared Vehicle Control

    The authors propose Dream2Assist, a model-based reinforcement learning framework that adds intent inference to a recurrent state-space world model, enabling an assistive policy to anticipate and support discrete human goals without taking over human control. The work also presents a complete deployment pipeline from world-model policy to a safe human-machine system, addressing the sim-to-real-human gap through policy down-selection and reward restructuring, and validates it in two human-subjects studies, high-speed racing (N=30) and urban driving (N=10), plus a preliminary on-vehicle deployment, where intent conditioning produced measurable improvements in torque alignment and collision rate.

  13. arXiv cs.RO71

    RobotAgentBench released: a benchmark for open-world robot autonomy

    The authors propose RobotAgentBench (RAB), a benchmark for evaluating robot agents in realistic open-world household environments. Using an agent-assisted, human-in-the-loop real-to-sim pipeline, it constructs 50 household layouts and includes 90 atomic tasks, 18 short-horizon tasks, 8 long-horizon tasks, and 4 open-ended scenarios.

  14. arXiv cs.RO60

    What Does World Modeling Actually Improve in Robot Action Learning?

    The study conducts controlled future-video prediction loss ablations in FastWAM and OpenWAM, raising success rates by 24.38 and 72.91 percentage points, respectively, under clean conditions on seen tasks. Trajectory clustering shows higher success rates within several coarse trajectory clusters, but cluster frequencies change little, indicating the gains do not come solely from changes in the distribution of trajectory patterns.

  15. arXiv cs.RO46

    Dual Gradient Consensus (DGC): An Adaptive Learning Method for Transient Out-of-Distribution Disturbances in Quadrotor Control

    To address the problem that short-term out-of-distribution residuals can corrupt the learned disturbance model during quadrotor trajectory tracking, DGC improves the online update step size while retaining the random Fourier feature model, the offline BPTT policy, and least-squares adaptation (DAPL). It distinguishes adaptation states by comparing long- and short-term gradient memories and uses a confidence map to suppress unreliable updates.

  16. arXiv cs.RO52

    A Multimodal Social Perception Framework for Human-Robot Interaction in Virtual Reality

    The study proposes a multimodal framework integrating speech, vision, and touch to generate contextually appropriate robot speech and movement. The team tested it in virtual reality with 18 participants across five social interaction scenarios and compared it against a human teleoperation baseline: the autonomous system scored above the neutral midpoint on both naturalness and social appropriateness, but the teleoperation baseline received significantly higher ratings on both measures. Behavioral coding and feedback showed that conversational timing, clarification of ambiguous requests, and spatial coordination were key factors behind the gap.

  19. arXiv cs.RO59

    AuRA: Reconstructing Digital Twins from Atoms to Bits through Autonomous Robot Interaction

    AuRA is an automated robot interaction and data preparation system for digital-twin reconstruction, combining execution-aware grasp selection, gripper-motion measurements derived from robot states, and registered pre-interaction geometry and appearance priors. Across 13 deformable objects, it reduces active human labor and wall-clock collection time per successfully reconstructed asset by 88.5% and 55.4%, respectively; compared with the human-collected PhysTwin baseline, it improves 7 of 9 metrics for future prediction under known gripper motion, including a 28.1% reduction in one-way L1 Chamfer distance.

  20. arXiv cs.RO65

    ThinkLocoManip: A Whole-Body Control Agent Framework for Legged Loco-Manipulation

    The authors propose ThinkLocoManip, a whole-body control agent framework for legged loco-manipulation that uses an interleaved reasoning-and-acting paradigm to handle long-horizon tasks in large-scale environments. The framework orchestrates perception, navigation, manipulation, and locomotion as composable tools, and it designs cross-domain coordination skills and a keyframe thinking mechanism to support error recovery. In real-robot experiments, its success rate was 77%, more than double the 37.1% of the strongest plan-then-act baseline.

  21. arXiv cs.RO77

    RynnVLA-Latent: Scaling Robotic Manipulation Foundation Models with 3D-Grounded Latent Actions

    The research team proposes the RynnVLA-Latent framework, which uses RynnLAM to learn continuous, 3D-grounded latent actions from visual changes without native robot control signals, converting 30.1K hours of human, robot, and simulation video into 745.52M latent action descriptors for policy pretraining.

    Why it matters: This research uses 3D-grounded latent actions to turn heterogeneous interaction videos into pretraining supervision, offering a reusable path for VLA models to bypass the scarcity of action annotations.

  22. arXiv cs.RO58

    ScheNav: A Sketch-Guided Embodied Object Navigation Framework

    The authors propose ScheNav, a closed-loop navigation framework that turns rough sketch guidance into exploration and stopping decisions. Its core is SchemaField, a learning module that fuses guidance with accumulated observations and predicts spatial value fields. A supporting data pipeline converts 2D floorplans into navigation training environments, producing the Furnished-RPLAN dataset, which contains over 33k furnished layouts with paired navigation episodes and noise-controllable sketches.

  23. arXiv cs.RO58

    AdaPolicy: Action-Driven Anticipatory Robot Policies via Next-Chunk Prediction

    AdaPolicy proposes an action-driven anticipatory policy framework that shapes future-aware representations through cascaded next-chunk prediction, requiring no encoding of future observations during training and adding zero inference-time overhead. The method achieves consistent gains on MetaWorld, RoboTwin 2.0, LIBERO, and real-world deployment, with improvements of up to 14 percentage points on millimeter-precision tasks such as headphone assembly and shoelace threading.

  25. 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.

  26. 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).

  27. 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.

  29. 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.

  30. 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.

  31. 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.

  32. 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.

  33. 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.

  35. 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.

  36. 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.

  37. 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.

  39. 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.

  40. 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.