Skip to content
#Robot models and learning
Oct 28Wed
  1. arXiv cs.RO60

    Reinforcement Learning-Based Autonomous Fixed-Wing Drone Interception

    A team at Delft University of Technology in the Netherlands proposed a fixed-wing interceptor controlled by a neural network that maps filtered monocular vision observations directly to actuator commands; the policy was trained in a custom Python environment using reinforcement learning and proximal policy optimization (PPO). In simulation, the policy outperformed two classical guidance controllers under most conditions, and it has successfully intercepted maneuvering fixed-wing targets in real flight. This work provides validation for the autonomous control of fixed-wing interception drones.

  2. arXiv cs.RO70

    MARL-Exo: Multi-Agent Reinforcement Learning Co-Adapts Hip Exoskeleton Control to Reduce Metabolic Energy Cost

    The study proposes MARL-Exo, a multi-agent reinforcement learning framework in which a full-body musculoskeletal model and a bilateral hip exoskeleton policy are co-trained in simulation in two stages; the exoskeleton observes only delayed hip kinematics and past commands. In simulation, the metabolic policy reduced modeled muscle metabolic power by 13.8% compared with walking without the device, and by 8.8% after adding mechanical-power shaping.

  3. arXiv cs.RO62

    QCP: Learning Quality-Controllable Robot Policies with Iterative Hindsight Relabeling

    The authors propose Quality-Controllable Policy (QCP), which enables inference-time control over individual quality attributes such as execution speed and motion smoothness, as well as their combinations. To extend the controllable quality range beyond the initial demonstrations, the authors design an iterative hindsight relabeling loop: collect online rollouts under target quality labels, add successful trajectories to the dataset, then relabel the accumulated trajectories in hindsight according to their measured quality and fine-tune the policy.

  4. arXiv cs.RO65

    PDRVLA: A vision-language-action model with physical dynamics representations

    PDRVLA proposes a lightweight Physical Dynamics Module that extracts physical tokens from vision-language features to condition action generation, without online video generation. The module learns through direct prediction of future RGB, depth, and optical-flow features, with auxiliary supervision from a frozen pretrained video world model. All training data comes from public data and was expanded through human-to-robot synthesis, subtask annotation, and embodied chain-of-thought (ECoT) annotation.

  5. arXiv cs.RO62

    VLA-Stem: Adding a Lightweight Safety Mechanism to Vision-Language-Action Models with Reinforcement Learning

    Researchers propose VLA-Stem, a lightweight reinforcement-learning-based safety mechanism that does not fine-tune the underlying VLA policy; instead, it converts its action chunks into motion references and, at inference time, adaptively corrects execution using proprioception and local obstacle information. Across six constrained simulation settings, VLA-Stem achieved an average collision-free task success rate of 52.8%, versus 23.7% for the strongest safety baseline, while preserving the original policy's performance in obstacle-free environments; real-robot experiments also showed reduced collisions under static and dynamic obstacles.

  6. arXiv cs.RO58

    Tokens of Motion: Primitive Motion Learning for Human-Humanoid Skill Transfer

    The paper proposes the Adaptive Failure-Informed Learning (AFIL) framework, which uses a pretrained VLA to generate failure trajectories online as adaptive negative guidance for diffusion- and flow-based VLA policies, avoiding the need for handcrafted failure modes or human intervention for recovery. The method jointly trains two action generators (DAGs) for success and failure on a shared vision-language backbone, and at sampling time adjusts the guidance strength according to the distance between the success and failure distributions at each diffusion step. On in-domain and out-of-domain, short-horizon and long-horizon robotic manipulation tasks, AFIL improves task success rate and robustness compared with existing VLA baselines.

  7. arXiv cs.RO55

    Spiking Neuromorphic Controller Stabilizes Periodic Orbits in Underactuated Robots: A Greedy Lyapunov-Based Approach

    This paper investigates event-driven spiking controllers implemented as spiking neural networks for trajectory tracking in underactuated robots, with the goal of stabilizing a prescribed periodic orbit. The authors reduce orbital stabilization to stabilizing the trivial solution of a periodic linear time-varying system via transverse linearization, and extend a Lyapunov-based greedy design originally developed for LTI systems to this periodic LTV transverse dynamics, thereby constructing a feedback controller. The method's effectiveness is demonstrated on the trajectory stabilization problem for the butterfly robot.

  8. arXiv cs.RO71

    Retrieve to Act: Improving VLA Execution Robustness with Motion Primitive Graph Retrieval

    The authors propose Retrieve to Act (R2A), a test-time scaling framework that organizes successful demonstrations into an offline spatiotemporal motion primitive graph. At inference, a query-dependent graph neural network locates the current execution state and retrieves structurally compatible future motion primitives, which are then fused with the frozen VLA's predicted actions through dimension-wise gating.

  9. arXiv cs.RO53

    LSMM: Learning Spatial Memory for End-to-End Legged Robot Exploration

    The authors propose Local Spatial Matrix Memory (LSMM) based on Hadamard memory, which lets reinforcement learning agents store task-relevant information in position-based memory cells. On an autonomous maze exploration task, LSMM's maze coverage is comparable to strong recurrent networks and finite-history baselines, and it is the only architecture that remembers the maze layout, reducing reconstruction error to one third. The work targets the problem that humanoid and legged robots in complex environments need to remember information long after it leaves their field of view before it is used.

  11. arXiv cs.RO67

    How Much Can Best-of-N Improve a World Action Model: Measuring the Selectable Space and Realizing It in Closed Loop

    The paper proposes a denoised counterfactual protocol that uses bitwise-reproducible branching, repeated continuations, held-out oracles, and common random numbers to measure the selectable space a world action model (WAM) can achieve when selecting candidate actions at test time, and builds a fail-gated selector on that basis.

  14. arXiv cs.RO54

    State Estimation for a Hybrid Rigid-Continuum Space Manipulator System

    This study proposes a state reconstruction pipeline for a hybrid rigid-continuum space manipulator whose distal link is a tendon-driven continuum robot. The method formulates the estimation problem as maximum a posteriori inference conditioned on joint encoders and vision measurements, transcribes it as a factor graph, fuses heterogeneous measurements with the rigid-soft kinematic model, and uses the Geometric Variable Strain (GVS) approach to obtain a minimum-order parameterization of the configuration. Closed-loop trajectory simulations verified the state reconstruction accuracy and computational performance.

  15. arXiv cs.RO64

    Omni-IGN: Omnidirectional Vision Onboard Image-Goal Navigation System for Drones

    The authors propose Omni-IGN, a fully onboard drone image-goal navigation (ImageNav) system that integrates quad-fisheye perception, an attention-based reinforcement learning policy, and a collision-aware safety layer. The policy adaptively fuses goal-relevant features across multiple views to provide directional guidance for continuous navigation, and the quad-fisheye RGB images support both ImageNav and obstacle avoidance without a dedicated depth sensor.

  16. arXiv cs.RO56

    When Black-Box Robot Policies Fail: Online Sensor Diagnosis and Repair with a Transparent Surrogate

    The study proposes a structured surrogate model that separates periodic gait from sensor-dependent correction terms, diagnosing and repairing corrupted inputs externally around a frozen locomotion policy. On simulated Go2, offline diagnosis raises walking success from 259/640 to 429/640, while online repair reaches 382/640, with no repair triggered in 64 held-out fault-free episodes. The authors note that the evidence remains limited to a single policy and a partly screened simulated fault set, and that robot foundation model and hardware validation have not yet been conducted.

  18. arXiv cs.RO62

    HC2Robot: Multi-Terrain Humanoid Parkour with Predicted Foresight and Hyper-Connection Transformers

    HC2Robot proposes a deployable perceptive humanoid parkour framework composed of a Predicted Foresight Module (PFM) and a Hyper-Connection (HC) Transformer. The PFM predicts, from deployable observations, the timing, location, and support quality of each foot's next touchdown event, converting privileged supervision into a deployable latent representation.

  20. arXiv cs.RO36

    Integrated Design and Reproducible Feasibility Analysis of a Four-Wheel Rover: Packet-Commanded Imaging, LiDAR Ranging, and Learning-Assisted Target Recognition (NASA HERC 2027)

    For the high-school remote-control division of the 2027 NASA Human Exploration Rover Challenge, a study presents an integrated, reproducible design assessment of a four-wheel educational rover: 10.919 kg, 215 mm wheels at 200 RPM give a theoretical rolling speed of 8.105 km/h, and the nominal motor specifications are insufficient to demonstrate compliance with the 8 km/h design ceiling.

  22. arXiv cs.RO58

    An Auditable Framework for Autonomous Driving Laboratories: LAM, OpenSDL, Safe-SDL, and L3-Bench

    The authors propose a traceable framework comprising four components—LAM, OpenSDL, Safe-SDL, and L3-Bench—to connect experiment planning, automated execution, and measurement feedback. LAM is a measurement-conditioned policy architecture that extends the vision-language-action paradigm to generate both continuous robot trajectories and typed instrument API calls; OpenSDL maps role-level operations to heterogeneous devices and records timestamped causal events.

  23. arXiv cs.RO55

    Learning Side-by-Side Navigation for Personal Mobility Vehicles from Human Drivers

    The paper proposes a method for enabling two personal mobility vehicles (PMVs) to drive autonomously side by side: after a global path is given, each vehicle runs its own sampling-based model-predictive planner, the two exchange only the odometry information each can observe of the other, and a common rule is used to make a joint decision on a utility matrix over pairs of candidate trajectories.

  24. arXiv cs.RO67

    TOPS: Reshaping Flow Source and Transport for Multi-Task Robot Manipulation

    TOPS proposes a task-oriented prior and scheduling method that adapts the flow-matching source prior and interpolation schedule for each task without modifying the network architecture. On LIBERO-10, single-step inference raises the success rate from 53.44% for standard flow matching with ten steps to 69.67%, with inference time about one-sixth of the baseline. The method improves performance on LIBERO, LIBERO-Plus, and CALVIN, and when transferred to the π0.5 policy, and on a real robot it raises the success rate on an upper-drawer-opening task from 20% to 80%.

  25. arXiv cs.RO60

    RoboSTD: A Bimanual Learning Method for Zero-Shot Single-to-Dual Transfer via Sagittal Mirroring

    RoboSTD is a zero-shot single-to-dual data construction method for bimanual robot learning. It applies sagittal-plane mirroring to observations, proprioception, and actions to generate executable contralateral pseudo-demonstrations, then uses an LLM to generate structured spatio-temporal coordination constraints that guide the reconstruction of pseudo-bimanual supervision across workspace regions and task stages. The study evaluates RoboSTD on the AgileX ALOHA platform in RoboTwin simulation and real-world settings. Without additional bimanual demonstrations, it achieves an average of 2.11 times the number of cups collected by the single-arm baseline in the cup-collection task.

  26. arXiv cs.RO60

    From Biological Curves to Modular Soft Robots: Learning-Based Inverse Design with Uncertainty-Guided Refinement

    The paper proposes a learning-based real-to-sim-to-real framework that converts biological curves into cable-driven modular soft robot configurations through surrogate-assisted inverse design and uncertainty-guided refinement. The learned forward surrogate reduces deformation prediction time from 1.53×10^5 ms to 0.0008 ms, a speedup of about 1.9×10^8 times; inverse design time drops from 1.99×10^6 ms to 4.18 ms.

  27. arXiv cs.RO58

    SWIM: Learning State Weights from Behavior to Train JEPA World Models

    SWIM proposes using an inverse dynamics model to learn state feature weights from offline trajectories for state-supervised training of JEPA world models. The method predicts recorded actions from current states and observed future displacements, quantifies each state feature's contribution to action prediction, and uses this to build a weighted state metric that aligns the latent space with state distances. The inverse dynamics model is used only during training; planning still relies solely on images. On five simulated robotic manipulation tasks, SWIM's planning success rate consistently outperforms uniform state weighting.

  28. arXiv cs.RO67

    HumiTransfer transfers human-human interactions from in-the-wild videos to humanoid robots

    The authors propose HumiTransfer, a framework that converts in-the-wild human-human interaction videos into physically feasible human-humanoid interaction motions. The method first reconstructs 3D human motion with foot-ground contact, then transfers one participant to the target humanoid embodiment through interaction modeling, and uses post-retargeting optimization to ensure balance, contact, and collision avoidance.

  29. arXiv cs.RO71

    REACH: An RL-based hierarchical humanoid crawling framework for confined-space traversal

    REACH is an RL-based hierarchical framework that enables a Unitree G1 humanoid robot to traverse height-constrained environments through variable-height bipedal walking and crawling. Its policies regulate body height over complementary command ranges: 1.27–0.50m for bipedal walking and 0.50–0.20m for crawling, and a bidirectional transition policy connects crouched walking and crawling. The framework also combines a LiDAR planner that converts point clouds into collision-free three-dimensional paths and locomotion commands. Real-world experiments validated height and speed tracking, terrain traversal, mode switching, and confined-space traversal; the code and videos have been released.

  30. arXiv cs.RO64

    CARD: A Contact-Anchored Retargeting Framework for Dexterous Manipulation

    CARD is a simulator- and object-model-free framework that retargets human demonstrations into robot actions by explicitly preserving contact force targets, and it handles video, motion capture, and tactile glove data in a unified way. On video, glove, and motion-capture data and the GRAB dataset, CARD maintains contact on 43%–78% of contact frames, whereas purely kinematic and object-aware retargeting methods stay below 10% and 22.5%, respectively.

  32. arXiv cs.RO58

    MyoAcross: A Generalist Muscle-Driven Control Policy for Musculoskeletal Models Across Individuals

    The study proposes the MyoAcross framework, which combines individual musculoskeletal model construction, morphology-conditioned policy learning, and virtual model augmentation to achieve muscle-driven reference-motion tracking without per-individual fine-tuning. Across 31 held-out individual models, it achieved a 99.24% completion rate and reduced tracking error by 22.60% relative to an unconditioned multilayer perceptron. Policy reuse also extended to 22 held-out individual models constructed from single-image shape estimates, height, and body mass, as well as to a model with unilateral lower-limb shortening after asymmetry augmentation.

  33. arXiv cs.RO69

    RoboAuto: Autonomous Robot Program Search through Multi-Level Physical Understanding

    RoboAuto proposes a robot program search framework that organizes physical understanding at three levels—object, scene, and interaction—to guide LLM agents in generating and revising manipulation programs. The method achieves better performance across five benchmark suites, reaching an 89.1% success rate on short-horizon LIBERO-PRO tasks and averaging more than twice the success rate of ASPIRE on LIBERO-PRO Long and LIBERO-Plus.

  34. arXiv cs.RO62

    GRAFT: Geometry-Aware Residual Adaptation for Frozen Humanoid Behavior Trackers

    GRAFT proposes a geometry-aware residual adaptation framework that, without changing the parameters of a frozen humanoid behavior foundation model (BFM) or its motion-reference interface, uses a lightweight adapter to apply terrain-conditioned corrections only to lower-body execution targets. The adapter is trained with task and foot-geometry rewards, is regularized so it produces no unnecessary corrections on flat ground, and is distilled from privileged terrain observations to onboard depth.

  35. arXiv cs.RO71

    Whole-Body Humanoid Control via Stochastic Motion Tokenization

    This study proposes using a stochastic motion tokenizer to generate concise discrete motion representations that can be fed directly into a whole-body tracking policy to execute dance, fighting, and walking motions on the Unitree G1. At matched capacity, the tokenizer reconstructs Unitree G1 motions better than FSQ and VQ-VAE. In real-robot demonstrations, each generated motion lasted 16.8 seconds, and a 35-second fighting-to-dance sequence was completed.

  36. arXiv cs.RO62

    HIDRA: A Hierarchical Dual-Routing Attention Framework for Replay-Free Lifelong Imitation Learning

    The authors propose HIDRA, a hierarchical dual-attention routing framework for replay-free, single-pass lifelong imitation learning that decouples instruction-level expert retrieval from context-dependent refinement. The method introduces key-level geometric regularization that constrains within-task alignment and between-task separation to mitigate routing interference under representation drift. On several LIBERO benchmark suites, including variants with heterogeneous task sequences and paraphrased language instructions, HIDRA outperforms replay-free and replay baselines on AUC and forward transfer while maintaining low forgetting, and the code is publicly available.

  37. arXiv cs.RO42

    ASPEN: Adaptive Semantic Reasoning via Command Persistence Prediction to Improve Navigation VLA Efficiency

    ASPEN is an event-triggered scheduler based on command persistence prediction, designed to reduce repeated semantic reasoning calls in navigation VLA policies. On frozen NaVILA in the cross-dataset RxR-CE val-unseen evaluation, ASPEN reduces VLA calls by 9.5% and estimated inference time by 8.1% (including scheduling overhead), while preserving action sequences identical to the baseline in 92.1% of episodes.

  38. arXiv cs.RO65

    Learning Multi-Phase Robot Loco-Manipulation via Phase-Conditioned Policy Optimization

    The study proposes extending PPO into phase-conditioned policy optimization, using a phase-conditioned critic, phase-gated rewards, and a multi-phase reset curriculum to train a single policy for long-horizon loco-manipulation tasks. The method adopts a decoupled hierarchical control architecture that separates low-level balancing dynamics from high-level phase execution, and validates multi-stage skills on real hardware on the Unitree G1 and a Boston Dynamics Spot equipped with a manipulator arm, while testing design choices through ablation experiments.

  39. arXiv cs.RO55

    Program-Level Adversarial Co-Adaptive Self-Play for Robot Security

    The paper proposes a program-level adversarial co-adaptive self-play framework that lets coding agents generate attacks and defenses based on opponent code and task feedback, without model fine-tuning. The framework provides external program interfaces for autonomous driving and SDK-level defense interfaces for humanoid robot locomotion; two eight-generation case studies show that the driving defender introduced direct-actuation recovery, while the humanoid defender revised incoming-state validation. These responses provided only local task recovery and did not establish sustained robustness.

  40. arXiv cs.RO72

    Rollback: Using World-Model Experience Backtracking to Improve Robot Policies

    Rollback proposes a policy-improvement paradigm in which a world model rewinds to critical decision points, branches to try different actions, and compares the outcomes, distilling counterfactual results into action-level supervision without a separately learned value model. On three real-world tasks—block extraction, ball throwing, and tabletop curling—success rates improved by 10.67%, 11.11%, and 11.85% in absolute terms over the strongest baseline, and ball throwing and tabletop curling surpassed the average human player success rate within 8 hours of training.