Skip to main content
Search

# Robot Manipulation

23 related articles

Subscribe to updates
  1. arXiv

    A general model commands the arm directly: AMP1 hits 85.6% zero-shot on RoboLab-120, ahead of VLA and WAM policies

    Robot manipulation has relied on specialist policies trained on demonstrations, such as vision-language-action models (VLAs) and world-action models (WAMs). AMP1 keeps a generalist AI model in the control loop and has it command arm motion through code, reaching 85.6% zero-shot success on RoboLab-120 against 49.2% and 42.9% for leading VLA and WAM policies. It succeeded in 104 of 108 real-world trials across twelve tasks and three robot instances, with no demonstrations or task-specific training.

  2. arXiv

    A big model's evolved robot harness still works at 4B: 62.8% success after distillation

    TinyAstral moves the improvement target outside the robot agent's planner: an open-weight teacher is frozen while frontier LLM agents read its trajectories and propose harness edits, each accepted only if it passes a paired statistical test, lifting the teacher's strict RoboTwin success from 33.4% to 63.6%. The final harness is then frozen and the teacher's reasoning and failure-recovery trajectories are distilled into a 4B student, which reaches 62.8% in that same harness and tracks its corresponding teacher across four harness versions.

  3. arXiv

    Robot memory stops guessing from keyframes: per-tick text lines lift success from 44.5% to 66.0% under a fixed policy

    Gamma replaces pixel keyframe memory with auditable text lines, raising average success on RoboMME's sixteen memory tasks from 44.5% to 66.0% under a fixed execution policy and matching the privileged subgoal ceiling on three tasks.

  4. arXiv

    Async replanning stops restarting from independent noise: a shared noise trajectory follows real execution displacement, reaching 96.7% on real-robot cloth folding

    Generative robot policies no longer restart from independent Gaussian noise at every asynchronous replanning step; adjacent action chunks now share an initial noise trajectory aligned to actual execution displacement and correlated along action time, reaching 96.7% success on real-robot cloth folding and 90.0% on object storage.

  5. arXiv

    Incomplete scans become simulation-ready articulated assets: an LLM reads the mesh in meters and cuts joint-axis error to 6.8 degrees

    Robot policies trained in simulation only transfer if the simulated asset matches the real object. USDCraft has a pretrained LLM turn a source mesh into a metric text description, write an executable asset program, and revise it against that same representation without task-specific training. On USDCraft-bench it reaches 84.83% part-matching F1 and 6.833 degrees joint-axis error; on Lightwheel it leads part matching and joint location but trails mesh-segmentation methods on whole-object Chamfer distance.

  6. arXiv

    FOCUS uses KL-guided modality switching to transfer privileged-state training to an RGB-D policy, raising average test success from 0.71 to 0.93 across five IsaacLab manipulation tasks

    FOCUS is a single-stage PPO framework that keeps the critic on privileged state while an actor switches its rollout source between privileged-state and RGB-D latents according to the KL divergence between the two modalities' action distributions, combined with representation alignment; across five IsaacLab manipulation tasks it raises average test success from 0.71 to 0.93 and end-to-end budget-normalized training-success AUC from 0.47 to 0.65 relative to the strongest RGB-D-at-test baseline on each task.

  7. arXiv

    OpenRUA turns an off-the-shelf coding agent into a zero-shot visuomotor policy via ROS 2 terminal access alone, reaching 99.0% on CaP-Bench and 87.0% on LIBERO-PRO

    OpenRUA introduces a zero-abstraction harness that gives an off-the-shelf coding agent (Claude Code powered by Claude Opus 5) only terminal access to the robot's native software interface ROS 2, plus ROS 2 documentation and basic tools, with no prescribed workflow or bespoke interfaces, recasting perception as file I/O and manipulation as coding; it achieves 99.0% success on CaP-Bench and 87.0% on LIBERO-PRO, showing that an off-the-shelf coding agent can serve as a zero-shot visuomotor policy without bespoke primitives or task-specific training.

  8. arXiv

    PatternDex guides bimanual dexterous manipulation with object-dictated interaction patterns, raising Allegro success from 52.2% to 92.8%

    PatternDex learns an embodiment-agnostic "interaction pattern" token sequence from human-object demonstrations and, combined with a target robot hand description, decodes wrist motions and contact points that fit that hand to guide a reinforcement learning policy; on 6 unseen demonstrations from the ARCTIC dataset it reaches a 92.8% mean success rate with Allegro hands versus 52.2% for the baseline ObjDex, transfers to Shadow, Paxini, and XHand hands with 84.0%, 73.5%, and 71.6% after fine-tuning alone with a frozen pattern encoder, and opens a real microwave in 7 of 10 trials.

  9. arXiv

    Vela replaces fixed-rate action chunks with continuous spline trajectories, reaching 45.3% average success on LIBERO-X, 49.7% overall on EBench, and 67.5% mean subtask success on a real dual-arm egg-cake task

    Vela is a vision-language-action foundation model pretrained in trajectory space that represents future motion with a fixed number of cubic B-spline control points plus a motion-dependent temporal span; pretrained on roughly 20,000 hours of public and roughly 20,000 hours of private robot data, it reaches 45.3% average success on LIBERO-X, 49.7% overall success and a 66 task-progress score on EBench, and 67.5% mean subtask success on a real wheeled dual-arm egg-cake cooking task.

  10. arXiv

    SWAP routes stepwise among VLA policies via offline RL, lifting real-world DROID task success from 63.3% to 96.7%

    SWAP casts robot policy routing as an offline reinforcement learning problem, training an IQL routing critic that picks, after each action chunk, the candidate pretrained VLA policy with the highest predicted long-horizon return given the current observation; on real-world DROID tabletop manipulation it raises success from 63.3% for the best single policy to 96.7% while cutting average action steps for successful trajectories from 269 to 193 (a 28.3% reduction), and it outperforms fixed-policy and random-routing baselines in LIBERO and LIBERO-Plus simulation.