RoboDoctor
Paper arXiv CodeBibTeX

RoboDoctor

Training-Free Recursive Self-Improvement
for Robot Policies

Heyang Yu1,*, Haomin Li1,*, Sihang Li2, Jixin Yan3, Shiyun Peng1, Wanqi Zhong1, Tianyu Xin1, Yuan Yao1, Xiaoyu Tian1,†, Yiming Li1,†

1 Tsinghua University2 New York University3 Northwestern University

*Equal contribution†Equal correspondence

Paper arXiv CodeBibTeX
pour_liquid_into_cup · iter3 · tick 0

robodoctor / agent

01 / 03

02 / 03

03 / 03

01 / 03
Harness test · revise · select control mode HybridDirect Agentinspect · ground · verify WAMweights fixed Robot action chunk · unedited prefix review or: bounded correctiongrounded action Fresh observation verify re-ground, then reuse Experience memoryInterventionOperation

head RGB-D grounded targetx, y, z proposed path FK preview offset bound released prefix suffix discarded → new proposalchunk withheld action chunk

D0002 ↔ I0002

01 / 05

Control compatibility and intervention cost

Hybrid control does not eliminate the mismatch between an agent’s intended correction and the policy’s next action. Excessive intervention can disrupt useful motion, while repeated grounding, planning, and verification add latency. Future work should study intervention budgets, controller compatibility, and the trade-off between outcome quality and execution cost.

Control changes do not resolve every task

Changing how control is organized does not resolve every task. In several tasks, switching to agent-led Direct control after a weak warmup still produced no successful episode. In such cases the remaining difficulty may lie in perception, precise contact or task-specific success conditions rather than in who controls the robot. Addressing them will require stronger low-level skills, richer outcome measurements and more explicit checks of task completion.

Generalization and evaluation

The current page contains local results for ten tasks, whereas official reference bars use full-benchmark aggregates. These figures do not establish a matched-coverage ranking or isolate the contribution of each RSI component. Incorporating the complete local results, isolating harness, intervention and operation updates through controlled ablations, and testing across policies and environments would clarify robustness and transfer.

  1. PrismWAM

    PrismWAM: Decoupling State Representations in World Action Models

    Tianyu Xin, Wanqi Zhong et al. · 2026 · Project website

  2. RoboDojo

    RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

    Tianxing Chen et al. · 2026 · arXiv:2607.04434

  3. GPT-as-Policy

    Systematically Exploring the Capabilities of GPT-6 Astra as Embodied Policies

    Galbot Team · 2026 · arXiv:2609.38537

  4. PhysicalRSI

    Physical RSI 1.0: Recursive Self-Harness for Scaling Embodied Skills

    MMLab & PhysicalRSI contributors · 2026 · Project website

  5. RoboICL

    RoboICL: Embodied In-Context Learning with GPT-6 Astra

    Fangcheng Liu et al. · 2026 · arXiv:2609.34261

  6. GPT-6 Astra on RoboDojo

    An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond

    Wenbo Zhang et al. · 2026 · arXiv:2609.24170

  7. EmbodiedRSI

    EmbodiedRSI: Active Continual Robot Learning Through Hypothesis-Guided Co-Evolution

    Python Song et al. · 2026 · arXiv:2610.10498

@misc{yu2026robodoctor,
  title        = {RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies.},
  author       = {Heyang Yu, Haomin Li, Sihang Li, Jixin Yan, Shiyun Peng, Wanqi Zhong, Tianyu Xin, Yuan Yao, Xiaoyu Tian, Yiming Li},
  howpublished = {https://robodoctor-rsi.github.io/},
  url          = {https://robodoctor-rsi.github.io/},
  note         = {Project page},
  year         = {2026}
}