rsi.watch

Heyang Yu et al. (10 authors) · 2026-10

RoboDoctor: Training-Free Recursive Self-Improvement for Robot Policies

“The overall score increases from 49.74 in Warmup to 57.10 in Refinement, corresponding to a gain of 7.37 points.”
Who
Heyang Yu et al. (10 authors)
Date
2026-10 (month only) The technical report and project page give no publication date. The code repository was created on 2026-10-21, per the GitHub API.
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ selective reporting. The archive picks among candidate outcomes by score, keeping the highest valid one. The authors list retrospective selection among the limits on attribution. “The archive includes score-based selection among candidate outcomes, retaining the highest valid score with deterministic tie-breaking.”
  • ↓ risk of bias. Warmup and Refinement draw on several source runs; the report says a uniform agent version, checkpoint and inference budget across them is not established. “The local archive combines multiple source runs for which a uniform agent version, checkpoint, and inference budget are not established.”
  • ↓ imprecise. Refinement has at most six episodes per task and 301 of 324 are valid; the README says missing or invalid episodes are excluded. No uncertainty is given for the scores. “n: valid Refinement episodes out of six nominal slots”
how we rate evidence
Replication
No rerun found. The code is public. We found no rerun by another group. Searched up to 2026-10-26. all claims
Checked
2026-10-26, against the source page
Links
Source Code

Also on the page

  • Scores are on 54 RoboDojo-Sim task entries. Policy and agent weights stay fixed. In the benchmark runs, a Warmup-score rule picks one of two predefined controllers; continuing the hybrid controller comes "with harness self-revision based on the Warmup stage".
  • The report says Warmup "already incorporates Doctor assistance and experience updates", so it is the system's initial stage, not an unassisted baseline.
  • The project page ranks the Refinement score #1 / 9 against RoboDojo's official agent entries. The report compares against the leaderboard as archived on 2026-10-20 and says interaction budgets are not matched. When we checked, RoboDojo's public leaderboard did not list RoboDoctor.
  • Authors' stated limits: "All results are in simulation, and transfer to other policies or physical robots is unmeasured." Across tasks, 24 improve, 20 decline and 10 stay unchanged from Warmup to Refinement.

← Ledger Report an error