PhysEvo: recursive self-improvement around a frozen model hits 62% on RoboDojo, 84% on real robot
Wenqing Tian · hf · 2026-10-08
PhysEvo is a framework for physical recursive self-improvement around a single frozen model: a task agent executes robot tasks while a meta-agent diagnoses failures from trajectories, revises tools and skills, and tests corrections — even improving its own diagnostic tools — with no weight updates or separately trained policy.
Results:
- 62.00% success and 68.14/100 average across 42 held-out RoboDojo tasks vs 47.17% for RoboDawn's one-shot Astra agent.
- 55.00% success on 8 tasks challenging direct Astra vs 1.25%.
- Deployed on a real AgileX PiPER with continued skill revision: 84.00% success over 25 trials on 5 real-world tasks.
The core idea: turn the consequences of action into persistent, testable changes to how a frozen model acts and improves.
More from Embodied
- Humanoid robots now walk in 3-inch Louboutin high heels, a brutal balance test — kscottz · 2026-10-08
- RAI Institute's Roadrunner robot blends wheels and legs with a single unified policy — Olivier__OG · 2026-10-08
- Burkov: a humanoid robot that can't do this with its hands is useless — burkov · 2026-10-08
- KingKong Robotics open-sources software toolkit for its six-legged crab robot Jumper — kevin_zakka · 2026-10-08
- Developer uses Codex for CAD and build docs in physical robot hardware project — OpenAIDevs · 2026-10-08
- Twitter user dance-battles a Unitree robot in real life — zealcaiden · 2026-10-08