EmbodiedEvo: Self-Evolving Coding Agents Lift Robot Task Success from 13% to 35%
机器之心 · wechat · 2026-10-15
NaversLab's EmbodiedEvo brings coding agents into the physical world. Agents control robots via Python programs through four primitives (getinstruction, getobservation, step, reset), first exploring freely in a Playground scene, then distilling experience into code Skills (how to act) and Lessons (when and where things go wrong). The resulting harness is frozen and handed to a fresh agent, which tackles unseen scenes without any weight updates.
Key results:
- 45.08 average score and 38.95% success rate on RoboDojo, beating all VLA/WAM, agent, and agent+VLA baselines; Memory and Open tasks reach 69.67% and 53.00% success.
- With the same model config, self-evolving lifts success from 13.36% to 35.08%; stacking and Tic-tac-toe jump from 0 to 100%, Imitation sorting from 0 to 80%.
- Experience transfers across models and tasks: GPT-6-Astra's experience boosts GPT-5.6-Sol by 8.86 points; even mismatched-task experience adds 18.65 points; Skills/Lessons cut token usage by 11.3% vs. full interaction history.
- Fine contact remains unsolved: intubation and key-insertion improve in score but stay at 0% success.
The work adds an accumulable, transferable knowledge path to embodied AI, complementing large-scale VLA trajectory training. Paper: arXiv 2610.18867, code open-sourced.
Related event: EmbodiedEvo: Coding Agent That Evolves Through Physical-World Experience(2 posts)→
More from Embodied
- WEDGE uses navigation agent failures to generate targeted training environments — mohitban47 · 2026-10-16
- Tevel's robotic system harvests fruit around the clock using computer vision — LexiLove · 2026-10-16
- SHOES: 10 minutes of real-robot RL lets quadrupeds walk in shoes, CoRL 2026 paper — kscottz · 2026-10-16
- Tesla Cybercab robotaxis logged zero crashes in first two weeks of public rides, NHTSA data shows — elonmusk · 2026-10-16
- Robotics LLMs tested on 16 tasks: GPT-6 Astra hits 17.2% success vs Fable's 9.1% — pabbeel · 2026-10-16
- WEDGE paper generates training environments conditioned on navigation agent failures — _akhaliq · 2026-10-16