Report
EmbodiedEvo gives fresh coding agents reusable robot-control skills without updating model weights
Researchers say it builds reusable code and lessons in one simulated scene for fresh agents to use in unseen scenes.
TLDR
The EmbodiedEvo researchers describe a coding agent that revises robot-control programs using feedback from a simulation, then passes its code skills and diagnostic lessons to fresh agents. Their paper reports a 44.48 average score and 37.82% success rate on RoboDojo-Sim, while noting that fine contact control remains difficult.
Combined views
791
1 Source, first seen ago
