Coding agents that control robots and keep what they learn as reusable skills and lessonsIntermediate
EmbodiedEvo: Self-Evolving Coding Agent For Physical Intelligence
A coding agent that first practises on one scene per task and saves what worked lifted its success rate on the RoboDojo-Sim benchmark from 8.50% to 30.17% in paired tests, and to 37.82% on the full benchmark, without changing any model weights. The preprint from Einsia.AI and Tsinghua University suggests robot know-how can be stored as code and notes rather than trained into a network.