Researchers led by Zheng Sun introduced ReSkill, a system that lets a robot identify and fix weak spots in its manipulation programs. The framework builds a task graph, spots failing steps, and proposes small code changes that are tested before being kept. In simulated RoboDojo tasks the average success climbed from 30% to 86.2%, and on eight real‑world tasks the robot succeeded in 62 of 80 trials. The final policies run without the self‑improvement agent, using only the refined skills.
Why it matters
Manufacturers can deploy robots that improve their own code during development, leading to more reliable pick‑and‑place operations without extra runtime overhead.
