ReSkill uses a coding agent to fix only the failing parts of robot manipulation programs

Researchers led by Zheng Sun have presented ReSkill, a framework in which a coding agent improves robot manipulation programs by revising only the parts that fail, in a paper posted on arXiv.

The agent builds a task graph centred on objects from an instruction and a library of skills, then uses execution records for each node to find failures. It can change how stages are organised, handed over and recovered, or strengthen perception, motion, manipulation and checking skills that keep failing. Changes are kept only after full-task tests and paired comparisons, and the deployed robot runs the retained skills without the agent.

Across ten RoboDojo tasks, average success rose from 30.0% for the first versions to 86.2% for the final programs. On eight real-world tasks, ReSkill succeeded in 72 of 80 trials in rebuilt simulations and 62 of 80 on a real robot, or 77.5%. The median revision changed three functions, about 0.5% of a program.