Event

Test results of coding agents acting as closed-loop robot policies

1 reports · 1 sources

A research team tested on the RoboDojo-Sim benchmark whether coding agents can directly act as closed-loop robot policies: the agent receives RGB observations and robot state, writes and executes Python control programs, and then revises the next program based on actual results, with no learned motion policy generating actions at any point. During testing, successful operation procedures and failure records were distilled into executable Skills and diagnostic Lessons, which were frozen and passed to new agents working on unseen layouts. Results show that the method performs relatively well on memory-dependent and open-ended tasks, but achieves a success rate of only 13.25% on precision tasks.