Study of 28,000 coding-agent runs finds robot self-learning hinges on reading failures

A team including Kaiyuan Chen, Spencer Huang and Ken Goldberg has published a controlled study of agentic robotics on arXiv, the approach in which coding agents write robot policies and keep improving them from simulated and real-world interactions. The study covered nine manipulation tasks, 631 experiment configurations and 28,000 coding-agent runs, varying model capability, skill priors, feedback structure and curriculum.

The data suggest the main obstacle is often not getting enough information but turning it into the right change to the policy. A working self-learning loop depends on how well agents find the cause of failures, reason about the spatial and geometric structure of the task and rewrite the right parts of the code, the authors write.