Hands-Off Agentic Robot Learning in MuJoCo

A case study where an agent did not merely write robotics code, but ran the experimental loop: it built a MuJoCo OpenArm manipulation world, chose what to test, rejected misleading reward curves, launched training jobs, judged rollouts through a separate vision-language critic, localized bugs, and decided what to try next. The human supplied goals and behavioral constraints over chat; the agent made most of the tactical judgments. In roughly 36 hours and 28 logged iterations, the loop went from an empty simulator and repeated 0% reward-hack failures to chained teachers at 76–80%, a language-conditioned π0.5 policy recovered from a silent action-harness bug to 56/60 grasps and 49/60 correct-color grasps, and a final RABC checkpoint at 219/250 (87.6%) strict simulated success. The result is not solved robotics. It is evidence that hands-off agentic systems can begin to own the messy judgment loop of robot learning: build, watch, doubt, diagnose, retry, and remember.