AI coding agents can now autonomously train physical robots to near-perfect performance on fiddly manipulation tasks — and they did it without a human babysitting the process.
Researchers introduced ENPIRE, a framework that gives coding agents a repeatable physical feedback loop: reset the scene, run the robot, check whether it succeeded, then rewrite the policy and try again. The system has four modules handling environment resets, policy refinement, parallel robot rollouts, and a final stage where agents read logs, consult research literature, and patch their own training code. Using this setup, frontier coding agents trained a manipulation policy to a 99% success rate on tasks including organizing a pin box, fastening a zip tie, and basic tool use — with no human intervention during training.
The bottleneck in robotics has never been the robots; it has been the researchers. Every failed grasp traditionally meant a human stepping in to diagnose, adjust, and re-run. ENPIRE replaces that loop with software, which means robot learning can run overnight, on weekends, and across a fleet of machines simultaneously — compressing iteration cycles that used to take weeks.
This is still a lab result, and "99% on a pin box" is a long way from a warehouse floor full of edge cases. But the direction is clear: the next wave of robotics progress may be gated less by mechanical engineering and more by how well AI agents can conduct their own experiments.