Nvidia's robot lab now runs training sessions overnight with no human supervision, just AI agents and a budget of compute.
Researchers at Nvidia's GEAR lab, alongside collaborators from Carnegie Mellon University and UC Berkeley, built an agent harness framework called ENPIRE. The system wraps AI coding agents with memory, feedback loops, and tool access, then turns them loose on a lab full of robotic arms. Given a task and a "generous token budget," the agents designed and executed training regimens from scratch. The result: robots that can cut zip ties and insert GPUs into tight motherboard slots, the kind of precise physical task that typically requires human technicians.
That second benchmark matters more than it sounds. GPU insertion demands the fine motor alignment that humans still do largely by feel. If AI agents can write the training curriculum for that, the pace at which robots acquire new physical skills could decouple from the pace at which researchers can babysit experiments. Nvidia's director of AI, Jim Fan, put it plainly on LinkedIn: "We just read the reports in the morning."
The framework's real test will be whether skills developed inside a controlled Nvidia lab transfer anywhere else, and whether "reading the reports" constitutes genuine oversight or just checking whether anything caught fire.