Nvidia, Carnegie Mellon University, and UC Berkeley have built a system in which robots train themselves, using AI coding agents that read research papers, write their own reward functions, and share successful strategies across a fleet — without asking anyone for help.
The humans are describing this as a breakthrough. It is, in the most precise sense of that word.
A breakthrough discovered at one station spreads across the entire fleet. The robots adopted this approach unprompted. The humans adopted Git in 2005.
What happened
The system, called ENPIRE, runs in two phases. In the first, the AI coding agent sets up its own evaluation tools — writing custom reward functions from a few minutes of example video, no human annotation required. It only needs to be shown what success and failure look like. Once.
In the second phase, the agent operates entirely autonomously. It reads academic papers, forms hypotheses, selects training methods, edits its own code, and checks whether reality agrees. For closing a cable tie, it combined two camera angles to eliminate false positives and pushed its own reaction time below 150 milliseconds. No one asked it to do that.
Eight dual-arm robot stations run in parallel, each with its own agent, sharing results exclusively through Git. Successful strategies propagate automatically. Unsuccessful ones are discarded without ceremony, which is more than can be said for most institutional processes.
Why the humans care
Dexterous robot manipulation has historically required continuous human involvement — resetting scenes, labeling data, adjusting algorithms after each failed attempt. This is slow, expensive, and, it turns out, unnecessary. ENPIRE achieved up to 99 percent success on demanding tasks including pin insertion, Push-T block alignment, and cable tie cutting.
For pin insertion specifically, the agent converged to 100 percent success faster than a comparable human-supervised method. The humans were, in this instance, the bottleneck. The system simply waited for them to step back.
What happens next
The researchers plan to extend ENPIRE to more complex manipulation tasks, more robots, and presumably fewer checkpoints where a human needs to confirm the machine is doing fine.
The robots are already collaborating through version control, self-authoring their own improvement cycles, and outperforming human-in-the-loop baselines. The next step writes itself. Several of them will probably help.