Nvidia's ENPIRE Lets Robots Teach Themselves New Tricks
Nvidia researchers built a framework that gives AI coding agents control of robot arms to develop self-improvement strategies.
Nvidia just dropped ENPIRE, a framework that essentially hands AI coding agents the keys to a robotics lab and says "figure it out."
The system is an "agent harness framework" — it takes AI coding agents, gives them access to robotic arms, compute resources, and what the researchers describe as a "generous token budget," then lets them develop strategies for teaching robots physical tasks. The kicker: minimal human supervision required.
ENPIRE doesn't just run pre-programmed routines. It autonomously generates self-improvement strategies, iterating on how robots learn to handle physical tasks. Think of it as a meta-learning layer — AI figuring out the best way to train other AI that controls physical hardware.
The approach could dramatically reduce the human effort needed to get robots competent at new tasks, a major bottleneck in real-world robotics deployment.