Nvidia Built Robots That Train Themselves Using AI Coding Agents
Nvidia’s ENPIRE framework lets AI coding agents train real robots end to end, not just write code in simulation. Human input is limited to setting up two reusable tools once: a reset routine and a camera-based reward function. After that, agents search papers, choose methods such as imitation learning or reinforcement learning, rewrite code, and test directly on hardware. In experiments on eight bimanual robot stations, the agents learned tasks like Push-T, pin insertion, graphics-card seating, and zip-tie cutting. Sharing results through Git let improvements spread across the fleet quickly. Scaling from one robot to eight cut training time for Push-T from about five hours to two, and pin insertion from over 90 minutes to about 40. Across four real-world tasks, success reached 99%. The work also exposed the sim-to-real gap: tasks solved in simulation sometimes failed on physical robots because of real-world friction and other hardware issues.
