Nvidia's ENPIRE uses AI agents to code robots to 99% dexterity
nvidia researchers used coding agents like codex to train robots in dexterous tasks, achieving 99% success on complex manipulations.
researchers from nvidia, cmu, and uc berkeley developed enpire, allowing coding agents to autonomously train robots on dexterous manipulation tasks by writing their own evaluation code and sharing findings via git. testing three models—codex, claude code, and kimi code—against eight dual-arm robots, the system hit 99% success on the push-t test and beat simulations by 100% in speed on pin insertion. the system struggles more with real-world physics than simulation, but significantly accelerates mastery of tasks like cable cutting.