ENPIRE: NVIDIA GEAR Lab's AutoResearch System with 8 Codex Agents on Robot Fleet

ENPIRE gives 8 Codex agents a fleet of robots, GPUs, and token budget to autonomously solve physical tasks. The agents learn visual clues, reset scenes, practice skills, read papers, and debate. They mastered high-precision tasks like tying zip-ties, organizing fine pins, and installing GPUs. A new "physical scaling" effect emerged: 8 robots exploring in parallel improve significantly faster than fewer robots. The system runs at NVIDIA GEAR lab, self-improving overnight, and the team plans to open-source everything.