Nvidia showcases robots installing GPUs autonomously

NVIDIA GEAR Lab, jointly with Carnegie Mellon University and UC Berkeley, published ENPIRE (ENvironment, Policy Improvement, Rollout, Evolution), a harness framework that enables frontier coding agents to autonomously run the full loop of real-world robot learning: reset the scene, execute a policy, verify the outcome, refine code, and repeat, per the NVIDIA research page. Demonstrations include a robot arm performing high-precision GPU insertion into a motherboard, organizing pins into a pin box, tying zip-ties, and a Push-T shape-matching task. Linxi 'Jim' Fan (NVIDIA Director of AI & Distinguished Scientist, GEAR Lab co-lead) described the system as enabling "AutoResearch in the physical world for the first time." Trade press and the project page indicate an open-source release is planned.

ENPIRE's four modules work as a closed loop: the Environment (EN) module handles automatic scene reset and outcome verification; Policy Improvement (PI) generates and revises policy code from rewards, videos, and failure traces; Rollout (R) runs budgeted robot trials; and Evolution (E) has coding agents analyze logs, consult literature, and improve training infrastructure and algorithm code. The NVIDIA research page reports that agent teams of up to 8 robots, guided by frontier coding agents including Codex (GPT-5.5), Claude Code (Opus 4.7), and Kimi Code (Kimi K2.6), achieved a 99% pass@8 success rate on showcased dexterous manipulation tasks. Scaling from 1 to 8 agents improved research progress speed but increased total token consumption; the team introduces Mean Robot Utilization (MRU) and Mean Token Utilization (MTU) metrics to track efficiency. The project also includes simulation evaluation in the RoboCasa benchmark.

Open-source frameworks that combine automatic scene-reset, self-verification, and repeatable hardware rollouts materially lower the engineering overhead for robotics policy research. If the planned open-source release reproduces the reported results broadly, it could accelerate iteration on manipulation policies and physical RL reproducibility for academic and small-lab teams that previously lacked the staffing to operate continuous robot experimentation.

Verify the released codebase and documentation for reproducible benchmarks, instrumentation for safety checks during automated trials, and constraints on the robot hardware required. Observers should examine measured task success rates under diverse initial conditions, sample-efficiency metrics, and details on how the system verifies correct assembly steps. The planned open-source release timing has not yet been publicly confirmed.