NVIDIA GEAR Lab, led by Jim Fan, has unveiled ENPIRE, a groundbreaking framework that enables coding agents to autonomously conduct robotics research on real hardware. The system deploys coding agents onto a robot fleet, assigning GPU compute and token budgets to solve tasks efficiently while keeping robots busy and safe.
ENPIRE consists of four core modules: Environment Module for automatic scene reset and verification, Policy Improvement Module for launching policy refinement, Rollout Module for evaluating policies across parallel robots, and Evolution Module where coding agents analyze logs and rewrite code.
In dexterous manipulation tasks like PushT, pin insertion, zip-tie cutting, and GPU insertion, ENPIRE-powered agents achieved 99% success rates. Agents solved PushT in under 2 hours using heuristic methods without neural network training. The team observed a physical scaling law: increasing parallel robots from 1 to 8 reduced pin insertion time from 1.5 hours to 40 minutes.
Researchers proposed MRU and MTU metrics measuring robot and token utilization efficiency. Jim Fan stated ENPIRE will be fully open-sourced. Collaboration between NVIDIA GEAR Lab, CMU, and UC Berkeley.
