Overview
- Researchers from Nvidia GEAR with Carnegie Mellon and UC Berkeley built ENPIRE, a four-module harness that gives coding agents automated reset, policy refinement, multi-robot evaluation, and failure-analysis capabilities.
- The team demonstrated bimanual robot stations performing high-precision tasks such as inserting GPUs into motherboards, cutting zip ties, and sorting fine pins under agent-directed training.
- The paper uploaded on June 16, 2026, reports experiments using three coding-agent stacks (Codex/GPT-5.5, Claude Code/Opus 4.7, Kimi Code/Kimi K2.6) that produced different algorithmic solutions and kept changes that improved success rates.
- Scaling tests showed that running multiple robots in parallel sped learning substantially, with eight robots solving tasks far faster than single or smaller fleets while sharing winning code changes across stations.
- The authors say they will open-source the code and materials so other labs can reproduce the setup, and they stress ENPIRE remains a research demonstration rather than a commercial product.