Nvidia's ENPIRE harness lets AI coding agents train robots autonomously
Nvidia's GEAR lab built ENPIRE, an agentic harness that lets teams of AI coding agents autonomously direct robot training — achieving a 99% success rate on tasks like GPU insertion and zip-tie cutting.
Score breakdown
ENPIRE demonstrates that teams of AI coding agents can autonomously run and improve robot training overnight — outpacing a human-in-the-loop method developed by the same researchers on at least one task — and the planned open-source release extends that capability beyond Nvidia's own lab.
- 01ENPIRE is an agent harness framework developed by Nvidia GEAR lab with Carnegie Mellon University and UC Berkeley.
- 02The harness has four modules: automatic reset/verification, policy refinement, parallel multi-robot evaluation, and failure analysis.
- 03Three coding agents were tested: OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot AI's Kimi Code with Kimi K2.6.
Nvidia's GEAR (Generalist Embodied Agent Research) lab, alongside collaborators from Carnegie Mellon University and UC Berkeley, has developed ENPIRE — an agent harness framework that wraps around AI models to give them tools, memory, context, constraint, and feedback loops for autonomously directing robot training. Given a lab of robotic arms, compute resources, and a "generous token budget," teams of AI coding agents were able to independently develop algorithmic approaches, test them in real-world experiments, and iteratively retain changes that improved robot performance across repeated self-directed cycles.
ENPIRE was tested with three coding agents: OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot AI's Kimi Code with Kimi K2.6.
The harness comprises four modules covering automatic task reset and verification, policy refinement, parallel multi-robot policy evaluation, and failure analysis that includes ingesting research papers and improving training code. ENPIRE was tested with three coding agents: OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot AI's Kimi Code with Kimi K2.6. Across manipulation tasks — including the standard Push-T block-positioning task, pin box organization, zip-tie cutting, and GPU insertion into a motherboard — the agents achieved a 99% success rate. On the pin insertion and organization task, the AI agents reached near-100% success faster than a "frontier human-in-the-loop method" developed by many of the same human researchers. Scaling agent team size also showed clear gains: an eight-agent team completed the Push-T task at 99% success in two hours, compared to three hours for a four-agent team and nearly five hours for a single agent. Nvidia's Jim Fan announced plans to open-source the full framework so anyone can host their own self-running robot lab. The research paper was uploaded on June 16, 2026, though the source text is truncated before the article's discussion of the system's limitations.
Key facts
- 01ENPIRE is an agent harness framework developed by Nvidia GEAR lab with Carnegie Mellon University and UC Berkeley.
- 02The harness has four modules: automatic reset/verification, policy refinement, parallel multi-robot evaluation, and failure analysis.
- 03Three coding agents were tested: OpenAI's Codex with GPT-5.5, Anthropic's Claude Code with Opus 4.7, and Moonshot AI's Kimi Code with Kimi K2.6.
- 04Agents achieved a 99% success rate across manipulation tasks including Push-T, pin organization, zip-tie cutting, and GPU insertion.
- 05On pin insertion, AI agents reached near-100% success faster than a 'frontier human-in-the-loop method' from the same researchers.
- 06An eight-agent team completed the Push-T task at 99% success in 2 hours; a four-agent team took 3 hours; a single agent took nearly 5 hours.
- 07Nvidia's Jim Fan said the team plans to open-source everything so anyone can run their own self-directed robot lab.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 18, 2026 · 10:40 UTC. How this works →