AI Coding Agents Train Robots to Install GPUs and Cut Zip Ties
▼ Summary
– Nvidia’s ENPIRE agent harness enables AI coding agents to autonomously train robots to perform tasks like cutting zip ties and inserting GPUs into sockets.
– ENPIRE was developed by Nvidia GEAR lab researchers with collaborators from Carnegie Mellon University and UC Berkeley.
– The harness includes four modules for automatic task reset and verification, policy refinement, parallel robot evaluation, and failure analysis and code improvement.
– ENPIRE was tested with three AI coding agents: OpenAI’s Codex, Anthropic’s Claude Code, and Moonshot AI’s Kimi Code.
– The team plans to open-source ENPIRE so anyone can host their own “self-running robot lab at home.”
What happens when you give AI coding agents unrestricted access to a robotics lab, some hardware, and a generous token budget for training? They figure out how to teach robots to cut zip ties and insert GPUs into tight motherboard slots, all without human intervention. That’s the premise behind a new agentic harness framework called ENPIRE, developed by researchers at the Nvidia GEAR lab in collaboration with Carnegie Mellon University and UC Berkeley.
ENPIRE wraps around AI models, giving them tools, memory, context, constraints, and feedback loops. This lets the coding agents autonomously design and execute training regimens for physical robots. “A part of our NVIDIA GEAR lab now self-improves tirelessly overnight,” wrote Jim Fan, director of AI at NVIDIA, on LinkedIn. “We just read the reports in the morning.”
Fan jokingly added that the goal is to make the lab so self-sufficient that “we all take a holiday and Jensen wouldn’t even notice,” referencing NVIDIA CEO Jensen Huang. More seriously, the team plans to open-source everything, so anyone can host their own self-running robot lab at home.
The ENPIRE harness contains four modules. One handles automatic reset and verification for tasks. Another refines the policies that guide robotic behavior. A third evaluates those policies across multiple physical robots working in parallel. The fourth addresses failures by analyzing logs, ingesting research papers, and improving both training infrastructure and algorithm code. A detailed research paper was uploaded on June 16, 2026.
The harness was tested with three different AI 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. Each team of coding agents independently developed different algorithmic approaches to robot training. They tested those approaches in real-world experiments, then retained whatever changes helped raise the overall success rate over repeated cycles of self-directed testing.
(Source: Ars Technica)