Imagine handing a coding agent a lab of robotic arms, compute credits, and a mandate to teach robots new tasks. According to NVIDIA’s GEAR lab and collaborators from CMU and UC Berkeley, that agent can autonomously design a training regimen—successfully teaching robots to cut zip ties and insert GPUs into motherboard slots. The key is a new agent harness framework called ENPIRE, which wraps around large language models to provide tool access, memory, constraint enforcement, and feedback loops. In practice, this means the AI can iterate overnight: it writes policy code, runs experiments across multiple physical robots, evaluates success rates, and self-corrects by analyzing logs or even ingesting research papers. Jim Fan, NVIDIA’s AI director, noted their lab now self-improves without human intervention, joking that the team could take a holiday and CEO Jensen Huang wouldn’t notice. The harness was tested with three coding agents—OpenAI’s Codex (GPT-5.5), Anthropic’s Claude Code (Opus 4.7), and Moonshot’s Kimi Code (K2.6)—each independently developing and refining algorithmic approaches across repeated cycles. From a system design perspective, ENPIRE’s four modules handle reset/verification, policy refinement, parallel evaluation, and failure recovery. The team plans to open-source everything, enabling anyone to host a self-running robot lab. For data and ML engineers, this shifts the bottleneck from manual experiment design to infrastructure orchestration: managing token budgets, parallel hardware, and feedback loops becomes the primary engineering challenge. The full technical paper was uploaded on June 16, 2026.

Source: Ars Technica