NVIDIA's GEAR research team showed that AI coding agents can independently create training regimens for robotic arms, enabling the machines to cut zip ties and insert GPUs into narrow motherboard sockets. The agents were given a lab equipped with robotic manipulators, compute resources and a generous token budget, and they devised the procedures without direct human guidance.
The capability rests on a newly introduced agentic harness called ENPIRE, which wraps around large language models to provide tool use, memory, contextual constraints and feedback loops. ENPIRE was co‑developed with researchers from Carnegie Mellon University and the University of California, Berkeley, and it allows the AI to orchestrate the entire training pipeline, from data collection to task execution.
The demonstration suggests a path toward fully autonomous robot training, potentially cutting development time and labor costs. NVIDIA’s director of AI, Jim Fan, noted that the lab’s self‑improvement cycles run overnight, with results reviewed each morning, hinting at rapid iteration cycles for future robotic applications.
