AI coding agents can autonomously direct robot training

Imagine equipping AI coding agents with a laboratory brimming with robotic arms, computational resources, and a substantial token budget to instruct robots on various tasks. Remarkably, these agents can devise a training schedule that enables robots to adeptly cut zip ties and insert GPUs into motherboards' narrow sockets. This breakthrough in autonomous robot training stems from a novel agent harness framework known as ENPIRE. Developed by the innovative NVIDIA GEAR (Generalist Embodied Agent Research) lab, in collaboration with experts from Carnegie Mellon University and the University of California, Berkeley, this software wraps around AI models, enhancing their functionality with abilities such as memory, context, constraints, and feedback loops. Jim Fan, director of AI at NVIDIA, shared insights on LinkedIn, highlighting the lab's ability to self-improve continuously. He humorously noted, “A part of our NVIDIA GEAR lab now self-improves tirelessly overnight. We just read the reports in the morning.” Fan jokingly added that the goal of this AI-driven robot training is so efficient that, “We all take a holiday, and Jensen wouldn’t even notice,” referring to NVIDIA’s founder and CEO, Jensen Huang. The implications of this technology extend beyond NVIDIA; Fan announced that the entire project would be open-sourced, allowing anyone to create their own "self-running robot lab at home." The ENPIRE harness features four distinct modules that empower AI coding agents to conduct automatic resets and verifications on tasks, refine robotic behavior policies, evaluate these policies across multiple robots operating simultaneously, and troubleshoot failures by analyzing logs, studying research papers, and enhancing training infrastructure and algorithm code. In testing, the ENPIRE harness was evaluated with three AI coding agents, including 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 coding agents collaborated independently to develop various algorithmic strategies for robot training, conducting real-world tests and refining their approaches based on success rates over multiple cycles of self-directed examinations.