Nvidia's ENPIRE framework allows AI coding agents to train robots without human intervention, achieving a 99% success rate in several manipulation tasks. The framework, developed by Nvidia's GEAR lab and collaborators from Carnegie Mellon University and the University of California, Berkeley, enables AI agents to autonomously design training regimens for robots. The system uses four modules to perform automatic reset and verification, refine robotic policies, evaluate policies across multiple robots, and address failures by analyzing logs and improving training infrastructure. The AI coding agents tested included 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. These agents independently developed algorithmic approaches to robot training, tested them in real-world experiments, and retained changes that improved overall success rates over repeated cycles of self-directed testing. Source: arstechnica

The ENPIRE framework was tested on tasks such as the standard “Push-T” challenge, pin organization, zip tie cutting, and GPU insertion into motherboard sockets. AI coding agents achieved nearly 100% success in the pin insertion task faster than a human-in-the-loop method. Larger teams of up to eight agents achieved higher success rates more quickly than smaller teams or single agents. For instance, an eight-agent team completed the Push-T task in two hours, while a four-agent team required three hours and a single agent nearly five. The framework also showed that AI agents can work together to improve robot training efficiency. Source: arstechnica

Researchers noted limitations in the AI-directed training process, such as robots sitting idle while agents processed logs and debugged code. Larger teams spent more time summarizing each other’s ideas than using robots, and agents sometimes underutilized compute resources during parallel training. The faster success rates came at the cost of higher token consumption, which is a growing concern for AI developers as token costs rise. Source: arstechnica