AI coding agents taught robots how to install GPUs and cut zip-ties (arstechnica.com)

A groundbreaking development in autonomous robot training has emerged from researchers at the NVIDIA GEAR lab, Carnegie Mellon University, and UC Berkeley with the introduction of the ENPIRE framework. This software harness allows AI coding agents to independently design training regimens for robotic arms, enabling them to perform complex tasks such as cutting zip ties and inserting GPUs into motherboards without human intervention. The framework's innovative modules enhance the operational capabilities of these AI agents by incorporating memory, context, feedback loops, and a structured approach to evaluating and refining robotic behaviors. Significantly, this advancement marks a step towards fully automated robot training environments where AI can learn and improve without constant oversight. By leveraging agents like OpenAI’s Codex and Anthropic’s Claude Code, the ENPIRE framework demonstrates the potential for diverse algorithmic approaches to be tested and optimized in real time, fostering a cycle of continuous learning. The open-sourcing of this technology further extends its impact, allowing any interested party to establish their own self-running robot labs, thereby democratizing access to advanced AI-driven robotic training. This approach could revolutionize the efficiency and scalability of robotic applications in various fields.