Home / Technology / AI Tutors Robots: Overnight Self-Improvement Lab
AI Tutors Robots: Overnight Self-Improvement Lab
18 Jun
Summary
- AI agents autonomously train robots for tasks like cutting zip ties.
- New ENPIRE harness enables self-improving AI-directed robot training.
- AI teams achieved nearly 100 percent success faster than humans.
Researchers have developed the ENPIRE harness, a framework that allows AI coding agents to autonomously train robots. This system, developed by NVIDIA GEAR lab in collaboration with Carnegie Mellon University and UC Berkeley, enables AI agents to manage training regimens overnight.
The ENPIRE harness includes modules for automatic task reset, policy refinement, parallel evaluation, and failure analysis. It was tested with agents like OpenAI's Codex and Anthropic's Claude Code.
AI agents using ENPIRE achieved a 99 percent success rate on tasks including zip tie cutting and GPU insertion, notably completing a pin organization task faster than humans.
Larger teams of AI agents, such as an eight-agent team, demonstrated accelerated success rates on tasks like Push-T compared to smaller teams.
However, limitations were observed, including idle robot time and inefficient compute resource utilization by the AI agents. Increased token consumption also emerged as a concern with larger AI teams.
NVIDIA continues its push into physical AI, recently announcing a partnership with Unitree for research humanoid robots and discussing AI robot manufacturing with Hyundai Motor Group.
