NVIDIA's ENPIRE Framework Achieves a 99% Success Rate in Robot Self-Learning

ME AI News: According to monitoring by Beating, the ENPIRE framework jointly released by NVIDIA, Carnegie Mellon University, and the University of California, Berkeley, has enabled robotic training to achieve fully autonomous evolution without any human intervention. Previously, fine-tuning robot movements required humans to constantly reset props and manually write and debug control code. Now, the team has directly connected large language model programming tools like Codex and Claude Code to a robot cluster, allowing these tools to autonomously generate motion control programs and evaluate success or failure through on-site cameras—analyzing error logs and modifying code just like human researchers. In a series of millimeter-precision tasks—including organizing scattered pins, inserting and removing motherboard components, tying zip ties, and cutting zip ties with a utility knife—the robots achieved a 99% success rate in testing with zero human intervention. The experiment demonstrated that this form of physical autonomous learning has strong scalability: when the robot count was expanded to eight, different branches of large model agents could automatically share and iteratively improve each other’s optimal algorithms via Git branches, reducing the training time for pin insertion from 1.5 hours to approximately 40 minutes. However, the autonomous evolution process has also revealed new bottlenecks. While a single device operates with 85% effective motion time, when all eight devices run simultaneously, robots frequently pause to wait for large model tools to read massive operational logs, rewrite code, and await API responses, causing hardware utilization to drop sharply to 35%. Meanwhile, frequent synchronization of optimal solutions among multiple robot agents has caused token consumption to rise sharply. The project team has announced plans to open-source the related code soon. (Source: BlockBeats)