Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley have developed a system that allows robots to train themselves using AI coding agents. The project, called ENPIRE, enables robots to perform dexterous grasping and manipulation in the real world with minimal human intervention. A fleet of eight robots achieved up to 99% success on challenging tasks, including pin insertion and cable tie closing, demonstrating the potential of autonomous robot learning.
The system operates in two phases. In the first phase, the AI coding agent sets up a working environment with some human feedback, including safety boundaries, automatic resets, and automated success checks. The agent builds its own evaluation tools, such as a reward function for pin insertion that combines visual alignment, gripper height, and estimated force. For cable tie closing, the agent used two camera angles to avoid false positives and reduced reaction time below 150 milliseconds. These tools are built once and reused without modification.
The study highlights the challenges of real-world environments, noting that while agents solved tasks in simulation, two out of three failed in the real world due to unpredictable conditions like friction and object movement. The researchers also introduced two metrics to measure efficiency: Mean Robot Utilization (MRU) and Mean Token Utilization (MTU). They tested three coding agents—Codex with GPT-5.5, Claude Code with Opus 4.7, and Kimi Code with Kimi K2.6—with Codex performing best in most cases.
Source: thedecoder
