Researchers from Nvidia, in collaboration with the University of Carnegie Mellon and the University of California at Berkeley, have introduced the ENPIRE framework, which allows AI agents to automatically improve their robot management policies on real-world equipment. The system runs a closed loop: the robot performs the task, the environment automatically evaluates the result and returns it to its original state, and the AI agent analyzes errors, rewrites the code, and launches the next test series. Key elements - automatic validation of the result and reset of the scene, which allows the system to work without the constant involvement of engineers.

In real-world experiments, ENPIRE achieved 99% success on manipulation tasks when an agent was given up to 8 attempts. Scaling up to a fleet of 8 robots with sharing of results via Git reduced training time from 5 hours to 2 hours for the Push-T task and from 90 minutes to 40 minutes for Pin Insertion. However, the authors point out limitations: the increase in the number of robots increases the consumption of tokens and reduces the average equipment load, and the system has so far been tested only on a limited set of manipulation tasks.