Crypto news

A research group combining specialists from Nvidia, Carnegie Mellon University, and the University of California, Berkeley, has introduced ENPIRE — an innovative framework that fundamentally changes the approach to robot training. This system allows AI agents specializing in programming to independently improve robot control algorithms by working directly with physical equipment.

The key feature of ENPIRE is the implementation of a closed-loop learning cycle. The robot performs a task, the environment automatically evaluates the result and returns to its initial state, and the AI agent analyzes errors, corrects the code, and launches the next iteration of tests. This eliminates the need for constant human intervention, which traditionally slowed down and increased the cost of the process.

How ENPIRE Works

In robotics, training on real equipment has always been expensive and slow. Each failed attempt requires manually resetting the scene to its initial state, checking the result, and changing the algorithm. ENPIRE transfers the concept of AutoResearch to the physical world, where AI agents write code, test it, and improve it. However, unlike digital simulation, each experiment here is subject to real physical constraints: friction, grasping errors, and sensor imperfections.

The framework consists of four modules:

• Environment — handles automatic scene reset, result verification, logging, and safety.
• Policy Improvement — launches the process of improving the control policy.
• Rollout — evaluates the current policy on one or more physical robots.
• Evolution — allows agents to analyze logs, search for ideas in literature, modify the training infrastructure, and fix code.

After the initial environment setup, the cycle can operate without constant human supervision. The agent receives data from video, trajectories, and reward functions, formulates a hypothesis, modifies the code, tests the result on the robot, and saves changes if they improve performance.

Automation of Verification and Reset

A key element of ENPIRE is the automation of two critically important operations: result verification and scene reset to its initial state. The first allows the system to independently determine whether the task has been completed. For example, in a cable tie scenario, the evaluation function combined a detector, a segmentation model, and verification from two cameras. This allowed the agent to receive a success or error signal without manual labeling of each run.

Automatic reset enables running multiple attempts in succession. After a failed action, the robot returns the object or scene to a state suitable for the next experiment. Without this, training on real equipment quickly hits the barrier of requiring constant human involvement.

Results on Real Robots

In experiments, the team tested ENPIRE on several manipulation tasks: Push-T (pushing a T-shaped object into a target zone), Pin Insertion (inserting pins into holes with a diameter of 4 mm), GPU installation, and cable tie operations. The system successfully completed the task in 99% of cases when the agent was given up to eight attempts, accounting for previous errors. This metric reflects the system's ability to recover from failures, not the accuracy of a single isolated attempt.

For programming agents, Codex on GPT-5.5, Claude Code on Opus 4.7, and Kimi Code on Kimi K2.6 were compared. The evaluation was conducted in the AutoEnvBench benchmark on Push-T and Pin Insertion tasks. In the RoboCasa household task simulator, ENPIRE outperformed Nvidia's GR00T and CaP-X — an agent system that does not run a full automatic research cycle.

Scaling to a Robot Fleet

A separate section of the work is dedicated to scaling. Nvidia conducted an experiment on eight robotic stations with two manipulators each. The stations shared results via Git, which reduced training time. Transitioning from one robot to eight reduced the learning time for Push-T from approximately five hours to about two hours, and for Pin Insertion from over 90 minutes to around 40 minutes.

Limitations and Conclusions

Scaling does not solve all problems. When agents read logs, write code, or wait for a language model response, robots and computing resources are not fully utilized. As the number of robots increases, GPU activity rises, but the average utilization of the robots themselves decreases. Agent teams spend more time on coordination rather than physical runs. Token consumption also increases.

ENPIRE has so far been demonstrated on a limited set of manipulation tasks. Its results do not mean that robots can already autonomously master arbitrary physical skills in an open environment without engineering preparation. Nevertheless, this is a significant step forward in the automation of robotics.

My comment: ENPIRE is not just another tool, but a paradigm shift towards truly autonomous robotics. Automating the "error-analysis-correction" cycle on physical equipment is what separates modern laboratory prototypes from industrial solutions. If Nvidia manages to scale this approach to a wider range of tasks, we will witness a sharp acceleration in the deployment of robots in real-world production processes.