Crypto news

A research group combining experts from Nvidia, Carnegie Mellon University, and the University of California, Berkeley, has introduced the ENPIRE framework. This is not just another robotics library—it is a full-fledged ecosystem in which AI agents specializing in code writing take over the process of improving control policies for real robots. The system operates in a closed loop: the robot performs a task, the environment automatically evaluates the result and returns to its initial state, and the agent analyzes errors, rewrites code, and launches a new series of trials.

How ENPIRE is structured: four automation modules

In traditional robotics, training on real hardware is expensive and slow. Each failed attempt requires manual scene reset, result verification, and algorithm adjustment. ENPIRE brings the AutoResearch concept, already tested by Nvidia in digital simulations, into the physical world. The framework consists of four key modules:

• Environment — handles automatic scene reset, result verification, logging, and safety interfaces.
• Policy Improvement — launches an iterative process to improve 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, change training infrastructure, and fix code.

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

Automatic verification and reset: the key to scaling

The most important element of ENPIRE is the automation of two operations: result verification and scene reset to its initial state. Without this, training on real hardware quickly runs into the need for constant human involvement. For example, in a cable tie scenario, the evaluation function combined a detector, a segmentation model, and a two-camera check, allowing the agent to receive a success or error signal without manual labeling of each run. Automatic reset, in turn, enables many consecutive attempts, which is critical for effective learning.

Experiments on real robots: 99% success rate

In real 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 4mm diameter holes), GPU installation, and cable tie operations. The project page states that 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. Evaluation was conducted in the AutoEnvBench benchmark on Push-T and Pin Insertion tasks. Additionally, ENPIRE was tested in RoboCasa, a simulator of household tasks, where it outperformed Nvidia's GR00T and CaP-X.

Eight robots: scaling accelerates learning

A separate section of the work focuses on scaling to a robot fleet. Nvidia conducted an experiment on eight stations, each with two manipulators. The stations shared results via Git: a successful idea or code change could quickly spread among agents. Moving from one robot to eight reduced the time to master Push-T from approximately five hours to about two hours, and for Pin Insertion from over 90 minutes to around 40 minutes.

Limitations and a look to the future

The authors emphasize that scaling does not solve all problems. When agents read logs, write code, or wait for a response from the base language model, 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. Token consumption also grows. Importantly, ENPIRE has so far been demonstrated on a limited set of manipulation tasks. This does not mean that robots can already independently acquire arbitrary physical skills in an open environment without engineering preparation.

Expert commentary: ENPIRE is an important step toward autonomous robotics, but it shows that the "bottleneck" is becoming less about physics and more about computing resources and AI agent coordination. We are on the threshold of an era where robots can learn from each other, but the price of this learning is an exponential increase in token consumption and GPU power. Investors should take a closer look at companies offering solutions to optimize these processes.