Nvidia ENPIRE: AI agents take over robot training — a new standard for autonomous robotics
The robotics market is on the verge of a fundamental shift. A research group from Nvidia, together with colleagues from Carnegie Mellon University and the University of California, Berkeley, has introduced the ENPIRE framework. This is not just another programming tool—it is a full-fledged ecosystem in which AI agents autonomously manage the learning cycle of physical robots.
ENPIRE's key innovation is a closed loop: the robot performs a task, the environment automatically evaluates the result and returns to its initial state, while the AI agent analyzes errors, rewrites code, and launches the next series of trials. Human involvement is minimized, dramatically accelerating the learning process.
How ENPIRE Works
In traditional robotics, training on real equipment is an expensive and slow process requiring constant engineer involvement. Each failed attempt demands manual scene reset, result verification, and algorithm adjustment. ENPIRE brings the AutoResearch concept to the physical world, where AI agents write code, test it, and improve it in subsequent iterations. However, unlike in a digital environment, each experiment here faces real physical constraints: grasping errors, friction, and equipment imperfections.
The framework consists of four key modules:
- Environment: automatic scene reset, result verification, logging, and safety interfaces.
- Policy Improvement: launching policy improvement.
- Rollout: policy evaluation on one or more physical robots.
- Evolution: log analysis, literature search for ideas, modification of the training infrastructure, and code correction.
Automatic Verification and Reset: The Foundation of Autonomy
Automating two operations—result verification and scene reset to its initial state—is the cornerstone of ENPIRE. The system itself determines whether the task is completed using a combination of detectors, segmentation models, and multi-camera verification. This allows running multiple attempts in a row without manual labeling of each run. Automatic reset, in turn, eliminates the need for constant human involvement, which is critical for scaling.
Results on Real Robots
In experiments, ENPIRE was tested on several manipulation tasks: Push-T (pushing a T-shaped object to a target zone), Pin Insertion (inserting pins into 4 mm diameter holes), GPU installation, and cable tie operations. In real-world tasks, 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 and repeat actions with context awareness.
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. In the RoboCasa household task simulator, ENPIRE outperformed Nvidia's GR00T and CaP-X.
Scaling: Eight Robots—Less Time
A separate section of the work focuses on scaling to a robot fleet. An experiment with eight robotic stations showed that sharing results via Git reduces training time. Moving from one robot to eight cut 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 Prospects
The authors rightly note that scaling does not solve all problems. As the number of robots increases, GPU activity rises, but the average utilization of the robots themselves decreases due to time spent on coordination and result aggregation. Token consumption also increases. Additionally, ENPIRE has so far been demonstrated on a limited set of manipulation tasks.
My expert opinion: ENPIRE is not just a step forward—it is a paradigm shift. We are moving from manual programming to autonomous learning, where AI agents themselves explore the physical world. This opens the path to creating truly universal robots capable of adapting to new tasks without human intervention. However, like any breakthrough, ENPIRE raises new questions: how to manage computational costs and how to ensure safety when scaling to hundreds of robots.