Nvidia introduces ENPIRE: an autonomous AI framework for training robots on real hardware
A research group combining specialists from Nvidia, Carnegie Mellon University, and the University of California, Berkeley, has introduced the innovative ENPIRE framework. This system marks a new stage in robotics, allowing AI agents for programming to autonomously improve robot control policies by working directly on physical hardware.
The ENPIRE concept is built on a closed loop: the robot performs a task, the environment automatically evaluates the result and returns to its initial state, and the AI agent analyzes errors, rewrites code, and launches the next series of trials. This approach fundamentally changes the traditional process, where each failed attempt required manual intervention from an engineer to reset the scene and adjust the algorithm.
ENPIRE Architecture and Key Modules
The framework consists of four key modules: Environment handles automatic scene reset, result verification, and safety; Policy Improvement focuses on enhancing the control policy; Rollout evaluates policy effectiveness on one or more physical robots; Evolution enables agents to analyze logs, search for ideas in literature, and fix code. After the initial environment setup, the cycle can function with virtually no human involvement.
Automation and Test Results
The key innovation is the automation of two critically important operations: result verification and scene reset to the initial state. The system uses a combination of detectors, segmentation models, and cameras to autonomously determine task success, eliminating the need for manual labeling of each run. In real-world experiments, including tasks from pushing T-shaped objects to precisely inserting pins into 4mm diameter holes, the system demonstrated impressive reliability — 99% task success rate when the agent was provided with up to eight attempts considering previous errors.
A comparison of various AI agents, including Codex on GPT-5.5, Claude Code on Opus 4.7, and Kimi Code on Kimi K2.6, showed that ENPIRE surpasses existing solutions such as Nvidia's GR00T and CaP-X in the RoboCasa household task simulator. Of particular interest is the scaling experiment on a fleet of eight robotic stations. Using Git for sharing results between agents reduced the training time for the Push-T task from five to two hours, and for Pin Insertion from over 90 to 40 minutes.
Limitations and Prospects
Despite the breakthrough, the technology has limitations. Scaling increases the load on GPUs and token consumption, as agents spend time reading logs and coordinating. Furthermore, ENPIRE has so far been demonstrated on a limited set of manipulation tasks, and its results do not guarantee that robots can autonomously master arbitrary physical skills in an open environment without prior engineering preparation.
My expert conclusion: ENPIRE is not just another framework, but a fundamental shift in the paradigm of robot learning. Automating the "error-analysis-correction" cycle on real hardware brings us closer to an era where robots can autonomously improve their skills, much like humans learn from their own experience. However, the path to fully autonomous physical agents capable of operating in unstructured environments still requires solving problems of scaling and resource intensity.