Nvidia introduces ENPIRE: AI agents independently train robots in the real world — a breakthrough in robotics
Nvidia's research group, together with colleagues from Carnegie Mellon University and the University of California, Berkeley, has announced the ENPIRE framework — an innovative system that allows AI agents to autonomously improve robot control algorithms by working directly on physical hardware. This is a significant step forward in the field of robotics automation, where training on real machines has traditionally been an expensive and slow process.
ENPIRE implements 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, rewrites code, and launches the next series of trials. This approach, called Nvidia AutoResearch, transfers the methodology of digital simulations into the physical world, where each experiment must account for real-world constraints — friction, grasping errors, object physics.
ENPIRE Architecture and Modules
The framework consists of four key modules. The Environment module is responsible for automatic scene reset, result verification, logging, and safety interfaces. Policy Improvement launches the process of improving the control policy. Rollout evaluates the current policy on one or more physical robots. Finally, the Evolution module allows agents to analyze logs, search for ideas in scientific literature, modify the training infrastructure, and fix code. After the initial environment setup, the cycle can operate without constant human intervention, fundamentally changing the approach to robot training.
Automation of Verification and Reset — The Key to Autonomy
A critical innovation of ENPIRE is the automation of two operations: result verification and scene reset to its initial state. For example, in a cable tie task, the evaluation function combines a detector, a segmentation model, and a check using two cameras, allowing the agent to receive a success or error signal without manual labeling of each run. Automatic reset, in turn, enables running multiple attempts in succession, eliminating the need for constant engineer intervention.
Experimental Results and Performance
In real-world experiments, ENPIRE demonstrated impressive results. On manipulation tasks such as pushing a T-shaped object (Push-T) and inserting pins into holes with a diameter of 4 mm (Pin Insertion), the system successfully completed the task in 99% of cases when the agent was given up to eight attempts, accounting for previous errors. This reflects the system's ability to adapt and recover from failures, rather than the accuracy of a single attempt.
For programming agents, Codex on GPT-5.5, Claude Code on Opus 4.7, and Kimi Code on Kimi K2.6 were compared. Testing in the AutoEnvBench benchmark confirmed the effectiveness of the approach. In the RoboCasa household task simulator (opening cabinets, turning appliances on/off), ENPIRE outperformed Nvidia's GR00T and CaP-X — an agent system without a full automatic exploration cycle.
Scaling and Accelerating Learning
A separate experiment on eight robotic stations with two manipulators showed that scaling significantly accelerates learning. The stations exchanged results via Git, allowing successful ideas to be rapidly disseminated. The transition 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. This demonstrates the potential for industrial application.
Limitations and Conclusions
The authors emphasize that scaling is not without its challenges. As the number of robots increases, the time spent on reading logs, coordination, and waiting for the language model's response also increases, reducing the average utilization of the robots themselves. Token consumption also grows. Furthermore, ENPIRE has so far been demonstrated on a limited set of manipulation tasks, and its results do not imply that robots can autonomously master arbitrary physical skills in an open environment without prior engineering preparation.
My comment as an analyst: ENPIRE is not just another framework, but a potential turning point in robotics. The ability for autonomous, scalable learning on real equipment directly brings us closer to creating truly adaptable robots capable of adapting to changing conditions without constant human intervention. However, the key challenge will remain balancing computational costs with learning efficiency, especially when transitioning to more complex, multi-task scenarios. Keep an eye on the development of this technology — it could radically change the market for industrial and service robotics.