NVIDIA's ENPIRE Lets Robots Teach Themselves in the Real World

NVIDIA researchers have developed ENPIRE, a framework that brings the self-improving experiment loops of AI coding agents into the physical world of robotics. The software supervises real robots as they attempt tasks, evaluate their own performance, reset their environment, and iterate — all without human input.

The system works through four modules: one that automatically resets the physical setup and verifies outcomes, another that refines the robot's policy, a rollout module that evaluates performance across single or multiple robots running in parallel, and an evolution module where coding agents analyze failures, consult research literature, and improve the underlying code. Each workstation runs on an NVIDIA RTX 5090 and controls a pair of bimanual robot arms equipped with cameras.

On simpler tasks, it works impressively well. Frontier coding agents achieved 99% success rates on dexterous manipulation challenges like pushing objects, organizing pins, and cutting zip ties — even inserting GPUs into a motherboard. Larger teams of agents explored more of the solution space and sometimes outperformed single-agent setups, with GPT-5.5 and Opus 4.7 trading the top spot depending on the task.

But scaling has friction. When coding agents are busy reading logs, debugging, or waiting on language model responses, robot hardware sits idle. As the number of robots grows, utilization rates drop even as GPU demand climbs — an infrastructure bottleneck that still needs solving before fleets of self-improving robots become practical.