Nvidia's Autonomous Robotics Research

Implementing sophisticated robot manipulation skills requires massive human supervision and algorithm engineering, which is a key bottleneck in achieving general Physical Intelligence. Existing coding agents have mostly been limited to code generation in digital environments.

To address this, Nvidia has introduced a framework called ENPIRE. ENPIRE automates the physical feedback loop of 'environment reset - policy execution - outcome verification - iterative improvement.' The system consists of four core modules.

• Environment (EN): Performs automatic reset and outcome verification
• Policy Improvement (PI): Runs the policy improvement process
• Rollout (R): Evaluates policies using a single robot or multiple robots
• Evolution (E): A coding agent that performs log analysis, literature reference, and infrastructure and code improvements

Policies trained through ENPIRE achieved a 99% success rate (pass@8) on challenging manipulation tasks such as PushT, Pin Insertion, GPU Insertion, and Cut Ziptie. The system supports various learning methods including heuristic learning, Behavior Cloning, and Reinforcement Learning (RL), and can dramatically accelerate research speed by leveraging fleets of robots.