Nvidia, in partnership with Carnegie Mellon and UC Berkeley, has launched ENPIRE, enabling AI agents to autonomously teach robots new skills with a 99% success rate.
Summary
- Nvidia, in collaboration with Carnegie Mellon University and UC Berkeley, has launched ENPIRE, a framework that allows AI coding agents to autonomously teach robots new skills without human input.
- An eight-robot team achieved a remarkable 99% success rate in tasks like pin insertion and GPU installation.
- While expanding from one to eight robots cut task mastery time by over 50%, the associated token costs increased at a faster pace.
Nvidia's GEAR lab has successfully demonstrated a group of eight robotic arms that can train themselves to perform various tasks, including pin insertion and GPU seating, with minimal human oversight required for post-experiment evaluations.
This self-training capability is made possible by the ENPIRE framework, detailed in a recent paper co-authored by researchers from Nvidia, Carnegie Mellon, and UC Berkeley. ENPIRE allows AI coding agents to manage the entire training process, enabling them to write and test their own code directly on physical robots.
Utilizing coding agents such as OpenAI's Codex, Anthropic's Claude Code, and Moonshot's Kimi Code, the researchers engaged in a process termed autoresearch, where agents autonomously write, test, and refine code without human intervention. This represents a significant shift from prior experiments conducted on screens, where failures had no real-world implications.
Developing ENPIRE
The ENPIRE system operates in two main phases. Initially, a human assists the agent in creating key tools: a reset routine for workspace restoration and a reward function that evaluates performance through camera footage. This setup is established once and reused in future attempts.
Once these tools are in place, the agent gains complete control to explore existing research, choose training methods such as reinforcement learning, and autonomously rewrite and test code on the robot. This self-management raises questions about the implications of robots operating independently.
In their experiments, Nvidia utilized eight bimanual robot stations, each equipped with hardware, computing resources, and coding agents. The stations collaborated using Git, a version control system, to share successful strategies rapidly.
The researchers focused on two tasks: the “Push-T” task, where a robot pushes a T-shaped block into a designated area, and pin insertion, which involves threading pins into small holes. The transition from one to eight robots significantly reduced the time needed to master these tasks, with the Push-T task time dropping from five hours to two, and pin insertion time reducing from over 90 minutes to about 40 minutes.
The coding agents achieved a 99% success rate across four practical tasks, demonstrating near-perfect reliability in pin insertion and outperforming traditional human-assisted methods that still require daily supervision.
Jim Fan, co-lead of the GEAR Lab and Nvidia's AI research head, highlighted the groundbreaking nature of this project, stating, "Today, we enable AutoResearch in the physical world for the first time!" He described the process as giving the agents a fleet of robots and GPU resources, allowing them to operate autonomously.
Today, we enable AutoResearch in the physical world for the first time! Introducing ENPIRE: we give 8 Codex agents a fleet of robots, an allocation of GPUs, and generous token budget. We set them free with a simple goal: solve the task as quickly as possible, keep the robots busy… pic.twitter.com/zC0OQNzDBs
— Jim Fan (@DrJimFan) June 16, 2026
Discrepancies between simulated and real-world results quickly emerged, as two coding agents that succeeded in the simulation failed during physical trials due to unaccounted friction issues.
Nvidia also tested ENPIRE in RoboCasa, a simulated kitchen environment, where it outperformed existing models in tasks such as opening cabinets and turning off stoves, demonstrating the potential of its self-training capabilities.
Building on concepts introduced in its earlier Eureka system, which used a language model for simulation-based reward functions, ENPIRE advances the self-improvement process to real hardware, allowing agents to design their own experiments.
This announcement aligns with Alibaba's launch of the Qwen-Robot Suite, highlighting the growing competition in the robotics sector, with both companies exploring how AI agents can autonomously manage research processes in their respective hardware environments.
