Nvidia Research, Carnegie Mellon University, and UC Berkeley researchers introduced ENPIRE, a real-world robot learning framework that lets AI coding agents iteratively improve robot manipulation behavior. The team reports up to 99% success on tasks such as GPU insertion, pin insertion, and zip-tie cutting using an...
Nvidia Research, working with researchers from Carnegie Mellon University and UC Berkeley, has introduced ENPIRE, a framework that uses AI coding agents to improve robot behavior on real hardware.
According to Nvidia Research’s ENPIRE project page, the system is designed for “agentic robot policy self-improvement in the real world.” In practice, that means coding agents can propose changes to robot control code, test those changes through repeated physical trials, and keep modifications that improve performance.
The work focuses on dexterous manipulation, a long-standing challenge in robotics because small physical errors can cause a task to fail. Nvidia Research says ENPIRE was tested on difficult real-world tasks including GPU insertion, pin insertion, and zip-tie cutting.
Nvidia Research describes ENPIRE as a harness framework that combines automatic reset, verification, rollouts, and evolution. The system allows AI coding agents to run experiments on physical robots, evaluate results, and revise robot behavior without requiring a human to hand-tune every attempt.
The Decoder reports that the project uses AI coding agents to teach robots dexterous grasping in the real world. Decrypt similarly reports that coding agents including Codex, Claude Code, and Kimi Code were given control of an eight-robot fleet to train and test robot policies on real hardware.
That real-world emphasis is important. Many robotics systems are first trained in simulation, where data can be generated quickly and cheaply. But physical robots face friction, imperfect sensing, shifting objects, and hardware variation. ENPIRE’s reported contribution is to let software-writing AI systems iterate directly against those real-world conditions.
Nvidia Research reports that ENPIRE achieved up to 99% success on dexterous manipulation tasks. The project page names GPU insertion and zip-tie cutting among the examples, while Decrypt adds pin insertion to the set of reported tasks.
Tom’s Hardware reports that Nvidia showed robots learning to insert a graphics card into a motherboard, describing the demonstration as a high-precision task. The publication also connects the demonstration to the ENPIRE research paper and its comparisons between coding agents.
The reported results do not mean general-purpose household or factory robots are solved. The sources describe a controlled research setup with a fleet of eight robots and specific manipulation tasks. Still, the results suggest that AI coding agents may be useful not only for writing software, but also for improving robotic behavior through repeated physical experimentation.
Robotics research often depends on expert engineers who design control methods, collect data, inspect failures, and revise system behavior. ENPIRE points to a different workflow: AI coding agents generate and modify the robot-control logic, while automated testing on physical robots provides feedback.
If the approach proves reliable beyond the reported tasks, it could reduce some of the manual effort required to adapt robots to precise manipulation jobs. The examples cited by Nvidia Research and covered by The Decoder, Decrypt, and Tom’s Hardware are all tasks where millimeter-scale errors matter: aligning pins, cutting zip ties, and seating a GPU into a motherboard.
For now, ENPIRE is best understood as a research result rather than a commercial robot product. The key finding from Nvidia Research, CMU, and UC Berkeley is that coding agents can be integrated with real robot testing loops and can improve performance on specific dexterous manipulation tasks, with reported success rates reaching as high as 99% in the team’s experiments.
The work focuses on dexterous manipulation, a long standing challenge in robotics because small physical errors can cause a task to fail.
Nvidia Research says ENPIRE was tested on difficult real world tasks including GPU insertion, pin insertion, and zip tie cutting.
How ENPIRE Works Nvidia Research describes ENPIRE as a harness framework that combines automatic reset, verification, rollouts, and evolution.
Continue reading
