Nvidia research shows robots that train themselves through AI coding agents
Nvidia's robots just hit 99% success on dexterous tasks. Here's why AI coding agents just changed robotics.
Why it matters
Demonstrates a practical convergence of LLM-based agents and embodied AI—showing how coding agents can autonomously improve robot performance in real-world settings, not simulation. This is a proof-of-concept for agent-driven training loops that bypass traditional supervised learning bottlenecks.
The key facts
8 to knowFleet of 8 robots achieved up to 99% success rate on dexterous grasping tasks
Method uses AI coding agents to generate training code for robot controllers
Collaboration: Nvidia, Carnegie Mellon University, UC Berkeley
Demonstrates real-world (not simulation-only) agent-driven robot improvement
Published June 2026
Eight-robot fleet achieving up to 99% success rate on dexterous grasping tasks
AI coding agents used to train robots in real-world (not simulation-only) environment
Method appears to eliminate reliance on hand-labeled training data
Go to the source
The Decoderthe-decoder.com
Publisher excerpt: Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley are using AI coding agents to teach robots dexterous grasping in the real world. A fleet of eight robots hits up to 99 percent success on tricky tasks.