UC Berkeley and Nvidia say teams of AI agents can now write, test and fix robot programs with little new training data
A Berkeley warehouse-robot company says the method lets each of its sorting robots handle more than 15,000 extra packages a year.
Researchers at UC Berkeley and Nvidia, the Santa Clara chipmaker, say teams of AI agents can now write robot-control programs, test them in a computer simulation and fix them when they fail, according to a UC Berkeley release published Nov. 9. They call the approach “agentic robotics.” An AI agent is an AI program that can plan steps and use other software tools on its own to finish a multistep job.
The problem it tries to solve
Robots are usually built one of two ways, the release says. Engineers can hand-write every instruction, which takes a lot of work. Or a robot can learn by copying many human demonstrations, which takes huge amounts of data. Last year, Berkeley roboticist Ken Goldberg warned of a “100,000-year data gap” between what robots need and what exists, because there is no robot version of the internet text that trained chatbots like ChatGPT.
How it works
Goldberg’s team breaks a task into small skills, such as finding an object, planning a grip and moving without hitting anything. AI agents then snap those skills together into a program and rehearse it thousands of ways in Nvidia’s Isaac simulation software, a kind of video-game world for robots. When the real robot fails, the agents use video of the failure to improve the simulation and try again. Because the program is made of named skills, engineers can read and check it before it runs.
In one test, the team gave OpenAI’s Astra model a short video of a robot wiping ink off a metal bar with a sponge. In under an hour, it built a realistic simulation of the scene, and the robot program it then wrote worked on a real robot.
Who is using it
Ambi Robotics, a Berkeley company that makes warehouse package-sorting robots, said the method raised each robot’s capacity by more than 15,000 packages a year and is now used across 30% of its U.S. fleet. Goldberg said not to expect home robots that fold laundry yet.
Nvidia also introduced GR00T 2, a free-to-use AI model for robot hands and arms. The company says it succeeded 71% of the time on 50 tasks after seeing one demonstration of each, and 95% after five. Both projects are being presented this week at the Conference on Robot Learning, a yearly research meeting in Austin, Texas. The results are the researchers’ and companies’ own.
Source: UC Berkeley News release, Nov. 9, 2026 and Nvidia blog, Nov. 9, 2026.