For years, teaching a robot a new job meant one of two slow paths: hand-write the control software, or collect mountains of demonstrations. A UC Berkeley roboticist now says a third path is opening, one where teams of AI agents write the robot’s program, rehearse it in simulation and fix it when it fails.
Ken Goldberg, a professor of engineering at Berkeley, and Spencer Huang, director of robotics engineering at NVIDIA, call the approach agentic robotics. Berkeley News described it on November 9, 2026, as Goldberg presents the results at the Conference on Robot Learning.
Two bottlenecks, one proposed bridge
According to Berkeley, engineers traditionally take one of two routes. Model-based robotics means programming how a robot perceives objects, plans motion and controls its actions, which takes extensive engineering. Model-free robotics trains behaviors from demonstrations through imitation learning, which needs enormous amounts of data and time.
Goldberg has argued before that a “100,000-year data gap” stands between robots and the skill AI agents gained from internet text. He describes agentic robotics as a missing link between the two methods, one that, in his words, does not require any additional data.
How agents build a robot program
The team’s framework is called Graph-as-Policy. It splits a robot task into individual skills, such as finding an object, planning a grasp and moving the gripper without collisions. Goldberg compares the skills to building blocks.
AI agents then generate those skills and connect them into a program. Using NVIDIA’s Isaac simulation libraries, the agents rehearse the program across variations of the task, spot where it fails and refine the approach before anything is deployed. Because the output is structured code rather than an opaque model, Goldberg says human engineers can still interpret it for fine-tuning and verification.
The sponge-and-bar test
Berkeley highlights one demonstration. Postdoctoral researchers Eric Kaiyuan Chen and Shuangyu Xie gave OpenAI’s recently released Astra model access to their simulation interface and a short video of a real robot using a sponge to wipe ink off a metal bar. In less than an hour, the model built and tuned a physics simulation that reproduced the video, including realistic deformation of the sponge.
The AI then took that simulation plus Graph-as-Policy, generated a control program, and tested and improved it repeatedly. When run on a real robot, it worked. In one trial the robot pushed the bar over because it had not reliably inferred its mass. That failure became new evidence: the agent used video of the miss to improve the simulation and revise the program. Goldberg said this was when he got goosebumps.
A warehouse data point
The release also names an industry user. Ambi Robotics, a Berkeley company focused on logistics and warehouse order fulfillment, applied the methods earlier this year to analyze parcel-sorting data and test updates to robot behavior. Berkeley reports tests showed added capacity of more than 15,000 packages sorted per year per robot, and the change is now deployed across 30% of the company’s U.S. fleet. Ambi’s co-founder and CTO, Jeff Mahler, said production data was converted into behavior updates while engineers stayed in control.
The team is also releasing open-source code built on NVIDIA’s Isaac framework, along with an industry-standard benchmark, called VABAR, for evaluating robot performance on industrial tasks.
What it means
If the approach holds up, robot builders could spend less time hand-tuning behaviors and more time defining tasks and checking results. Repetitive, high-volume industrial settings look like the first fit. Goldberg is explicit that home assistants folding laundry should not be expected yet. The figures and demonstrations come from the researchers and a partner company, and independent replication will matter, but open code and a shared benchmark make that testing easier.
The bottom line
Agentic robotics reframes robot development as a loop of generate, simulate, fail and revise, run largely by AI agents. Goldberg calls it a true paradigm shift; the next year of open-source experiments will show how far it travels beyond the lab and the sorting line.