Nvidia and Berkeley Say Coding Agents Hit 99% on Robot Tasks

UC Berkeley and Nvidia researchers say teams of AI coding agents can write robot-control programs that reach up to 99 percent success on industrial tasks, against about 20 percent for imitation learning on the same tasks, according to IEEE Spectrum.

IEEE Spectrum senior editor Evan Ackerman reported the work on Nov. 9, 2026, ahead of its presentation at the Conference on Robot Learning. The 99 percent figure is the research team's own, from its controlled tests, and the article reports no independent replication.

How the method works

Agents write a control program, run it in a simulator, read the failure and revise the code. The researchers pair this with a framework they call Graph-as-Policy. Nvidia's Isaac Lab-Arena and its Newton physics engine supply the simulated tests.

Ken Goldberg, a robotics professor at UC Berkeley and cofounder of Ambi Robotics, said "The imitation-learning paradigm has hit a plateau." He added: "It's not clear that adding more data is going to get you there." Imitation learning trains a robot from recorded human demonstrations.

Spencer Huang, Nvidia's director of product for robotics, described setups where one agent writes code and another checks it. "We all know that AI agents can create slop," Huang said.

The numbers the article gives

Measure	Figure	Source
Peak success, agentic method	99%	Research team
Success, imitation learning, same tasks	About 20%	Research team
Retasking one robot for a new product	8+ hours by hand	Wistron, per IEEE Spectrum
Retasking with agents	A minute or two	Wistron, per IEEE Spectrum

1. Agentic robotics, peak99 %
2. Imitation learning, baseline20 %

Source: IEEE Spectrum, citing the Berkeley and Nvidia team's own tests, accessed 2026-11-11

One task the piece describes is semiconductor placement, where parts must land within 0.3 millimeters. The retasking comparison comes from Wistron in Taiwan, which the article cites as needing at least 8 hours per robot for each new product by hand.

Where the claim is weakest

Torsten Kroeger, chief science officer at Alphabet's Intrinsic and not part of the research, raised two objections. "There's just no physics engine out there that can accurately model physical contact with the world," he said, which is the gap between simulation and a real factory floor. He also asked "how do we know that these agents are really doing the right thing all the time?"

IEEE Spectrum said the approach has been tested mainly in controlled settings. Nvidia's GR00T humanoid platform, covered in the earlier report on GR00T 2 and Isaac Lab 3.0, is described as not yet precise enough for such tasks. Huang's own framing is narrower than general intelligence: "What you get with agentic robotics is near-term applications."

What the earlier CaP-X paper found

Goldberg is a co-author, alongside Linxi "Jim" Fan, on a separate arXiv paper, CaP-X, first submitted March 23, 2026. It tested 12 frontier models on robot manipulation and found performance falls when human-designed building blocks are removed. The Decoder reported on April 2 that none of the 12 matched human-written programs in a single attempt.

Its training-free agent, CaP-Agent0, matched or beat human-written code on 4 of 7 tasks. A reinforcement-learning variant lifted a 7-billion-parameter Qwen model from 4 percent to 44 percent on simulated cube stacking, and reached 76 percent on a real Franka arm without extra real-world training, The Decoder said. Those are different tasks from the 99 percent industrial figure, and the two should not be compared directly.

Another robotics platform is covered in the report on Boston Dynamics' Atlas wheeled research testbed. Goldberg's own summary of the Berkeley result was cautious: "I don't want to say that we have the solution, but I'm very excited about it." IEEE Spectrum gives no formal title for the new paper and links only a project page.