Berkeley and NVIDIA use coding agents to write robot code

At the 2026 Conference on Robot Learning, UC Berkeley and NVIDIA showed agentic robotics. In this method, AI coding agents write robot code that a person can read. Professor Ken Goldberg of Berkeley tells that the data method is at a plateau for industrial work, where high reliability is mandatory. The source gives no test numbers for the new method.

How it works

A task, for example to put groceries in a bag, is divided into smaller tasks. Each one goes to one coding agent, for example for view, hold and arm movement. The agents operate their code in a simulation, find errors and change the code until it works.

The proof

An agent made a simulation the same as a video, Goldberg tells. Goldberg tells that the difference was hard to see. This is one example, not a measured test.

What is not known

Torsten Kroeger, chief science officer at Intrinsic of Alphabet, tells that no physics engine can model physical contact correctly. The distance between simulation and the physical world is not solved.

Insight: NVIDIA tools serve both robot methods

Provision Labs, started on 26 October, collects data for robot training and tests. Spencer Huang of NVIDIA helped to make the code method known. On 9 November NVIDIA showed software for lower agent cost, and a vice president tells that token use is 2 times more each month. This reads as a toll gate for NVIDIA: the data method and the code method each use its tools.