# Berkeley AI agents write robot programs and test them in simulation first

> UC Berkeley researchers are putting teams of AI agents in charge of writing robot programs, which they first try out in simulation. According to Interesting Engineering, a warehouse robotics company has already rolled out the improvements to 30% of its US fleet.

Published on 2026-11-10

According to Interesting Engineering, UC Berkeley researchers have developed a method that is meant to speed up significantly the programming and training of robots for real-world tasks.

The framework, called Graph-as-Policy, breaks a complex task into simpler, reusable skills, which AI agents then assemble into a complete program.

In one demonstration, a robot had to wipe ink off a metal bar with a sponge. In under an hour, the AI system produced a physical simulation of the scene, including the deformation of the sponge.

In logistics, a warehouse robotics company applied these methods to parcel sorting. According to the article, the trials increased sorting capacity by more than 15,000 parcels a year per robot, and the company extended the improvements to 30% of its US fleet.

The researchers are also releasing the code as open source, built on NVIDIA's Isaac simulation tool, along with a benchmark designed to compare robot performance on industrial tasks.

## Sources

- [Interesting Engineering](https://interestingengineering.com/ai-robotics/new-model-to-transform-robots-programming)

---

HTML version: https://vuntum.com/en/news/uc-berkeley-ai-agents-write-robot-programs-and-test-them-in-simulation

Factual data: CC BY-NC 4.0. Attribution: “Source: Vuntum (https://vuntum.com)” with a link to the page quoted. Editorial text: all rights reserved, short quotation with attribution.
