# UC Berkeley and NVIDIA present agentic robot programming research at CoRL 2026

> Original title: Can AI agents program industrial robots without training data? Berkeley says yes

- Source: The Robot Report
- Published: 2026-11-11T12:55:16.000Z
- Bad Influence: https://badinfluence.news/items/kc9ld1a37i80kxmdszmxe2w2e
- Original: https://www.therobotreport.com/can-ai-agents-program-industrial-robots-without-training-data-berkeley-says-yes/

## Summary

UC Berkeley and NVIDIA presented Graph-as-Policy, a framework that splits robot tasks into skills that AI agents generate, connect, and test in simulation, at CoRL 2026 in Austin. They also introduced the VABAR benchmark, which covers six industrial tasks with standardized variations and difficulty tiers, built on Isaac Lab-Arena and Newton. Goldberg says a graduate student's RoboSimStudio interface let a frontier model produce a Newton simulation closely matching a video within about an hour.