UC Berkeley, NVIDIA release open-source Graph-as-Policy and VABAR for AI robot coding

The release stems from a practical bottleneck in industrial robotics: programming, tuning, and troubleshooting workcells has traditionally taken weeks to months, and software teams often start only after hardware is installed, putting them under deadline pressure.

The tools address this by combining a modular framework, Graph-as-Policy, which splits tasks into skills like perception, grasp planning, and collision-free motion, with a benchmark, VABAR, that measures progress on six practical manipulation tasks. A key enabler is RoboSimStudio, an interface built by graduate student Eric Chen that makes the hundreds of parameters in the Newton physics engine tractable. According to Goldberg, when Chen gave GPT-6 Astra the tool and a video in September, it produced a Newton simulation closely matching the video within about an hour.

Industry collaborators are already evaluating the approach. Bosch Research is assessing agentic robotics for diverse industrial automation scenarios, and Ambi Robotics reported increased throughput across its U.S. fleet while keeping engineers in control.