Berkeley and Nvidia open robot-programming tools for AI agents

Claude Opus 5.5 averaged 77% per-object success on the new VABAR benchmark after two rounds of execution feedback, five percentage points below the 82% human-written average and above other models at 44% to 64%. Simulation iteration was therefore part of the score, not an afterthought.

Researchers at UC Berkeley and Nvidia released three open-source components at the CoRL 2026 robotics conference in Austin: Graph-as-Policy, or GaP; RoboSimStudio; and VABAR. Together, the tools let teams of AI coding agents write, simulate and validate industrial robot programs without collecting robot training data.

RoboSimStudio lets agents tune simulations in the open-source Newton physics engine. VABAR supplies industrial manipulation tasks for evaluating generated programs. The workflow closes a loop: write code, rehearse it in simulation, observe failure and revise the relevant step.

Ken Goldberg, who holds the William S. Floyd Jr. Distinguished Chair in Engineering at Berkeley, said: “Agentic robotics is a fundamental shift in how robots are programmed.” Goldberg credits a harness of thousands of prompts and structures that constrain agent behavior, rather than free-form code generation.

What does GaP do with a robot task?

GaP targets what the team calls variational automation: a robot in a known workcell repeatedly handles parts whose shapes and positions vary. It splits a task into skills such as perception, grasping and motion planning.

Skill agents write and assemble pieces from a library of 51 starting skills, while an orchestrator connects them into one type-checked graph. Parallel simulation rehearsals pin failures to specific nodes, so agents can edit only the failing steps.

What did the real-robot trials show?

The paper reports 25 of 25 successful grocery-fulfillment trials, against eight of 25 for the TipTop baseline; on a UR5 arm, the system completed 143 of 150 USB-C cable insertions. In the reported comparison, a single model writing raw Python without GaP’s decomposition scored zero.

What does VABAR measure?

Berkeley, Nvidia and Siemens built VABAR and released it on Nov. 9. It covers tool sorting, tool hanging, gear meshing, syringe packing, cable insertion and cable routing. Across five of those tasks, average per-object success after two rounds of execution feedback was:

• Human-written programs: 82%
• Claude Opus 5.5: 77%
• Fable 5.1: 64%
• GPT-6 Sol: 53%
• DeepSeek V4: 44%

The authors observed that models starting within a few points of one another ended roughly 30 points apart after feedback. VABAR therefore tests the ability to improve from simulated failure, not merely produce a plausible first program.

Does generated code replace the controller?

GaP maps its graph to a Robot Operating System graph and outputs Python that calls existing vendor modules. KUKA modules, for example, provide low-level control, so the generated program does not replace the robot controller.

The commercial boundary matters: the framework can sit above existing controller software instead of competing with it. Teams can therefore adopt generated orchestration without replacing the vendor layer that executes robot motion.

Who is testing the approach outside the benchmark?

Bosch Research is evaluating the method for industrial automation. Ambi Robotics reports that agentic methods increased throughput across its U.S. fleet, although the company disclosed no figures.

At the same CoRL conference, Nvidia released its GR00T 2 robot model and Isaac Lab 3. The open research effort therefore arrives amid a broader set of Nvidia releases.

The near-term competitive test is whether agents can turn simulation gains into repeatable throughput on live industrial fleets. Until Bosch’s evaluation and Ambi’s deployment evidence become more precise, the commercial advantage remains a forecast rather than a measured result.