Coding agents now program factory robots within five points of a human, Berkeley and NVIDIA benchmark shows

AI coding agents that write robot programs, instead of learning behavior from collected data, came within five points of a human programmer on a new set of factory tasks from UC Berkeley and NVIDIA. On VABAR, a benchmark whose code the teams released on Nov. 9, eight models wrote programs for five tasks, including hanging tools, packing syringes and inserting USB-C cables. Each used the same library of 33 skills and got two rounds of feedback from a physics simulator. Claude Opus 5.5 succeeded 77% of the time overall; the human-written reference scored 82%. Ken Goldberg's group and NVIDIA presented the work at CoRL 2026 in Austin this week, The Robot Report reported on Nov. 11.

For jobs where parts vary but the workcell is known, that makes code the more practical route than end-to-end robot models. The choice of coding model is the lever. "Models that start within a few points of each other end up 30 points apart after two rounds of feedback," the VABAR page says. After those rounds the eight spread from 77% down to 44% for DeepSeek V4.

Shuffle the parts and the VLAs fall away

The harness behind VABAR, Graph-as-Policy (GaP), breaks a task into nodes for perception, grasping and motion planning. Agents write and connect them, then rehearse variants in simulation before the program runs on a real robot. Its output is Python that calls a vendor's existing modules, such as KUKA's, for low-level control. In the GaP paper's simulated tests, vision-language-action models did well when objects barely moved: π0.5 scored 0.96 and MolmoAct2 0.97 on LIBERO, against GaP's 0.95. When the authors permuted object positions, π0.5 dropped to 0.15 and MolmoAct2 to 0.26, while GaP held 0.97. On a real bimanual crate-washing task, GaP succeeded on 143 of 150 tries against 148 for a hand-engineered expert program, and managed 18.33 successful cycles an hour to the expert's 19.33.

Industry is already testing the idea. Bosch is evaluating it for industrial automation, according to The Robot Report. Jeff Mahler, co-founder of Ambi Robotics, which Goldberg advises, said it "converted production data into robot behavior updates, increasing throughput across our U.S. fleet."

The case against is in who did the grading. The same labs built the harness, the skill library, the tasks and the human baseline. The human beat every model on tool hanging, 87% to a best of 60%, and scored only 47% on cable routing, where all eight models reached between 72% and 83%. VABAR's authors say a harder test set is coming. Outside teams running it on their own cells will show whether the five-point gap holds.