Robot Foundation Models Work. What Breaks Next?

Notes from the Bay Area Robotics Symposium, 2025 vs. 2026

On October 30 I attended the Bay Area Robotics Symposium (BARS) for the second year. Faculty from UC Berkeley, Stanford, UC Santa Cruz and UC Davis gave dozens of short talks, and industry and student speakers added spotlights.

My notes from the two years show a shift. In 2025 speakers asked how to build robot foundation models. This year they asked why the models still fail.

What stayed the same

Robots still lack data. Language models train on the text of the internet in billions of hours; a leading robotic model two years ago was trained on ten thousand hours of robotic data. To accelerate, labs edit human videos to resemble robot demonstrations, build cheap teleoperation rigs, rebuild real scenes in simulation and generate synthetic data.

Robots still forget. Models lose track of earlier steps in a long task, and they lose old skills when they learn new ones. Researchers proposed retrieving only the key moments of a task and anchoring old skills in memory. No one claimed a solution.

Simulators remain hard to use. Researchers praised the leading physics engines and again complained about them. Contact, soft materials and small forces still resist accurate simulation.

What was new

Researchers now agree on why vision-language-action models fail. A model reaches about 80% success on the tasks it trained on; the last 20% remain the hard part. Speakers named three causes: The models do not master tasks, they do not generalize beyond their training, and they do not apply force precisely. Most of this year's presentations addressed one of the three.

Mastery. Last year researchers proposed adding reinforcement learning to these models. This year they showed models trained that way, and those models perform tasks more reliably. One lab simplified the training by stating the reward in plain language instead of building a separate reward model.

Generalization. Researchers showed models that reason in images and words before they move. An autonomous-driving company described two of them: One world model runs in the car and predicts how the scene will unfold. The other, built on a video model, generates unusual scenarios on which the driving system trains. Some researchers even call this latter language model an implicit world model, but they do not predict failure states of a scene from actions.

Force. Touch moved to the center. Last year tactile work meant mostly new robot hands. This year several researchers argued that the next advances will come from touch, not vision. Speakers also named hardware limits: actuators overheat, and designers must choose between three fingers for precision and four for force.

The problem as a whole. Two developments address all three causes at once. First, companies in driving and humanoids now split control into a fast reflexive layer and a slower reasoning layer, so that one layer supplies reliability and the other breadth. One humanoid maker even uses three layers. Second, coding agents entered robotics. Researchers asked whether AI agents can write robot software. Long programs confuse the agents, so some groups break the program into a graph of small parts that agents build and revise one at a time. Because the agents need no demonstrations, the approach sidesteps the data shortage.

Andra Keay of Silicon Valley Robotics gave a spirited afternoon keynote. She asked what is really happening with humanoids, whether robots should be generalists or specialists, and what animals can teach us about the choice. 

This year's research presentations supported answers to each choice. Foundation models push toward generalists, while layered designs and coding agents favor reliable specialists.