Foundation Models as Robot Brains
Mike Xia, CEO of Anvil Robotics, on humanoid vs. non-humanoid robots
The real prize in physical AI is becoming the default model layer that every robot application company builds on. In practice, that means training a general policy that can take in camera feeds, eventually force and tactile data too, and output robot actions, then selling that capability to teams that do not want to build core learning systems themselves. The model company owns the shared intelligence layer, while customers handle deployment, workflow fit, and customer specific data collection.
-
The stack is splitting into three jobs. Foundation model companies build reusable control models. Hardware providers supply robot bodies, sensors, and control infrastructure. Solutions companies combine both into a working product for a warehouse, factory, or kitchen, then integrate with the customer’s real workflow.
-
Owning the brain does not mean owning the customer. The company with the deployment often wins the relationship because it collects failure cases on site, tunes the system for a narrow task, and pushes performance from mostly works to reliable enough for production.
-
This looks a lot like the LLM market structure. A small number of model builders sit underneath a larger set of application companies, even when the base model is general. In robotics, that pattern is stronger because each deployment also needs physical setup, safety tuning, teleoperation, and site specific integration.
Over the next few years, the most valuable brain providers will be the ones that turn multimodal robot data into a standard developer layer for manipulation. As force and tactile sensing get cheaper and easier to integrate, the model layer should consolidate around a few suppliers, while many more vertical application companies grow on top of them.