Industrializing Data for Physical AI
Mike Xia, CEO of Anvil Robotics, on humanoid vs. non-humanoid robots
The bottleneck in physical AI is not just training better models, it is building the missing industrial base that makes rich robot data cheap and repeatable. In practice that means cheap cameras alone are not enough. The winning stack also needs affordable force and tactile sensors, reliable robot bodies, and the wiring, drivers, and supply chains that let hundreds of systems collect the same data without months of custom integration work.
-
Anvil sits at the hardware and systems layer because many teams still spend five to six months stitching together arms, cameras, compute, cables, firmware, and OS patches before they can even start collecting training data. That setup tax is the robotics equivalent of needing to build the internet before training LLMs.
-
The current bias toward vision first models is partly a sensor economics story. A camera stack can cost around $12, while force torque sensors often cost thousands of dollars each, which makes force and tactile data much rarer even if they are likely important for getting from decent demos to production grade manipulation.
-
This also explains the split between humanoid builders and component platforms. Humanoid companies like Figure, Apptronik, and Foundation are trying to drop robots into existing factories, while Anvil is betting that many early winners will be narrower manipulation systems and solution companies that need better bodies and sensing more than legs.
Over the next few years, the market is likely to converge on a small set of high value data types, most likely force and tactile among them, and then race to industrialize their collection. Once that happens, physical AI stops looking like isolated robot demos and starts looking like a scaled data supply chain, with clearer winners at the hardware platform and application layers.