Physical AI Mirroring Early LLMs
Mike Xia, CEO of Anvil Robotics, on humanoid vs. non-humanoid robots
The real signal is that physical AI is moving from a science project into a model race, but it is still in the stage where teams are proving that more data and better sensors improve capability before the market agrees on the winning recipe. In practice, that means labs can already get robots to do simple manipulation tasks, but the leap to production grade reliability still depends on finding which data types, like force and tactile, push success rates from good enough demos to fast, repeatable work.
-
The stack is splitting into three layers, brain companies like Physical Intelligence and Skild, hardware suppliers like Anvil, and solution companies that adapt robots to a warehouse, factory, or kitchen workflow. That is similar to early LLMs, where general models existed first, then application companies won by tuning for a specific job.
-
The bottleneck is not only model size, it is data collection. Cameras are cheap and everywhere, so vision led training dominates today, while force torque sensors still cost thousands of dollars and are harder to integrate across large fleets. That makes current progress look more like early prompt demos than stable production software.
-
Comparable research across humanoids shows the same pattern. Big names have raised heavily, but the real competition is building a data flywheel from deployments, teleoperation, and synthetic or video based training. Revenue is still thin across much of the space, which is exactly what an early platform shift looks like before the breakout products arrive.
The next phase is likely a fast collapse from broad experimentation into a narrower standard stack. Once a few labs show that adding the right non visual signals reliably lifts task performance, the market will standardize around those sensors, data pipelines, and deployment loops, and companies that already sit close to live customer workflows will have the clearest path to turning model progress into real revenue.