Resetting the Robotics Stack for Physical AI
Mike Xia, CEO of Anvil Robotics, on humanoid vs. non-humanoid robots
The reset means robotics is no longer being built around perfect repetition in fenced cells, it is being rebuilt around messy contact with the real world. Older industrial robots were optimized to move fast, lift heavy loads, and never touch the wrong thing. Physical AI flips the target. The hard problem now is handling slightly different boxes, cables, bags, and fixtures every cycle, then learning from those misses instead of failing outright.
-
This is a hardware reset, not just a software update. Mike Xia describes a factory era stack built around high reduction gearboxes, high payloads, and brittle no contact assumptions. That works for ABB and KUKA style cells, but not for assembly, packing, and light manipulation where small collisions and variation are normal.
-
It is also a data reset. Vision became the default input partly because cameras are cheap and everywhere, while force torque sensors have historically cost thousands of dollars and were hard for small teams to integrate. That pushed the field toward pixel to action models first, even though richer sensing likely matters for reliability.
-
The stack is splitting into three layers. Foundation model groups build the brain, hardware platforms supply bodies and sensors, and solution companies win deployments by fitting the system into a specific workflow, like a 3PL packing station or an electronics assembly line. That looks more like the modern AI app stack than old line automation.
The next phase is likely a narrowing from broad experimentation to a standard sensor and model recipe that reliably gets robots from good demos to production speed and uptime. As cheaper data capture spreads across real deployments, the winners should be the companies that can turn contact rich, variable work into repeatable learning loops.