Why the humanoid form factor keeps winning
Physical AI research could target any robot morphology, wheeled bases, robot arms bolted to a table, quadrupeds, yet humanoid robots have become the most popular platform for foundation model research specifically. The reason connects directly back to data. The world's buildings, tools, vehicles, and workflows were all designed around the human body, doorknobs at human hand height, stairs sized for human legs, tools shaped for human grips. A human-shaped robot can operate in that environment without it needing to be redesigned, and it can be trained on the enormous existing supply of human demonstration data, like videos of people doing tasks, or motion-captured human movement, in ways that map far more directly onto a humanoid's joints than onto a wheeled robot with a completely different kinematic structure.
This creates a flywheel: more humanoid hardware in the world generates more humanoid-specific data, which improves humanoid-specific foundation models, which makes humanoid robots more capable and commercially useful, which funds more humanoid hardware. It's not that the humanoid form is mechanically optimal for every task, a wheeled base is often more stable and efficient for flat-floor logistics, it's that the humanoid form is data-compatible with the human world in a way that pays compounding dividends for a data-hungry learning approach.