Humanoids win through data
Sankaet Pathak, CEO of Foundation, on why humanoids win in robotics
The bottleneck in humanoids is shifting from invention to collection. Once motors, cameras, onboard compute, and transformer based control are good enough, the hard part is getting huge amounts of real robot experience, the video, sensor traces, actions, and corrections needed to teach a machine how to handle messy factory reality instead of a scripted demo. That is why early deployments matter so much, they are product rollouts and data harvesting systems at the same time.
-
Older robots worked by forcing the world to be simple. Parts arrived in fixed positions, humans stayed out of the cell, and engineers hard coded motions. The new promise is end to end policies that can react inside unstructured spaces, but those policies only improve if fleets see enough edge cases in the wild.
-
That makes customer workflows the training set. Foundation describes starting with repetitive, high attrition factory jobs, then using teleoperation when the robot gets stuck, so each intervention becomes labeled failure data for retraining. In practice, the service contract buys labor replacement and also generates the examples needed to expand capability.
-
The same logic explains the competitive map. Tesla has enormous road data but not indoor autonomy data. Figure, Agility, and Foundation are all racing to place robots in real workplaces because whoever gets the densest deployment loop can compound data fastest. In humanoids, installed base is not just revenue, it is model fuel.
The next phase of humanoids will look less like a hardware unveiling race and more like a fleet learning race. The winners will be the companies that can keep robots useful enough to stay on the floor, collect intervention data every day, and turn those deployments into a compounding performance advantage across more tasks, more sites, and eventually more markets.