Indoor Work Data Is the Moat
Sankaet Pathak, CEO of Foundation, on why humanoids win in robotics
The real moat in humanoid robotics is not the robot body, it is access to indoor work data from real jobs. Tesla already has a giant driving data engine, but roads do not teach a robot how to carry parts through tight factory aisles, place sheet metal into fixtures, recover from mistakes, or work around people and equipment. That leaves every humanoid company racing to win deployments that generate labeled action data inside buildings.
-
Foundation describes the near term loop very concretely. Put robots on high attrition, repetitive factory tasks, use teleoperation when the model fails, label those interventions, then feed that data back into the action model. In robots, unlike cars, partial autonomy creates little value, so useful deployment and learning have to happen together.
-
Figure has already shown what this indoor loop looks like in practice at BMW. Its robot handled sheet metal loading on an active assembly line, moved 90,000 plus parts, and logged 1,250 hours. That matters less as a one off demo than as a stream of real manipulation data from a production floor.
-
Tesla still has the broadest real world autonomy dataset in transport. Tesla says FSD is trained on billions of miles of driving data, with more than 3 billion cumulative FSD miles as of January 2025 and 2.5 billion telemetry packages received in Q3 2025 alone. The strategic gap is that this data is overwhelmingly road shaped, not factory shaped.
The next phase of the market will be won by whoever turns early factory fleets into a compounding indoor data advantage. Once a company has robots working daily across many buildings, tasks, layouts, and edge cases, model improvement should start to look more like Tesla on roads, but for warehouses and factories, and that is when the category begins to tip from pilots to scaled deployment.