the robots are getting good. in the last ninety days alone, a foundation model operated an appliance it had never seen in training. another ran a commercial espresso machine autonomously for thirteen hours straight.

and yet most robotics deployments still fail quietly, in the gap between what a robot can do in a lab and what it can do reliably in production. the narrative says this is a data and training problem. the evidence suggests something else.

three camps are building physical AI right now. real-world data, simulation, foundation models. they're becoming layers of the same stack. the question is who owns the layer where it all comes together.

part i: teaching robots with the real world: what 'real-world data' actually means across the companies building it, and why the supplier tier is consolidating faster than most people expect.

part ii: more data won't fix this: why the bottleneck isn't data volume. normalisation, indexing, and observability matter more. and why three alternative approaches have emerged in parallel.

part iii: who owns the assembly: how the three camps converge into a single stack, and why the eval-and-deployment flywheel emerges as the defensible moat.