Closing Note
about 1 minutes
The honest summary of August 2026 is this. Robot hardware is getting cheap fast and demonstrably so — a 71% two-year decline in humanoid ASP is not a marketing claim, it is a filed financial statement. Robot intelligence is improving smoothly, measurably, and without any sign of the discontinuity the valuations assume. The binding constraint remains data, and the field has produced five genuinely different answers to it, of which the least-discussed — putting the human's demonstration directly into the robot's action space — may be the most important.
The three institutions that will look best in retrospect are probably not the ones with the highest valuations. They are the ones that built the measurement apparatus: TRI for insisting on blind evaluation, Stanford for building interfaces cheap enough that everyone could participate, and the applied companies in Chapter 6 that have been quietly accumulating real contact data while everyone else argued about form factors.
The three questions in §13.9 remain open. Anyone who tells you they are closed is selling something.