Robot training data used to mean a human teleoperating a robot. World action models learn physics from video, which means footage of a person doing the job counts. Your process video is an appreciating asset — and the footage you don't record is gone.

Search for humanoid deployment figures and you'll find confident numbers. Most don't come from the companies they describe, and several don't survive contact with corporate filings. Here's what humanoid robots verifiably do right now — measured in hours, totes, and parts.

Meet the Founders of 3GEN Robotics.

Everyone covered BCG's 60% productivity figure. The more useful number was 87 — the share of manufacturers who said talent becomes the constraint on sustaining a deployment. The report describes a race won by deploying faster, in a market where deployment specialists are unhirable.

The robot arm is 25 to 40 percent of what a working cell costs. The rest is the deployment layer, and it's where projects go 40 percent over budget in year one. Why getting robots working, and keeping them working, decides automation ROI.

A policy succeeds ten thousand times in simulation, then misses the mark on real hardware. Nothing is broken. "The sim-to-real gap" names four unrelated failures at once: physics, perception, actuation, and environment. The fix depends entirely on which one you have.

A world action model can adapt to new hardware on thirty minutes of play data, where teleoperation takes weeks. It also runs at roughly 140 milliseconds per decision. Both facts are true, and together they explain why the deployment architecture matters more than the model.

The bottleneck in physical AI isn't compute or model architecture. It's that someone has to physically produce every demonstration a robot learns from. That constraint is about to break.