HumanoidTraining
Menu

A FabSpeak project · AI × humanoid robots × the factory floor

A humanoid robot is not programmed. It is trained. Here is how.

Teleoperation, simulation, human video and foundation models, taken apart at the layer that decides whether the robot actually ships: the mechanism, the number, and the caveat.

Start with how it works What it means in a plant
Ram Agarwal on a factory floor, a robot visible behind him
Ram Agarwal · twenty years in the software that runs factories · fabspeak.com

The stack, in one picture

Human video scene priors, zero robot data Simulation scale, edge cases, locomotion Teleoperation a person drives the robot; every move becomes a labelled example. 5–50 episodes per operator-hour. Vision-language-action model sees a camera frame, reads an instruction, outputs joint targets 50–200 times a second Factory floor
The stack, bottom to top. Each layer fixes the weakness of the one below it. Nobody skips the teleoperation layer yet.
30,000+BMW X3 vehicles built with Figure robots on the Spartanburg line
540ksynthetic demonstrations NVIDIA generated from a few dozen human ones
~7–8kcommercial humanoids operating worldwide, mid-2026 estimate
$38BGoldman Sachs humanoid market estimate for 2035

Pick a door

How a humanoid learns

Why writing code for 40 joints stopped working, and what replaced it. The four-layer stack, explained once.

Where the data comes from

Teleoperation rigs, simulators, and plain video of people working. What each one is good for, and what each one gets wrong.

The factory floor

Which companies are really running robots in plants, what the robots actually do, and why a fab is a harder room than a car plant.
The hard part of a humanoid is not the body. It is the ten thousand hours of someone showing it what to do.

That one sentence is the whole site. The pages that follow are the evidence for it.