What the demo videos actually show

Two humanoid robots kit a set of tools, close the case, and put it away. They talk to each other while they do it. Each one runs its own copy of the same model.

You cannot build that. Not this year, not with any budget you have.

But about a third of it shipped to a public API, and a different third is on Hugging Face with open weights — and by the end of this course you will be talking to an arm that costs less than a phone, with the network cable unplugged.

It is two models, not one

This is the most common misreading in robotics right now.

The reasoning model (ER) takes an image and a prompt and returns text and coordinates. Where things are, what is occluding what, what order to do things in. This is on the public API. Any key, no application.
The action model (VLA) is what actually drives the joints. Early-access partners only.

When you watch that reel you are watching two products. One has a URL and a price; the other has an application form. That confusion is expensive, because it is why people buy the wrong hardware.

Every model id, price and access tier in this course moves. Check the date on any claim and follow the link to current docs.

Which one drives the joints?

Two models produced the behaviour in that reel, and only one of them is the reason you cannot rebuild it. Which one actually moves the arm — and what does it take to get at it?

The second wall

Suppose the gated half opened tomorrow. You still could not ship it, for two reasons.

Connectivity. The robot is in a vineyard, a barn, or the back of a warehouse under two feet of concrete. If a perception step needs a round trip to a data centre, the robot stops working exactly where it was supposed to earn its money.

Cost. One call costs a fraction of a cent, which feels free. Put it inside a machine that looks at the world several times a second, for eight hours, every day.

A robot billed per token gets more expensive the more work it does. That is exactly the wrong shape for a machine whose entire pitch is doing work cheaply.

So what is the frontier model for?

It is the ceiling. You measure against it so you know precisely what the local-only constraint costs you, on your task, as a number rather than a feeling.

The frontier model is your teacher and your measuring stick. It is never your runtime.

That inversion is the spine of this course, and it is the opposite of how most AI content teaches you to build things.

Teach back the system boundary

Without looking back, explain why the reasoning model can be useful even though it cannot be the robot's runtime.