PYTHON · NUMPY

What reaches the robot

Of a humanoid foundation model's 440.5 million parameters, reported as 'used for real tests', roughly how many does the robot's control loop run?

Find out how much of a humanoid foundation model runs on the robot, reconcile a released checkpoint to the byte, and see a file-size argument confirm the wrong answer.

01Challenge

A promptable foundation model drives a real humanoid, and the paper reports 440.5 million parameters 'used for real tests'. The authors released both the training checkpoint and the model that deploys. Work out how much of the trained model the robot actually runs, then predict what the deployed file contains before you open it.

Predict firstOf a humanoid foundation model's 440.5 million parameters, reported as 'used for real tests', roughly how many does the robot's control loop run?

02Model

The previous lessons in this module asked how a foundation model acts. This one asks how much of it is on the robot when it does. The model drives a Unitree G1 humanoid through three kinds of task, maximising a reward, reaching a goal pose and tracking a motion, all from one policy that takes a short task description called a prompt. The paper reports 440.5 million parameters used for its real tests, split across six networks, and the authors released both the full training checkpoint and the smaller model they export for deployment. That lets you do something rarely possible: check the paper's numbers against the actual files. You will find that most of those 440.5 million parameters are machinery for training one network, that the share running on the robot shrinks as the model gets bigger, and that the 3.4 gigabyte checkpoint can be accounted for to the last byte. Then comes the part worth remembering. Before opening the deployed file you will predict what it contains from its size, find a hypothesis that fits almost perfectly, and then open the file and find the hypothesis is wrong.

03Guided practice

1. The row, not the printed total. Table 3 prints totals rounded to one decimal. Divide by the sum of the row's entries so a share of the whole row is exactly 1. 2. One line. share(row, parts) is the named entries summed, over every entry summed. Check it on the full row, which must give 1. 3. Watch the share fall. Deepen the model and the actor's share drops, because each block adds 4.2M to the actor and 25.2M to each of three critics. 4. Then do the prediction before the reveal. The last block predicts the deployed file three ways from its size, and one fits to within a tenth of a percent. Decide which you believe before the file is opened for you.

04Feedback

PASS WHEN You found the robot runs 7.24% of a model reported as 440.5 million parameters used for real tests, reconciled a 3.4 GB checkpoint to the byte, and watched the best-fitting guess about the deployed file turn out to be wrong.

If your run fails, check:

If the 6-block share is off in the fourth decimal place, you divided by the printed 440.5 rather than the row's sum, 440.501.
A share of the whole row must come to exactly 1. If it does not, you are summing the wrong entries.
Define share(row, parts) returning the fraction of the row's total that the named networks make up.

05Retrieve & space

Five of the six networks are never run on the robot. Say what each of them is for, and why a robot learning model needs so much that it never ships.
The actor's share falls from about 9.6% to 6.5% as the model deepens. Explain the slopes that cause it, and what it would take to reverse it.
'Actor plus B' predicted the deployed file's size better than the truth did. What exactly made the wrong answer fit, and what is the only way to have known?

06Mastery & project

You can take a published model size, separate what trains from what deploys, reconcile a released checkpoint against its own contents, and refuse to treat a close size match as proof of what a file contains.

Pick any released model you can download. Before opening it, predict its parameter count and file size from its paper or model card, and write down which components you expect it to contain. Then open it and list what is actually there. Record whether a size match would have led you to the right conclusion.

GO FURTHER · THE LIVE PLAYGROUND

The paper ↗ · BFM-Zero: one promptable policy for a Unitree G1, with its architecture in Tables 2 and 3.
The released model ↗ · The training checkpoint and the exported deployment model this lesson reconciles.
PAI-120 · What a learner has to hold ↗ · The same question from the other side: an on-device learner whose weights are a quarter of what it keeps.

← Action Chunking and the Two Systems Dance: Style and the Beat →

PYTHON · NUMPY

What reaches the robot

One foundation model, its paper's table, the files its authors released, and a size that points at the wrong answer.

You read

the arrays and values already in scope

You change

the code you write in each cell

Fixed

the dataset and the checks that grade you