PYTHON · NUMPY

Clone the expert

Fit a policy to recorded demonstrations and deploy the clone on goals it never saw.

01Challenge

Hand-writing a policy does not scale; learn it from demonstrations instead. The Bench records an expert driving a 2D tip to random goals, logging (observation, action) pairs, then leaves you a policy whose weights are zero. Solve the least-squares fit that maps observations to the expert's actions, and deploy the clone on thirty fresh goals with the expert switched off. This is behaviour cloning, the method behind every teleoperated robot arm.

02Model

Behaviour cloning is supervised learning on demonstrations: observations are the inputs, the expert's actions are the labels, the policy is the fit. Least squares finds the linear map that best reproduces the expert. Because the fit captures the rule and not the specific goals, the clone reaches goals that were never in the data -- it generalizes across the whole task.

03Guided practice

Step A (worked): the recording loop logs a few thousand (obs, action) pairs from the expert. Step B: the policy weights are zero, so the clone never moves -- the FAIL. Step C (the fix): Wc = np.linalg.lstsq(OBS, ACT, rcond=None)[0].T. Step D: the deployed clone reaches nearly every fresh goal, with no expert in the loop -- the same arc as the live SO-101 lab.

04Feedback

PASS WHEN PASS when the cloned policy reaches at least 27 of 30 fresh goals.

If your run fails, check:

Read the bench’s printed result — it passes when the cloned policy reaches at least 27 of 30 fresh goals. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
The clone does not move -- solve the least-squares fit Wc = np.linalg.lstsq(OBS, ACT, rcond=None)[0].T.

05Retrieve & space

What are the inputs and labels in behaviour cloning?
Why can the clone reach goals that were never demonstrated?
How does this least-squares fit scale up to a real arm and a real camera?

06Mastery & project

Pass the lab: a policy cloned from demonstrations reaches fresh goals on its own. That is the fourth rung -- learning to act.

This is the numeric core of the live SO-101 lab, where you teleoperate, record, train, and deploy on a simulated arm.

GO FURTHER · THE LIVE PLAYGROUND

Do the full loop: the SO-101 lab ↗ — teleoperate, record, train, and deploy a pick-and-place policy in your browser.

← Perception becomes a decision The generalization wall, and what a VLA adds →

PYTHON · NUMPY

Clone the expert

Fit a policy to recorded demonstrations and deploy the clone on goals it never saw.

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