PYTHON · NUMPY

From teleop to a released policy

Record demonstrations from an expert, clone a policy by least squares, and deploy the clone.

01Challenge

This is the top rung: teleoperate, record, train, deploy -- the arc a LeRobot arm follows. An expert controller drives the agent while every (observation, action) pair is recorded. Then clone the expert: fit a straight line action = w*obs + b to the demonstrations by least squares, and deploy that policy while the expert stands down. The starter leaves the slope and intercept at zero; solve them so the clone matches the expert and reaches on its own.

02Model

Behaviour cloning is supervised learning on demonstrations: the observations are the inputs, the expert's actions are the labels, and the policy is the fit. Here the fit is a one-line least-squares slope and intercept, but the shape is exactly what trains an arm from a few hundred teleoperated episodes.

03Guided practice

Step A (worked): the record cell captures 40 (error, action) pairs from the expert. Step B: with w and b at zero the policy commands nothing and never reaches -- the FAIL. Step C (the fix): solve w = (nSxy - SxSy)/(nSxx - SxSx) and b = (Sy - w*Sx)/n; because the expert was a 0.5 gain, the clone recovers w about 0.5. Step D: the deployed clone reaches the target with no expert in the loop.

04Feedback

PASS WHEN PASS when the cloned slope recovers the expert (about 0.5) and the deployed policy reaches the target.

If your run fails, check:

Read the bench’s printed result — it passes when the cloned slope recovers the expert (about 0.5) and the deployed policy reaches the target. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
w and b are still zero -- solve the least-squares fit so the clone matches the expert.

05Retrieve & space

What are the inputs and the labels in behaviour cloning?
Why can the cloned policy run after the expert stops?
How does this scale from a line-fit to a real robot arm?

06Mastery & project

Pass the lab: your clone recovers the expert's policy from its own demonstrations and reaches on its own. You have run the teleop-to-deploy arc that trains real robots.

The full version -- teleoperation, a released checkpoint, and a policy you can drive -- lives in the Robot Learning course and the Institute's released on-device checkpoint.

GO FURTHER · THE LIVE PLAYGROUND

The full arc: the Robot Learning course ↗ — demonstrations to a released, on-device policy.
A released policy you can run ↗ — the Institute's first on-device checkpoint.

← The magnet, switched Back to course →

PYTHON · NUMPY

From teleop to a released policy

Record demonstrations from an expert, clone a policy by least squares, and deploy the clone.

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