PYTHON · NUMPY

The demonstrator's secret

The identical expert controller drives four bodies from free pivot to very stiff. Which body gives it the best tracking of the moving target?

Measure how much of an expert demonstrator's competence belongs to its body, by running the identical controller on four bodies and reading region and tracking separately.

01Challenge

One full-state expert controller, four pivot stiffnesses, noisy episodes, a moving target. Measure the fraction of episodes survived and the tracking error, per body.

Predict firstThe identical expert controller drives four bodies from free pivot to very stiff. Which body gives it the best tracking of the moving target?

02Model

Every imitation dataset records an expert. Here is a question the dataset cannot answer: how good is the expert, and how good is the expert's body? The two are fused in every frame, because the camera sees only their sum. In simulation we can do what no dataset can: hold the controller fixed, swap the body, and difference the results. The same four gains drive a free pivot, a soft spring, a passively stable spring, and a very stiff one. Two numbers per body, and they answer different questions. Region is the fraction of episodes that end with the pole still up: can this pairing operate at all? Tracking is how closely the cart holds a moving target while alive: how well does it work when it works? Watch what the body does to each. And watch the stiffest body closely, because it is the first hint in this course that the substrate is not free either: at some point the spring the policy leans on becomes a spring the policy fights.

03Guided practice

Run the table, then read the two columns against each other: region rises with stiffness and saturates, tracking improves and then turns.

04Feedback

PASS WHEN You measured the same controller scoring 0.271 to 1.000 depending on the body it wears.

If your run fails, check:

Region 0.271 on the free body against 1.000 above the passive threshold: the demonstrator's reliability is mostly the spring's, and no frame of video says so.
Tracking improves too, 1.009 down to 0.376 metres, because a body that handles the pole frees the controller's authority for the target. Then kp=2.00 pays 0.407: past sufficiency, stiffness becomes a tax.
Tracking is averaged over surviving steps only. Keep that in mind: a metric conditioned on survival can flatter an early death, and a later lesson turns exactly that blade.

05Retrieve & space

Why do region and tracking answer different questions, and why does conflating them mislead?
The controller is identical on every body. Where does the 0.271 to 1.000 spread come from?
Why does the stiffest body track slightly worse than kp=1.00?

06Mastery & project

Given the same policy's scores on two bodies, attribute the difference to substrate against controller and say what evidence would change the attribution.

Take any published imitation-learning result you know. Write down what share of its reported success you can attribute to the robot's mechanics from the paper alone. If the answer is none, you have found this course's thesis in the wild.

← Balance with the motor unplugged Two demonstrations against one hundred and twenty-eight →

PYTHON · NUMPY

The demonstrator's secret

The expert's gains never change. Only the body does. Forty-eight noisy episodes per body, tracking accumulated only while the pole is up.

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