PYTHON · NUMPY

Capstone: Build the Whole Pipeline

Your pipeline lands short of 98 percent. You have normalisation off and the learning rate low. Which do you fix?

Assemble everything (normalize, split by episode, train, and verify on held-out episodes) into one shippable policy, then go make your own in Forge.

01Challenge

Run the full pipeline and reach at least 98% held-out variance explained. It only works if you apply both key lessons: normalize the data and use a large enough learning rate.

Predict firstYour pipeline lands short of 98 percent. You have normalisation off and the learning rate low. Which do you fix?

02Model

This is the whole course in one cell. You have every piece now: turn demonstrations into data, normalize it, split by whole episodes, train a neural policy by gradient descent, and score it on episodes it never saw. Here you put them together and hit a real bar, ninety-eight percent of the held-out action variance, the mark of a policy worth shipping. Two decisions are left to you, and they are the two that matter most: whether you normalize, and whether your learning rate is enough to converge. Get both right and the pipeline works. Then the last step is the real one. Open Forge, record your own demonstrations on the arm, train a policy, and export the dataset, the exact loop that produced the Institute's released checkpoint. You started by watching a policy; you are finishing by making one.

03Guided practice

1. Assemble. The cell already wires the full pipeline: episode split, train-only stats, the MLP, and the held-out score.

2. The two decisions. Set NORMALIZE = True (lessons 2 and 5) and raise LR to 0.2 (lesson 4). Either one wrong and it falls short of 98%.

3. Make it real. Then use the links below: record your own reach dataset in Forge and train a policy, run the released checkpoint live, and see the open dataset it learned from.

04Feedback

PASS WHEN PASS when the full pipeline reaches at least 98% held-out variance explained (requires NORMALIZE=True and a large-enough LR).

If your run fails, check:

Short of 98% - you need BOTH: NORMALIZE=True and LR=0.2.
Only one fix applied - check both decisions.
Read the bench’s printed result, it passes when the full pipeline reaches at least 98% held-out variance explained (requires NORMALIZE=True and a large-enough LR). Run the cells top to bottom; if one errors, the FAIL line names the step to fix.

05Retrieve & space

What two preparation choices most decide whether the policy learns? (normalizing the data, and a large-enough learning rate).
What score marks a shippable policy here? (at least 98% held-out variance explained).
What is the real final step? (record your own dataset in Forge and train a policy).

06Mastery & project

You can assemble a complete behavior-cloning pipeline to a shippable bar and reproduce the loop that made the released checkpoint.

The whole arc, yours to run: demonstrations to dataset to a held-out-verified policy to a checkpoint you can release.

GO FURTHER · THE LIVE PLAYGROUND

Record & train in Forge ↗ · Reach Trials → record the expert → train an MLP → export a LeRobot dataset. The exact loop that made the released checkpoint.
Run the released checkpoint ↗ · Watch the published weights drive the same arm live, on-device.
The dataset it learned from ↗ · 11 episodes, generated on-device.

← How Many Seeds Does a Ranking Need?Back to course →

PYTHON · NUMPY

Capstone: Build the Whole Pipeline

Assemble the entire course into one pipeline and reach a shippable held-out score.

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