PYTHON · NUMPY

Compose LeRobot — one adapter, any checkpoint

Drive different LeRobot architectures through one `.select_action` adapter, with action chunking.

01Challenge

LeRobot standardized .select_action(obs) across ACT, Diffusion Policy, and more. Write one adapter that drives any of them, consuming an action chunk before re-inferring, and show it works for two different checkpoints.

02Model

Don't rebuild the model hub — compose it. LeRobot already made ACT, DiffusionPolicy, TDMPC2, VQ-BeT, SmolVLA, and pi0 share one method: select_action(obs). So a single adapter drives them all; a skill just names a checkpoint. This is the flywheel: adding a skill is naming a string, and it inherits the same contract and envelope.

03Guided practice

1. The interface. Every checkpoint exposes select_action(obs). 2. Chunking. It returns a horizon of actions; consume them before inferring again. 3. One adapter. A closure that caches the chunk and steps a cursor. 4. Two checkpoints. Same adapter, different policy — proof it's architecture-agnostic.

04Feedback

PASS WHEN PASS when one adapter drives both checkpoints and chunking re-infers only when spent.

If your run fails, check:

Read the bench’s printed result — it passes when one adapter drives both checkpoints and chunking re-infers only when spent. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
The adapter must cache the chunk and re-infer only when the cursor passes the horizon.

05Retrieve & space

What method do all LeRobot policies share? (select_action(obs)).
What is action chunking? (a policy returns a horizon of actions; you consume them before inferring again).
Why compose instead of rebuild? (LeRobot already won the model+interface layer; ride it).

06Mastery & project

You can wrap any LeRobot checkpoint behind one adapter and explain chunking.

The habit: sit above LeRobot. A skill is a contract plus a checkpoint string.

GO FURTHER · THE LIVE PLAYGROUND

The registry's arm-pick-place & arm-stack compose real LeRobot checkpoints ↗

← One contract, any policy MCP for robots, done right →

PYTHON · NUMPY

Compose LeRobot — one adapter, any checkpoint

Write the adapter once; drive ACT and Diffusion with it.

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