PYTHON · NUMPY

One contract, any policy

Swap an analytic policy for a learned one without touching the contract, envelope, or wire.

01Challenge

Run an analytic controller and a stand-in for a learned policy through the SAME envelope and the SAME wire encoder, and show the robot sees identical, valid commands either way.

02Model

A skill is policy-agnostic: the contract, the safety envelope, and the transport don't care whether the action came from ten lines of control or a 450M-parameter VLA. Both are just obs -> action. Swap the policy; keep everything else. That's the 'switch the model in one line' the embodied stack was missing — supplied by the contract, not the model.

03Guided practice

1. The interface. A policy is obs -> action. 2. Two of them. An analytic drive-to-target and a stand-in learned policy. 3. Same tail. Both flow through one safety_clamp and one Feetech encoder. 4. Compare. The resulting joint commands — and the wire packet — match.

04Feedback

PASS WHEN PASS when both policies produce identical, valid commands through the shared contract.

If your run fails, check:

Read the bench’s printed result — it passes when both policies produce identical, valid commands through the shared contract. Run the cells top to bottom; if one errors, the FAIL line names the step to fix.
Both policies must flow through the SAME safety_clamp and feetech encoder.

05Retrieve & space

What is a policy, reduced to its interface? (obs -> action).
What stays fixed when you swap the policy? (contract, safety envelope, transport, eval).
Why does this matter for adoption? (you can drop in any model without rewiring the robot).

06Mastery & project

You can swap a policy under a fixed contract and show the deployment is unchanged.

The habit: the policy is a plug-in; the contract is the socket.

GO FURTHER · THE LIVE PLAYGROUND

The live demo swaps analytic ⇄ SmolVLA on one contract ↗

← The safety envelope lives in the runtime Compose LeRobot — one adapter, any checkpoint →

PYTHON · NUMPY

One contract, any policy

Two policies, one contract. Prove the robot can't tell which drove 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