PYTHON · NUMPY

The Bridge, Not From Scratch

How many demonstrations does a modern robot policy need to learn a new reach task, when it STARTS from a pretrained base?

Show why a robot policy learns from a handful of demonstrations when it starts from a pretrained backbone's features, versus needing far more data from scratch.

01Challenge

Fit a robot's action mapping two ways from the same 10 demonstrations: on top of a pretrained backbone's meaningful features (the bridge), and from raw high-dimensional inputs (from scratch). The bridge nails it; from scratch cannot.

Predict firstHow many demonstrations does a modern robot policy need to learn a new reach task, when it STARTS from a pretrained base?

02Model

The loudest worry in robotics is data scarcity: we have nothing like the internet-scale corpus that trained language models. Tedrake reframes it. Nobody trains a robot from scratch. You start from a model that already has common sense about the world, a vision-language or world-model backbone, and you learn only the bridge that adds one new output: actions. This is the shape of the 2026 stack, from Physical Intelligence's pi-0.5 generalizing to new homes, to NVIDIA's GR00T, to Google's Gemini Robotics. The data you need is not everything ever; it is enough to map an already-understood scene into robot trajectories. This lesson makes that concrete: the same demos that are hopeless in a raw 64-dimensional space fit cleanly on top of the 4 features the backbone already provides.

03Guided practice

1. Two fits. bridge should fit on the k=4 pretrained features (Ztr,Zte); scratch fits on the D=64 raw features (Rtr,Rte). 2. Fix the bridge. Change bridge = rmse(Rtr, Rte) to rmse(Ztr, Zte). 3. Why it works. With 10 demos, 4 features are well-determined; 64 are not. 4. Confirm. Bridge RMSE under 0.1 while from-scratch is several times worse.

04Feedback

PASS WHEN PASS when the bridge (head on pretrained features) reaches low error from 10 demos while from-scratch does not: the reframed data problem.

If your run fails, check:

Bridge equals from-scratch: you fit both on the raw features. Change the bridge to `rmse(Ztr, Zte)`.
Bridge still high: confirm you pass the k-dimensional `Ztr`/`Zte`, not the 64-dim raw arrays.
Both low: with only 10 demos the 64-dim fit should be underdetermined; check D and the demo count n=10.

05Retrieve & space

What is Tedrake's reframe of the robotics data problem? (up-train a backbone; learn only the action bridge).
Why do 10 demos fit 4 features but not 64? (determined vs underdetermined system).
Name two 2026 robot foundation models that follow this backbone-plus-actions pattern (e.g. pi-0.5, GR00T, Gemini Robotics).

06Mastery & project

Demonstrate the data efficiency of learning an action head over a pretrained backbone: the strategy the whole VLA field runs on.

The data posture behind the manipulation policies humanoids deploy, and the reason teleop and simulation data are bridges, not replacements.

← Sim-to-Real by Domain Randomization Action Chunking and the Two Systems →

PYTHON · NUMPY

The Bridge, Not From Scratch

Fit a robot action head on pretrained features vs raw inputs, from the same 10 demonstrations.

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