Start here

How to use zero2robotread like a book, run like a lab

zero2robot is an embodied-AI course where the code is the product. Every chapter is one small, real, readable file you can run — from a bare simulation loop to a policy you train, export, and drive in your browser. No black boxes, no framework to excavate.

You read a chapter like a textbook, then run exactly what you read. The browser demos are the same code you train locally; every number carries its own provenance; and everything on the learner path finishes on a free Colab T4 or a CPU laptop. Forty-three chapters are built and readable today, from foundations through the imitation and RL spines to a from-scratch practitioner's stack and a real-arm graduation — with a from-scratch physics engine waiting off the main line as optional Depth. The only pieces still gated are the leaderboard capstone chapters, which stay closed until the hidden-seed grading server ships.

How to read a chapter

Each chapter page is laid out the same way, top to bottom, so you always know where you are. Read it in order the first time through:

  1. See it work. A live concept-toy at the top shows the one idea the chapter is about — running the real policy or simulation in your browser. It is interactive when JavaScript is on and falls back to a labelled still poster when it is off, so the point survives either way.
  2. Objectives. A short "by the end you can…" plate states exactly what you will be able to do. Come back and check yourself against it.
  3. The reading. The prose column, with the actual code shown in dark panels beside it. Those panels are included by regionstraight from the chapter's runnable file — never pasted — and each carries the sha256 fingerprint CI re-checks, so what you read is provably what runs.
  4. Run it. An honest wall-clock ledger tells you how long the chapter takes on each compute tier. Every time shown is measured, never estimated; tiers nobody has measured yet say so plainly.
  5. Practice. Candidate exercises to work locally. Nothing is auto-graded here and nothing phones home — each exercise gives you a local run command to check yourself.

Predict, then run

Some exercises are predict-then-run. Pick the outcome you expect and commit your prediction first — the measured answer and the run command stay hidden until you do, because predicting after you already know the answer teaches nothing. Your committed prediction is the one signal the site records toward progress.

Concept toys, by hand and by key

The "See it work" toy exposes exactly one control so you can build intuition by poking at it. You can drive every toy from the keyboard, not just the pointer: focus the toy (Tab to it, or click it) and use the arrow keys. In the flagship covariate-shift toy, for example, the arrow keys nudge the block, R resets it, andO sends it out of the region the demonstrations covered — the same "aha" you get by dragging. Screen-reader users get the same in-distribution / out-of-distribution transition announced live.

Progress, search, and help

Your progress lives only in your browser — no account, ever. Reading a chapter marks it read (a ✓ appears next to it in the sidebar), committing a prediction counts toward completion, and the landing page shows an overall and per-phase dashboard with a "continue where you left off" link. Clear your browser storage and it resets; it is never sent anywhere.

Two keyboard shortcuts are worth learning now:

KeyDoes
/ or ⌘ KOpen search — jump to any chapter, heading, objective, or code region.
?Open the keyboard-shortcuts & how-to overlay from anywhere.
EscClose whichever overlay is open.

The honest constraints

Two things are true and stated plainly rather than hidden. First, every number you read is measured, not estimated — each figure traces to a recorded run, and where a measurement is still pending (some laptop wall-clock times) the page says so rather than guessing. Second, everything on the learner path is built to the free-tier floor — it completes on a free Colab T4 or a CPU laptop. Anything that needs more is marked as an optional Scale Lab, never slipped onto the required path.

Start from the first page0.0 · quickstart.pyJump to the flagship chapter →