Find your path

The learning mapwhat to read, and what it unlocks

The dependency map

How the chapters actually connect.

Every phase, left to right, one column each. Every node is a built chapter that links to its page; every thread is a dependency a chapter'smeta.yaml actually declares — not a line we drew by hand. Focus or hover a chapter to light up what it needs and what it unblocks.

The main line runs 0 → 1 → 2 → 4 → 5 — a complete “empty file → real robot” journey on its own. Columns markedoptional · Depth sit off it: self-contained electives you take when the itch strikes. The main line requires none of them.

builds ondirect sequel — reuses the prior chapter's code
requireshard map prerequisite
phase gatePhase-4 prerequisite — Post-Training assumes it
reuses policyoptionally runs an earlier chapter's trained policy
on focus: prerequisites
on focus: what it unblocks
optional · Depthoff the main line — a self-contained elective

Phase 0 · Foundations

Phase 1 · Imitation

Phase 2 · Reinforcement

Phase 3 · Advancedoptional · Depth

Phase 4 · Post-Training

Phase 5 · Practitioner

0.0Your First Robot Policy in a Few Minutes 0.1The Simulation Loop 0.2Bodies, Joints, and MJCF 0.3Spatial Reasoning Without Tears 0.4Teleoperation & Your First Dataset 0.5Seeing Like a Robot (rerun)1.1Behavior Cloning 1.2Data Is the Policy 1.3ACT 1.4Generative Policies I 1.5Generative Policies II 1.6Evaluation Is Hard 1.7Tokens Meet Torques 1.8Tokens Meet Torques 1.9Graduation Bridge I 2.1PPO 2.2SAC and the Off-Policy Bargain 2.34096 Robots at Once 2.4Reward Design Is Programming 2.5Locomotion 2.6Sim-to-Real Intuition Lab I 2.7Sim-to-Real Intuition Lab II 2.8Concepts of ROS, Without ROS 3.1World Models I 3.2World Models II 3.3Build a Physics Engine I 3.4Build a Physics Engine II 3.5Build a Physics Engine III 3.6Full Circle 3.7Datasets at Scale 3.8Reading the Frontier 3.9Plan Through Your Engine 4.1Offline RL Primer 4.2Corrections 4.3RL Post-Training 5.1Patches & Attention 5.2Why Aligned 5.3Control From Pixels 5.4The Production VLA Shape 5.5FAST 5.6LoRA From Scratch 5.7Quantize a Policy by Hand 5.8The Real Loop

The map as a list

Every chapter, by phase, with the threads it declares. This is the same data the diagram draws — readable without JavaScript.

Phase 0 · Foundations
Phase 1 · Imitation
- 1.1Behavior Cloning: The Dumbest Thing That Worksunblocks 3.6 Full Circle, 5.3 Control From Pixels, 5.7 Quantize a Policy by Hand, 5.8 The Real Loop
- 1.2Data Is the Policy
- 1.3ACT: Commit to the Chunkunblocks 5.5 FAST
- 1.4Generative Policies I: Diffusionunblocks 3.1 World Models I, 3.2 World Models II
- 1.5Generative Policies II: Flow Matchingunblocks 5.4 The Production VLA Shape
- 1.6Evaluation Is Hardunblocks 4.2 Corrections, 4.3 RL Post-Training
- 1.7Tokens Meet Torques: The Tiny VLA, Part I (the data)unblocks 5.1 Patches & Attention, 5.2 Why Aligned, 5.5 FAST
- 1.8Tokens Meet Torques: The Tiny VLA, Part II (train it)unblocks 4.2 Corrections, 4.3 RL Post-Training, 5.1 Patches & Attention, 5.3 Control From Pixels, 5.4 The Production VLA Shape, 5.6 LoRA From Scratch
- 1.9Graduation Bridge I: LeRobot for Real
Phase 2 · Reinforcement
- 2.1PPO: The Policy That Acts and Sees the Consequencesunblocks 2.8 Concepts of ROS, Without ROS, 4.2 Corrections, 4.3 RL Post-Training
- 2.2SAC and the Off-Policy Bargainunblocks 4.2 Corrections, 4.3 RL Post-Training
- 2.34096 Robots at Once: PPO on MJX
- 2.4Reward Design Is Programming
- 2.5Locomotion: The Quadruped Walks
- 2.6Sim-to-Real Intuition Lab I: Latency & Noise
- 2.7Sim-to-Real Intuition Lab II: Randomize to Generalize
- 2.8Concepts of ROS, Without ROS: a Pub-Sub Control Runtimeneeds 2.1 PPOunblocks 5.7 Quantize a Policy by Hand
Phase 3 · Advancedoptional · Depth
Off the main line (0 → 1 → 2 → 4 → 5) — self-contained electives; the main line requires none of them.
- 3.1World Models I: Learning the Simulatorneeds 1.4 Generative Policies Iunblocks 3.2 World Models II
- 3.2World Models II: Acting in Imaginationneeds 1.4 Generative Policies I, 3.1 World Models I
- 3.3Build a Physics Engine I: Unconstrained Dynamicsunblocks 3.9 Plan Through Your Engine
- 3.4Build a Physics Engine II: Joints & Constraints
- 3.5Build a Physics Engine III: Contactunblocks 3.9 Plan Through Your Engine
- 3.6Full Circle: Run Your ch1.1 Policy in the Engine You Builtneeds 1.1 Behavior Cloningunblocks 3.9 Plan Through Your Engine
- 3.7Datasets at Scale
- 3.8Reading the Frontierunblocks 5.6 LoRA From Scratch
- 3.9Plan Through Your Engine: Sampling-Based MPC (CEM / MPPI)needs 3.3 Build a Physics Engine I, 3.5 Build a Physics Engine III, 3.6 Full Circle
Phase 4 · Post-Training
- 4.1Offline RL Primer: Beat the Data With Its Own Reward
- 4.2Corrections: Human-in-the-Loop Dataneeds 1.8 Tokens Meet Torques, 2.1 PPO, 2.2 SAC and the Off-Policy Bargain, 1.6 Evaluation Is Hard
- 4.3RL Post-Training: HIL-SERL in Simneeds 1.8 Tokens Meet Torques, 2.1 PPO, 2.2 SAC and the Off-Policy Bargain, 1.6 Evaluation Is Hard
Phase 5 · Practitioner
- 5.1Patches & Attention: A ViT From Scratchneeds 1.7 Tokens Meet Torques, 1.8 Tokens Meet Torquesunblocks 5.2 Why Aligned
- 5.2Why Aligned: Contrastive Vision-Language Pretrainingneeds 5.1 Patches & Attention, 1.7 Tokens Meet Torquesunblocks 5.3 Control From Pixels
- 5.3Control From Pixels: Visuomotor Behavior Cloningneeds 5.2 Why Aligned, 1.1 Behavior Cloning, 1.8 Tokens Meet Torques
- 5.4The Production VLA Shape: Prefix, Suffix, and the Action Expertneeds 1.5 Generative Policies II, 1.8 Tokens Meet Torques
- 5.5FAST: Turning Torques into Tokens (DCT -> Quantize -> BPE)needs 1.3 ACT, 1.7 Tokens Meet Torques
- 5.6LoRA From Scratch: Adapt a Frozen Policyneeds 1.8 Tokens Meet Torques, 3.8 Reading the Frontier
- 5.7Quantize a Policy by Hand: INT8 Is a Scale, Not a Roundingneeds 1.1 Behavior Cloning, 2.8 Concepts of ROS, Without ROS
- 5.8The Real Loop: Teleoperate, Record, Train, Deploy, on a Real Arm's Body (in sim)needs 0.4 Teleoperation & Your First Dataset, 1.1 Behavior Cloning