The FCP method
Three layers run through every chapter, interleaved rather than separated. The Foundation layer states and derives the mathematics in full — if a proof is genuinely out of scope, the theorem is stated precisely and the proof is cited, never waved at. The Conceptual layer gives every hard idea something you can manipulate, because a parameter you have dragged is a parameter you understand. The Practical layer implements the algorithm in Rust, seeded and tested, so the reader ends with working code rather than pseudocode.
The visual language
One encoding scheme runs book-wide, so a color means the same thing in Chapter 20 as it did in Chapter 4. The palette is validated for colorblind separation, lightness banding and contrast in both light and dark themes — not chosen by eye.
| Quantity | Encoding |
|---|---|
| Value function | Sequential blue ramp |
| Policy | Directional arrows, slot-1 hue |
| Reward / TD error / advantage | Diverging blue↔red, gray midpoint |
| Uncertainty | Fan or band, series hue at low alpha |
Every chart also carries a table view, so the numbers are readable without relying on color at all, and every interactive widget degrades to a captioned static figure.
The Rust toolchain
Teaching code lives in the book's own crates rather than depending on third-party RL libraries — building the gym-style abstractions is itself curriculum. External crates are used where they are genuinely best in class:
| Domain | Crate | Why |
|---|---|---|
| Deep learning | burn | Pure Rust, multi-backend, trains natively and runs in the browser via WGPU. |
| Tensors & geometry | ndarray, nalgebra | ML-shaped data and SE(3)/Riccati geometry respectively. |
| Physics | rapier2d / rapier3d | Deterministic option, WASM-compatible — simulations run client-side. |
| Robot models | urdf-rs, k | URDF parsing plus serial-chain forward/inverse kinematics and Jacobians. |
| 3D scenes | bevy + bevy_rapier | WASM-capable 3D for locomotion and manipulation demonstrations. |
| Dashboards | egui + egui_plot | Immediate-mode training dashboards that compile to the browser. |
| Parallelism | rayon | Vectorized environment rollout farms. |
| Optimization | cmaes | Policy search over DMP weights and system-identification fits. |
The standard of evidence
Claims about what reinforcement learning achieves on real robots are graded against Tang et al.'s levels of real-world success, from L0 (simulation only) to L5 (shipping in a product). It is an unusually honest rubric for a field that often reports its best single trial, and this book applies it to its own capstone as strictly as to the literature.