First-principles implementations of the ML ideas underneath robot policies: backpropagation, autograd, transformers, reinforcement learning, imitation learning, vision-language-action models, and policy architectures.
This repository owns first-principles model and policy foundations for the Software-to-Robotics-Inference guide. Start with M0 — skill translation, then use the chapter mapping below to build the theory required by later artifacts.
This repository answers “why does the model work?” Runnable serving mechanisms
live in yanizhang-yz/robotics-inference-lab; bounded experiments live in
yanizhang-yz/robotics-experiments; the real-arm project lives in
yanizhang-yz/so-arm101-policy-platform.
Statuses describe committed artifacts, not intended work.
| # | Chapter | Focus | Status | Evidence |
|---|---|---|---|---|
| 0 | Foundations | Feedforward networks, CNNs, RNNs, manual backpropagation, and autograd | Complete | Three training notebooks plus the backpropagation and autograd implementations are committed. |
| 1 | Transformer | Causal attention and a decoder-only transformer from scratch | Complete | The four-stage implementation, causal-attention tests, validation result, and generated sample are committed. |
| 2 | RL core | Value iteration, Q-learning, and REINFORCE | Active | All three exercise skeletons are committed; their learner TODOs remain open. |
| 3 | Bridge to robotics | Behavior cloning, Diffusion Policy, and ACT | Planned | Only the chapter scope is committed. |
| 4 | VLAs | Vision-language-action models and world-model contrasts | Planned | Only the chapter scope is committed. |
| 5 | Policy architectures | Policy data flow and deployment concepts | Planned | Only the chapter scope is committed. |
See the dependency map and artifact gates for the order in which the chapters build on one another. See third-party notices for dataset, exercise-inspiration, and runtime-dependency attribution.
The repository targets Python 3.11 or newer.
python -m pytest