SELF-PACED WORKSHOP

Models and data

Every main-path lab uses explicitly named fixed files. Download once and reuse them across chapters.

Quick reference · No retraining required

Course model inventory

File	Purpose
`walk.onnx`	Fixed walking policy
`terrain-walk.onnx`	Chapter4 fixed terrain walking actor, about0.79 MB
`kick-teacher.onnx`	Base kicking action
`kick-residual.onnx`	Residual kick adjustment
`kick-checkpoint.pt`	Retained kick training checkpoint
`detector.pt`	Fixed visual detector
`detector.onnx`	ONNX detector export
`indoor-detector/detector.pt` / `indoor-detector/detector.onnx`	Retained previous indoor detector for the old 7 cm ball scene
`scene-detector/detector.pt` / `scene-detector/detector.onnx`	Official 10 cm ball and four-surface detector; current default
`official-browser/rollers.onnx`	Official four-passive-wheel actor, 14 servo actions
`official-browser/walking.onnx` / `sitstand.onnx`	Browser-version checks and standing tests; the complete walking task retains its fixed actor

Run the command below from the repository root to verify fixed-model sizes and hashes.

python3 scripts/workshop.py verify-models

These files are included in Git without Git LFS or a separate model account. models/manifest.json records sizes, sources, and SHA-256 identities. Model identity checks guard against accidental same-name replacements. The added detector has a separate models/scene-detector/deployment.json with relative paths and hashes; the original seven files are unchanged.

These weights support Chapters 1–5 and the Chapter 6 baseline. GR00T and Cosmos foundation models are not bundled. Future adaptation requires separate model access under the applicable terms and an isolated environment.

Practice data and your outputs

datasets/scene-sample/ is a small image-and-label practice set for detector inference and short training. It is not enough to establish real-world generalization. Student checkpoints and logs belong in outputs/; do not copy them over the course files in models/.

Sources and usage terms

Robot geometry, community policies, and dependencies retain their source terms. The course-code license does not override third-party asset licenses. Before redistributing changes, read the repository’s attribution and license notices.