References and Third-Party License Audit
This index centralizes the upstream systems, papers, datasets, models, and license boundaries that define the course. The detailed paper list for a topic remains in that lesson's References section. A link or citation acknowledges a source; it is not permission to copy, rehost, or relicense the linked material.
The repository's controlling records are the third-party notice, license policy, and compatibility matrix.
Core software and platform sources
| Source | Course use and pin |
|---|---|
| Genesis documentation | Simulation API reference; the course pins genesis-world==1.3.3. |
| Genesis source | Version-pinned source used to check scene, entity, solver, control, asset, and camera behavior. |
| Genesis 1.3.3 on PyPI | Package identity for the pinned release. |
| PyTorch HIP semantics | Explains why ROCm devices appear through the torch.cuda compatibility interface. |
| LeRobot source | Dataset, ACT, SmolVLA, training, and policy-loading behavior; course extras pin lerobot==0.6.0. |
| PyAV documentation | Video decode/readback behavior used by the data lessons. |
These packages are installed dependencies, not copied source trees in this repository. Any future copied code requires a file-level license review and a new NOTICE.md entry.
Robotics, data, and learning references
| Topic | Primary reference used by the course |
|---|---|
| Robot kinematics | Modern Robotics and the Foundations of Robotics book |
| YCB objects | Calli et al., “Benchmarking in Manipulation Research: Using the Yale-CMU-Berkeley Object and Model Set,” doi:10.1109/MRA.2015.2448951 |
| Dataset aggregation | DAgger proceedings record, PMLR v15 |
| ACT | Paper record arXiv:2304.13705 and reference implementation |
| SmolVLA | Paper record arXiv:2506.01844 and LeRobot v0.6.0 implementation |
| Foundational domain randomization | arXiv:1703.06907, arXiv:1710.06537, and arXiv:1910.07113 |
| Binomial uncertainty | Wilson's interval record, doi:10.1080/01621459.1927.10502953, and Newcombe's comparison record, doi:10.1214/ss/1009213286 |
L08's acquisition-method survey and the remaining lesson-specific papers stay next to the claims they support. They were retained as external links; no paper figures or full text were copied into the repository.
Audited YCB and Franka sources
Vendored YCB subset
The repository includes only 011_banana, 014_lemon, 018_plum, and 024_bowl under assets/third_party/ycb/.
- Immediate distributor: ManiSkill.
- Original collection: YCB Object and Model Set.
- Reviewed archive: ManiSkill2 YCB archive.
- Applicable recorded terms: CC BY-NC 4.0.
- Local provenance, object list, archive checksum, modification record, legal text, and per-file checksums are stored beside the assets.
The files are byte-for-byte copies of the reviewed source and are not covered by the project MIT License. The noncommercial restriction remains in force.
Franka model
The course does not vendor the source-course Franka bundle. It resolves the model included with pinned Genesis and retains attribution to:
- MuJoCo Menagerie Franka Emika Panda;
- the model-level Apache-2.0 license;
- franka_ros robot description.
“Franka” and “Panda” identify the source robot/model and do not imply endorsement or a trademark grant.
External datasets, checkpoints, and models
The following references are inputs named by lessons, not redistributed course artifacts:
| Material | Pinned reference | Redistribution decision |
|---|---|---|
| L13 fruit-pick dataset | Hugging Face dataset revision 475e8c6... | Link and opt-in workflow only; not vendored |
| L13 SmolVLA checkpoint | Hugging Face model revision cdf720f... | Link and opt-in workflow only; not vendored |
| SmolVLA base | revision c83c316... | Local pinned snapshot for verified training; not rehosted |
| SmolVLM2 base | revision 7b375e1... | Local pinned snapshot for verified training; not rehosted |
| RoboTwin dataset | revision a967b85... | L14 external input; not rehosted |
| LingBot-VLA-v2 6B base | revision 11c703b... | L14 external input; not rehosted |
M5 must review any proposed public course dataset or checkpoint separately, including creator, source data, base model, license, immutable version, size, SHA-256, hosting URL, generation configuration, and citation.
L14 companion sources
L14 is bound to exact external revisions:
- RoboTwin
266f3aa...; - LingBot-VLA-v2 source
951475a...; - XPolicyLab
c37109c...; - Robotwin-radeon-cloud companion
27b2f6f....
The companion remains the runtime source of truth. The course does not modify it, copy its container/runtime into the core installation, or convert the external model and dataset into MIT-licensed materials.
Repository license boundary
| Category | Current treatment |
|---|---|
| Original course code, prose, exercises, and original media | MIT License unless otherwise noted |
| Four vendored YCB objects | CC BY-NC 4.0; outside MIT |
| Installed software dependencies | Their upstream licenses; not vendored by installation metadata alone |
| Linked papers, datasets, checkpoints, and models | Citation/link only unless a separate audit approves redistribution |
| Generated datasets, checkpoints, logs, frames, and videos | Excluded from Git and not approved for redistribution by default |
| Datawhale logo, template animation, name, and trademarks | Outside the MIT grant; local brand-use authorization is not recorded and must be confirmed or the media replaced before release |
Do not paste a repository-wide MIT header onto third-party files or describe the project license as relicensing them.
M4 citation and license audit result
The M4 audit reviewed the bilingual lesson reference sections, public auxiliary pages, NOTICE.md, LICENSE_POLICY.md, the YCB directory records, and the fixed L13/L14 source identities. Its conclusions are:
- no new third-party binary, dataset, checkpoint, paper figure, or model was added by M4.6;
- the only vendored third-party course assets remain the four audited YCB objects;
- Genesis/Franka, LeRobot, papers, and model/data sources remain links or installed dependencies rather than copied source;
- L14 base models, RoboTwin data, and the companion image must not be rehosted by the course;
- generated outputs remain excluded unless M5 performs an artifact-specific audit;
- Datawhale brand/template-media authorization remains an explicit pre-release item.
External-link availability can change after this static audit. A broken link should be updated to the same authoritative source or archived revision; it must not be replaced with an unreviewed mirror.
Contributor checklist
Before adding an external reference or material:
- Prefer an official page, DOI, release tag, or immutable commit.
- State which claim or implementation the source supports.
- Determine whether the repository links, installs, embeds, modifies, or redistributes it.
- Record the applicable license and attribution at the file or directory boundary.
- Add checksum and modification records for vendored assets.
- Keep datasets, checkpoints, and generated media out of Git unless repository policy explicitly approves them.
- Update
NOTICE.mdwhen the disposition changes from link/dependency to copied material.