| --- |
| tags: |
| - embodichain |
| - robotics |
| - simulation |
| - neural-dynamics |
| --- |
| |
| # EmbodiChain research datasets |
|
|
| Versioned datasets directly usable by EmbodiChain research workflows. The repository groups releases by research type and method, with stable manifests and explicit per-artifact licenses. [Dataset index](index.json). |
|
|
| ## Neural dynamics |
|
|
| | Dataset | Relative directory | Training transitions | Validation/test trajectories | |
| | --- | --- | ---: | ---: | |
| | NeRD Cartpole / Newton | `neural_dynamics/nerd/cartpole/v1/` | 1,024,000 | 512 / 512 | |
| | NeRD Ant / Newton | `neural_dynamics/nerd/ant/v1/` | 1,024,000 | 512 / 512 | |
|
|
| Each artifact has 10,240 training trajectories of 100 steps, a directly loadable HDF5 manifest and card, and Apache-2.0 `LICENSE`/`NOTICE`. State/contact/action fields and relative shard checksums are defined in its manifest. These data are generated by current Newton Featherstone, with the NeRD reference scene profile, not by the official historical Warp teacher. No training logs are hosted here. |
|
|
| Use `snapshot_download(..., repo_type="dataset")` for the desired artifact path and the EmbodiChain `DynamicsWindowDataset` loader. See each version's README for code and coordinate/physics contracts. The dataset is approximately the repository example's 1M training scale; it is smaller than the paper's stated 10M scale. |
|
|
| ## Adding another research dataset |
|
|
| Create `<research-type>/<method>/<name>/<version>/` with a self-contained manifest, relative data paths, explicit splits, a concise README, checksums, provenance/coordinate schema and a version-specific license. Add its entry to `index.json`. Keep model weights, training histories, simulator binary packages, credentials and machine-specific paths outside the dataset package. |
|
|
