| --- |
| pretty_name: Sigma_0 RoboTwin Training Data |
| task_categories: |
| - robotics |
| tags: |
| - robotwin |
| - imitation-learning |
| - act-gazer |
| - diffusion-policy |
| - hdf5 |
| - zarr |
| size_categories: |
| - n<1K |
| --- |
| |
| # Sigma_0 training datasets |
| |
| **Release complete: all 30 archives uploaded and verified.** |
| |
| Original generated demonstrations and the ACT/DP preprocessed data used for the six-task Sigma_0 models. The full release consists of **30 tar archives**: 12 raw, 12 ACT, and 6 DP. These are different representations of the same demonstrations, not 1500 independent trajectories. |
|
|
| - Code: [Cuixxx/Sigma_0](https://github.com/Cuixxx/Sigma_0/tree/v1.0.0). |
| - Models: [Lucascui97/Sigma_0](https://huggingface.co/Lucascui97/Sigma_0). |
| - Robot configuration: `aloha_agilex`; joint action/state dimension 14. |
| - Tasks: `click_bell`, `place_empty_cup`, `place_container_plate`, `beat_block_hammer`, `lift_pot`, `stack_bowls_two`. |
| - Each task has 50 clean and 50 strong-distractor episodes (600 raw episode files total). Clean and strong are paired scene variants. ACT contains the corresponding 600 preprocessed episode files. Each of six DP Zarr stores contains 50 paired episodes and clean, strong, and gray-mask image arrays. |
|
|
| ## Formats and layout |
|
|
| | Archive location | Content | Extraction location | |
| |---|---|---| |
| | `raw/<condition>/<task>.tar` | Original HDF5, instructions, trajectory metadata, and available videos/visualizations | `data/<condition>/<task>/aloha_agilex/` | |
| | `processed/ACT/<condition>/<task>.tar` | ACT HDF5: RGB, qpos, action, auxiliary vision | `XPolicyLab/policy/ACT/processed_data/<condition>/<task>/aloha_agilex-joint/` | |
| | `processed/DP/<task>.tar` | Zarr: clean/strong/gray head-camera images, state, action, episode boundaries | `dp_data/<task>.zarr/` | |
|
|
| Conditions are `gazer_clean` and `gazer_distractor_strong`. Raw RGB is stored in native encoded form. ACT images are resized to 480 x 640 during preprocessing. DP images are stored at 240 x 320; the published high-resolution policies resize them to 480 x 640 in the encoder. Gray-mask inputs are derived from strong observations and foreground masks. All original data files are preserved byte for byte inside the archives. Temporary caches are excluded. |
|
|
| ## Download and restore |
|
|
| Download all archives, or use `allow_patterns` to select a representation/task: |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| snapshot_download( |
| repo_id="Lucascui97/Sigma_0", |
| repo_type="dataset", |
| local_dir="sigma0_training_data", |
| # Optional: allow_patterns=["processed/DP/click_bell.tar", "manifest.json", "README.md"], |
| ) |
| ``` |
|
|
| Extract into the root of your Sigma_0 code checkout (use a checkout with no conflicting local datasets): |
| |
| ```bash |
| find sigma0_training_data/raw sigma0_training_data/processed -name '*.tar' -print0 | |
| while IFS= read -r -d '' archive; do |
| tar -xf "$archive" -C /path/to/Sigma_0 |
| done |
| export DP_GAZER_DATA_ROOT=/path/to/Sigma_0/dp_data |
| ``` |
| |
| ACT preprocessing is already placed at the path used by the policy launchers. To recreate processed data instead, use `XPolicyLab/policy/ACT/detr/process_data.py` and `scripts/gazer/process_dp_gazer_data.py` from the code release. Adapt the documented cluster-specific environment and Slurm settings to your machine. |
| |
| ACT train/validation episode IDs are recorded in each released model's `training_config.json`. DP episode splitting is controlled by the embedded checkpoint config. This repository does not invent an additional train/test split. |
|
|
| ## Integrity |
|
|
| `manifest.json` records archive sizes and SHA-256 digests. `file_manifests/` lists every archived file and its SHA-256 digest. Archive upload checks compare the remote LFS digest with the local archive digest. HDF5 validation checks episode count, completion metadata where available, camera sequence lengths, and ACT action/state dimensions. DP checks validate completion metadata and array lengths against episode boundaries. |
|
|
