| --- |
| license: cc-by-nc-4.0 |
| language: |
| - en |
| tags: |
| - robotics |
| - lerobot |
| - labutopia |
| - transportbeaker |
| - franka |
| - image-in-parquet |
| --- |
| |
| # LabUtopia-Level3-TransportBeaker |
|
|
| The **LabUtopia Level3 TransportBeaker** demonstration dataset collected in a simulated scientific laboratory. |
|
|
| ## Dataset Summary |
|
|
| | Property | Value | |
| | --- | --- | |
| | Task | Level3 TransportBeaker | |
| | Robot | Single-arm Franka | |
| | Episodes | **100** | |
| | Frames | **86,336** | |
| | Sampling Rate | 60 Hz | |
| | RGB Cameras | 3 × 256 × 256 | |
| | Raw State | 8 dimensions | |
| | Raw Action | 8 dimensions | |
| | Data Storage | Approximately 9.45 GiB, embedded image bytes within Parquet | |
| | Format | LeRobot v3.0 metadata with images in Parquet (no separate MP4 files) | |
|
|
| The task prompt in the current episode metadata is: "Pick up the beaker from the table and place it at the target." |
|
|
| ## Data Structure |
|
|
| The repository stores 100 Parquet demonstration files under data/chunk-000/, one Parquet episode metadata table under meta/episodes/, and the following metadata: |
|
|
| - meta/info.json |
| - meta/tasks.parquet |
| - meta/stats.json |
| - meta/stats_canonical_grip.json |
| - meta/labvla_manifest.json |
| |
| Each frame contains observations from camera_1_rgb, camera_2_rgb, camera_3_rgb, observation.state, action, timestamp, frame_index, episode_index, index, and task_index. |
|
|
| Camera frames are image bytes embedded in the Parquet files. Do not assume an MP4 video folder is available. |
|
|
| ## Validation |
|
|
| Read-only inspection verified all 100 data-file references and 86,336 Parquet rows, consistent with the declared episode and frame counts. The source dataset was also successfully loaded and sampled with LeRobot 0.4.4. |
|
|
| The associated fine-tuned LabVLA checkpoint's norm_stats.json and this dataset's meta/stats_canonical_grip.json are byte-for-byte identical (SHA256 9e9da9599949fe24ad7c241cc63af930962e02faa4cd02d8a4c8096b645a4230). |
| |
| ## Action Semantics and Schema Versioning |
| |
| The 8-dimensional raw action comprises seven arm dimensions and one gripper-width dimension. In the matching LabVLA training pipeline, the arm action is transformed into Delta Action while the gripper width remains absolute. |
| |
| **Schema note:** This dataset's meta/labvla_manifest.json identifies labutopia_level3_transportbeaker_v1, whereas the trained high-performing checkpoint uses labutopia_level3_transportbeaker_v3. This difference is documented rather than silently changed. Consumers of the released model should use the actual checkpoint schema and the matching LabVLA transformation pipeline. |
|
|
| Do not replace canonical gripper normalization with generic Action Statistics without verification. |
|
|
| ## Loading |
|
|
| Use a compatible LeRobot version. Example for LeRobot 0.4.4: |
|
|
| from lerobot.datasets.lerobot_dataset import LeRobotDataset |
| dataset = LeRobotDataset("zjunlp/LabUtopia-Level3-TransportBeaker") |
| print(len(dataset)) # 86336 |
| sample = dataset[0] |
| |
| ## License and Attribution |
|
|
| **CC BY-NC 4.0** applies to the original LabUtopia data assets. Source code has a separate MIT License; the MIT License does not supersede the restrictions on data assets. |
|
|
| Original LabUtopia project: https://github.com/Rui-li023/LabUtopia |
|
|
| This release contains TransportBeaker demonstrations only. It intentionally excludes historical _old copies, other task datasets, backups, temporary scan files, and unrelated experimental results. |
| |
| Please cite the original LabUtopia and LabVLA papers when using this dataset. |
| |
