| --- |
| license: cc-by-nc-4.0 |
| task_categories: |
| - robotics |
| tags: |
| - lerobot |
| - embodied-ai |
| - manipulation |
| - human-demonstration |
| - stereo-vision |
| - grasping |
| - desktop-manipulation |
| language: |
| - zh |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # TRICOS.AI DeskOrganizationV1 |
|
|
| A 600-episode desktop manipulation dataset recorded from human demonstrations, |
| in LeRobot v3.0 format. |
|
|
| Published by TRICOS.AI, an embodied-AI data infrastructure company. |
|
|
| ## Overview |
|
|
| | Property | Value | |
| |---|---| |
| | Episodes | 600 | |
| | Tasks | 12 language-conditioned tasks (6 objects x left/right box), 50 each | |
| | Frames | 199,675 | |
| | Frame rate | 60 Hz | |
| | Vision | Stereo pair, 1920 x 1200, calibrated and undistorted | |
| | State | Head pose, both wrist poses (position + quaternion), gripper state | |
| | Format | LeRobot v3.0 | |
|
|
| ## Tasks |
|
|
| Twelve instructions pair six object categories (tissue, clear tape, glue stick, |
| pen, bowl, desktop items) with a left or right box. The instruction text is |
| stored in Chinese and is the task label in `meta/tasks.parquet`. |
|
|
| ## Structure |
|
|
| ``` |
| DeskOrganizationV1/ |
| ├── data/chunk-000 .. chunk-005/file-000 .. file-099.parquet # one per episode |
| ├── videos/ |
| │ ├── observation.images.stereo_left/chunk-000 .. chunk-005/*.mp4 |
| │ └── observation.images.stereo_right/chunk-000 .. chunk-005/*.mp4 |
| ├── meta/ |
| │ ├── info.json # version, fps, totals, chunks_size |
| │ ├── tasks.parquet # 12 language instructions |
| │ ├── stats.json # per-feature statistics |
| │ ├── camera.json # per-episode left-camera intrinsics |
| │ └── episodes/ # per-episode length, file locations, timestamps |
| ├── VIDEO_ROTATION.md |
| ├── README.md |
| └── LICENSE |
| ``` |
|
|
| Each chunk directory groups 100 episodes; inside it each episode has its own |
| parquet and its own video file per camera. |
|
|
| ## Features |
|
|
| | Feature | Dim | Description | |
| |---|---|---| |
| | `observation.images.stereo_left` | video | Left RGB stream, 1920x1200 | |
| | `observation.images.stereo_right` | video | Right RGB stream, 1920x1200 | |
| | `observation.head_pose` | 7 | Head pose (xyz + quaternion xyzw) | |
| | `observation.left_hand_pose` | 7 | Left wrist pose | |
| | `observation.right_hand_pose` | 7 | Right wrist pose | |
| | `observation.state` | 21 | Head + both wrists, concatenated | |
| | `observation.gripper` | 2 | Gripper state: 1 = grasping, 0 = open, NaN = unknown | |
| | `timestamp` | 1 | Time within the episode (seconds) | |
| | `frame_index` | 1 | Frame index within the episode | |
| | `episode_index` / `index` / `task_index` | 1 | Episode, global row and task indices | |
|
|
| Pose convention: positions in metres, quaternions in xyzw order, one shared |
| world frame that is not gravity-aligned. |
|
|
| ### Pose origin |
|
|
| `observation.head_pose` is the left camera's optical-centre pose. |
|
|
| The two `*_hand_pose` columns are **not** measured at the IMU: they are the |
| pose of the wrist AprilTag centre, in the canonical tag frame (tag +z points |
| away from the hand). The tracker used for this dataset mounts the 20 mm tag on |
| the outer face of the enclosure, 17.5 mm above the hand-side surface: |
|
|
| ``` |
| p_hand = p_tag + R_tag * (0, 0, -17.51) mm |
| ``` |
|
|
| For the IMU-origin transform, contact TRICOS.AI. |
|
|
| ## Usage |
|
|
| ```python |
| from lerobot.datasets.lerobot_dataset import LeRobotDataset |
| |
| ds = LeRobotDataset("DeskOrganizationV1", root="DeskOrganizationV1") |
| frame = ds[0] |
| ``` |
|
|
| > **Video orientation.** The camera is mounted upside down. Using the videos as |
| > stored needs no change. For 2D policies the left eye is the recommended RGB |
| > input. If you rotate the videos 180 degrees, or change their resolution, apply |
| > the matching transforms to the poses — see [VIDEO_ROTATION.md](VIDEO_ROTATION.md). |
|
|
| ## Quality |
|
|
| - Every frame is checked against the source capture; every video segment is |
| decoded and timestamp-checked. |
| - Vision, poses and gripper state are aligned per frame. |
| - Images are calibrated and undistorted. |
| - Gripper states were reviewed and confirmed for all 600 episodes. |
|
|
| ## Privacy & Consent |
|
|
| All demonstrations were collected with informed consent for public release. The |
| camera is mounted overhead and is not aimed at the operator's face. If you |
| believe a recording should not be public, contact us and we will remove it. |
|
|
| ## License |
|
|
| Released under the Creative Commons Attribution-NonCommercial 4.0 |
| International License (CC BY-NC 4.0). Commercial licensing: contact TRICOS.AI. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{tricos_deskorganizationv1, |
| title = {TRICOS.AI DeskOrganizationV1: A Desktop Manipulation Demonstration Dataset}, |
| author = {{TRICOS.AI}}, |
| year = {2026}, |
| howpublished = {Hugging Face Datasets}, |
| note = {600 episodes, 12 language-conditioned tasks, LeRobot v3.0} |
| } |
| ``` |
|
|
| ## Contact |
|
|
| TRICOS.AI — embodied-AI data infrastructure. |
|
|
