--- license: other license_name: robox-commercial-nonexclusive-v1 task_categories: - robotics tags: - LeRobot - robox - ego-hand - manipulation - human-demonstration - cross-embodiment configs: - config_name: default data_files: data/*/*.parquet --- # RoboXTechnologies/lerobot-dataset ![RoboX LeRobot dataset available on Hugging Face](./banner.png) Egocentric hand-manipulation dataset in **LeRobot v3.0** format, produced by the RoboX export pipeline from real on-device hand-tracking recordings. It is structured for native loading, visualization, and training with LeRobot-compatible tooling ([LeRobot](https://github.com/huggingface/lerobot)). ## What this is This release packages 28 egocentric human demonstrations, spanning 15 tasks across four activity groups (object transfer, organization, assembly, and kitchen activities), as a standard LeRobotDataset v3.0. It opens in the LeRobot visualizer, carries natural language task descriptions, and exposes documented state and action channels with per-frame provenance and imputation masks. Compared with the earlier format-demo release, the `action` channel is now a proper per-frame **delta action** (14 dimensions), the wrist rotation delta is a true relative rotation vector, quaternion orientation is unit-norm and hemisphere-continuous, and the export ships recommended normalization, feature semantics, and a passing structural and numerical validation report. It is ready for experimentation and policy training within the LeRobot ecosystem. ## Read before training: this is human hand motion The `observation.state` and `action` channels are derived from **tracked human hand motion**, not from commands issued to a physical robot. They describe where a human hand was and how it moved between frames. - **No robot retargeting is applied.** `target_embodiment` is null and no robot telemetry is present. Using this to train a policy for deployment on a specific arm (for example SO-ARM101) requires a retargeting and validation step that this dataset does not provide. - **Scale is small.** 28 episodes and 7,730 frames. This suits prototyping, imitation-learning experiments, and cross-embodiment research, not a production policy on its own. - **FPS is declared, not independently verified.** The stated rate is 30 fps. - **State comes from MediaPipe hand tracking** with roughly 3.7 percent of frames imputed. Per-frame imputation flags are in `meta/robox_frame_masks.parquet`. ## Visualize this dataset **Online (LeRobot visualizer):** [Open episode 0 in the LeRobot visualizer](https://lerobot-visualize-dataset.hf.space/RoboXTechnologies/lerobot-dataset/episode_0) Browse other episodes by changing the trailing `episode_0` to any index from 0 to 27. **Locally (with LeRobot installed):** ```bash python -m lerobot.scripts.visualize_dataset \ --repo-id RoboXTechnologies/lerobot-dataset \ --episode-index 0 ``` A tabular preview of the per-frame parquet data also appears through the Hugging Face Dataset Viewer on the dataset page. ## Dataset summary - **Episodes:** 28 - **Frames:** 7730 - **Tasks:** 15 - **Declared FPS:** 30 - **Robot type label:** `robox_ego_hand` (a human hand source, not a physical robot) - **Codebase version:** v3.0 - **State dim:** 15 - **Action dim:** 14 (per-frame delta of state) ## State and action `observation.state` has 15 dimensions describing absolute hand pose. `action` has 14 dimensions describing the per-channel delta to the next frame, where the wrist rotation delta is a 3-D relative rotation vector `rotvec(q[t+1] . inverse(q[t]))` rather than a quaternion subtraction. See `meta/robox_feature_semantics.json` and `meta/robox_training_hints.json` for units and recommended normalization. **observation.state (dim 15)** | Index | Name | Unit | Meaning | | --- | --- | --- | --- | | 0 | `wrist_x` | meters | Wrist position in the egocentric camera frame | | 1 | `wrist_y` | meters | Wrist position in the egocentric camera frame | | 2 | `wrist_z` | meters | Wrist position in the egocentric camera frame | | 3 | `wrist_qx` | unit quaternion | Wrist orientation, unit-norm, hemisphere-continuous | | 4 | `wrist_qy` | unit quaternion | Wrist orientation, unit-norm, hemisphere-continuous | | 5 | `wrist_qz` | unit quaternion | Wrist orientation, unit-norm, hemisphere-continuous | | 6 | `wrist_qw` | unit quaternion | Wrist orientation, unit-norm, hemisphere-continuous | | 7 | `finger_curl_thumb` | normalized 0..1 | Finger curl (0 extended, 1 fully curled) | | 8 | `finger_curl_index` | normalized 0..1 | Finger curl (0 extended, 1 fully curled) | | 9 | `finger_curl_middle` | normalized 0..1 | Finger curl (0 extended, 1 fully curled) | | 10 | `finger_curl_ring` | normalized 0..1 | Finger curl (0 extended, 1 fully curled) | | 11 | `finger_curl_pinky` | normalized 0..1 | Finger curl (0 extended, 1 fully curled) | | 12 | `pinch_thumb_index` | meters | Fingertip pinch distance | | 13 | `pinch_thumb_middle` | meters | Fingertip pinch distance | | 14 | `pinch_thumb_ring` | meters | Fingertip pinch distance | **action (dim 14):** `d_wrist_x`, `d_wrist_y`, `d_wrist_z`, `d_wrist_rotvec_x`, `d_wrist_rotvec_y`, `d_wrist_rotvec_z`, `d_finger_curl_thumb`, `d_finger_curl_index`, `d_finger_curl_middle`, `d_finger_curl_ring`, `d_finger_curl_pinky`, `d_pinch_thumb_index`, `d_pinch_thumb_middle`, `d_pinch_thumb_ring`. **Recommended normalization** (from `meta/robox_training_hints.json`): MEAN_STD for both `observation.state` and `action`, excluding the quaternion indices `[3, 4, 5, 6]` from state mean/std and treating them with IDENTITY quaternion mode. ## Features | Key | dtype | shape | | --- | --- | --- | | `observation.images.front` | video | [480, 640, 3] | | `observation.state` | float32 | [15] | | `action` | float32 | [14] | | `timestamp` | float32 | [1] | | `frame_index` | int64 | [1] | | `episode_index` | int64 | [1] | | `index` | int64 | [1] | | `task_index` | int64 | [1] | ## Validation `meta/robox_validation.json` reports the structural and numerical checks for this release: - Quaternion channels unit-norm (non-unit ratio 0.0). - Frozen (repeated) action frames 3.6 percent, within the 10 percent limit. - Imputed frames 3.67 percent, within the 40 percent limit. - Rotation spikes 0 (max rotation-vector magnitude 0.994 rad, below the 1.0 rad threshold). ## Structure ``` meta/info.json # schema, fps, feature spec (canonical) meta/tasks.parquet # task_index to natural language task meta/episodes/*.parquet # per episode index, length, video timestamps meta/stats.json # per feature statistics meta/robox_export.json # licence, provenance, action semantics (RoboX sidecar) meta/robox_feature_semantics.json # per channel names, units, missing-data policy meta/robox_training_hints.json # recommended normalization and action semantics meta/robox_frame_masks.parquet # per frame imputation flags meta/robox_provenance.parquet # per episode source provenance meta/robox_validation.json # structural and numerical validation report data/chunk-000/*.parquet # per frame state, action, timestamp videos/observation.images.front/chunk-000/*.mp4 # RGB, one shard per chunk ``` Episodes are packed into shared chunk shards (one `.mp4` and one `.parquet` per chunk). Per-episode boundaries are addressed by the row ranges and video timestamps in `meta/episodes`. ## Licence and provenance - **Licence:** `robox-commercial-nonexclusive-v1` - **Commercial use:** allowed - **Exclusivity:** non-exclusive - **Export id:** `robox-p1-20260702` - **Source recordings:** 4 (IDs listed in `meta/robox_export.json`) Full licensing, buyer, and source-recording provenance are in `meta/robox_export.json`. This dataset is distributed under its stated commercial terms, not an open-source licence, despite the Hub `license: other` tag. ## Created with This dataset was created using [LeRobot](https://github.com/huggingface/lerobot) and the RoboX LeRobot v3 export worker.