| --- |
| license: cc-by-nc-sa-4.0 |
| language: |
| - en |
| tags: |
| - lerobot |
| - robotics |
| - feeding |
| configs: |
| - config_name: episodes |
| default: true |
| data_files: |
| - split: train |
| path: episodes_preview.parquet |
| - config_name: frames |
| data_files: |
| - split: train |
| path: task3_phase2_munich/data/chunk-*/episode_*.parquet |
| --- |
| |
| # Task 3, Phase 2: Spoon Feeding |
|
|
| **Task:** Pick up the spoon, feed the mannequin, and place the spoon back in the bowl. |
|
|
| This dataset contains 177 spoon-feeding episodes, with 124,872 synchronized time steps at 20 Hz and three camera views per episode. Recordings were visually selected from two source disks; unrelated thermal-pad placement and tray-transport recordings are excluded. The source labels indicate 157 successful and 20 failed episodes. Both are retained. |
|
|
| ## Browse the recordings |
|
|
| The default **episodes** view has one row per episode (177 rows). Each row contains the head-camera video, left-wrist video, right-wrist video, task description, success label, frame count, and duration. The videos are referenced from the files already stored in this repository; they are not duplicated in the preview index. |
|
|
| The **frames** view contains the 124,872 original time-step rows with robot states, actions, timestamps, and annotations. LeRobot stores its videos separately, so the frame table does not contain inline video columns. All episodes are assigned to `train`, as recorded in the original LeRobot metadata; there is no held-out validation split. |
|
|
| ## Files and LeRobot usage |
|
|
| The complete **LeRobot v2.1** dataset is in [`task3_phase2_munich/`](task3_phase2_munich). It contains `data/`, `videos/`, and `meta/`. The root-level `episodes_preview.parquet` is a browsing index and is not the training table. |
|
|
| Download the dataset and point a LeRobot version supporting v2.1 at the nested dataset directory: |
|
|
| ```python |
| from pathlib import Path |
| from huggingface_hub import snapshot_download |
| from lerobot.datasets.lerobot_dataset import LeRobotDataset |
| |
| snapshot = snapshot_download( |
| repo_id="ebim-benchmark/task3_phase2", |
| repo_type="dataset", |
| allow_patterns="task3_phase2_munich/**", |
| ) |
| dataset = LeRobotDataset( |
| "ebim-benchmark/task3_phase2", |
| root=Path(snapshot) / "task3_phase2_munich", |
| video_backend="pyav", |
| ) |
| ``` |
|
|
| The preview index pins video references to uploaded revision `0b7882bea6172131fd0db10646a153e7fdbd7e63` for reproducibility. If recordings are replaced in a later revision, regenerate the index to reference those new recordings. |
|
|
| ## Robot data and labels |
|
|
| State has 42 dimensions and action has 17 dimensions, including the original converter configuration's synthetic spine height of 434.0. Joint velocities are not included. The supplied converter's `next.reward` convention (1 on the last frame of every episode) is retained; it does not indicate success. Use `annotation.human.validity` in the frame data, or `success` in the episode index, to distinguish source success and failure labels. |
|
|
| The user-supplied feeding task description overrides stale source task metadata. Source episode identifiers, labels, checksums, and synchronization offsets are recorded in `meta/source_episodes.jsonl`. |
|
|
| ## Videos, synchronization, and camera correction |
|
|
| There are 531 H.264 video files. They were re-encoded at CQ 18 to align with the fixed 20 Hz robot-data time grid. The common timeline includes only times covered by all required streams. Original image dimensions and square pixels are retained. Original MCAP files and checksums are retained separately from this upload. |
|
|
| **The head-camera correction is approximate, not a verified per-device calibration.** It uses left-eye factory parameters from a different ZED Mini, SN17064700, with alpha=0. The recording camera serial number and recorded eye were not independently verified. Uniform focal scaling removes invalid borders, with a top-anchored crop to retain the mannequin mouth. All output pixels sample valid recorded pixels; no generated pixels or anisotropic stretching are used. Wrist-camera geometry is unchanged. |
|
|
| Input/output camera parameters and the reference calibration are included under `task3_phase2_munich/meta/`. The all-white `meta/calibration/head_valid_pixel_mask.png` is a validity-check artifact, not a sample image. The explicit Viewer configurations above exclude it from the displayed dataset. |
|
|
| ## License |
|
|
| CC BY-NC-SA 4.0, as specified for this repository. |
|
|
