EgoExplorer-100

EgoExplorer-100 is a multimodal egocentric manipulation dataset in LeRobot v3.0 format, with episode-level instructions and time-aligned segment captions.

Dataset overview

• Episodes: 3,130
• Task categories: 10 (large and small cardboard box folding are grouped as Folding a cardboard box)
• Frames: 11,320,870
• Duration: 377,258 seconds (approximately 104.8 hours)
• Video: 1920 × 970, H.264, 30 FPS
• Split: train (0:3130)
• Primary format: LeRobot v3.0 (codebase_version: v3.0)

The available modalities vary by episode. Depending on the recording, an episode can include egocentric video, tracker poses, hand joint positions, tactile measurements, gravity direction, and SMPL-X data. See valid_modalities in each episode's metadata.json for the modalities available in that episode.

Caption annotations

Caption annotations are stored under each episode's episode_meta/episode_XXXXXX/annotations/ directory:

• instruction.json: bilingual Chinese and English task-level instruction, task ID, and episode level.
• segments.parquet: time-aligned segment annotations and captions for the episode.

For example:

episode_meta/episode_000001/annotations/
├── instruction.json
└── segments.parquet

The segment captions are stored separately from the standard LeRobot frame-level Parquet files and should be loaded from episode_meta when needed.

Repository structure

egoexplorer_100/
├── data/                         # LeRobot frame-level Parquet data
├── videos/                       # Encoded egocentric videos
├── meta/
│   ├── info.json                 # Dataset schema, totals, and conventions
│   ├── stats.json                # Dataset statistics
│   ├── tasks.parquet             # Task index
│   └── episodes/                 # LeRobot episode metadata
└── episode_meta/
    └── episode_XXXXXX/
        ├── metadata.json         # Episode timing, task, status, and modalities
        ├── annotations/
        │   ├── instruction.json  # Chinese/English task instruction
        │   └── segments.parquet  # Segment captions and annotations
        ├── calibration/          # Calibration data where available
        ├── quality/              # Quality-control data where available
        └── smplx/                # SMPL-X data where available

Data conventions

• Tracker orientations are unit quaternions in wxyz order.
• Six-dimensional rotations contain the first two columns of the rotation matrix.
• Tracker data is aligned to the 30 Hz video frame timeline.
• The recorded SDK world frame uses +Y as the vertical axis, with gravity direction (0, -1, 0).
• Missing hand or tactile values may be represented as NaN; consult the episode metadata and meta/info.json for coverage and processing details.

Use a LeRobot version compatible with the v3.0 dataset format. Refer to meta/info.json for the complete feature schema and data/video path templates.

Privacy and usage

The recordings contain person views and faces are not redacted. Review meta/info.json and the per-episode metadata for privacy and usage information before accessing or redistributing the data.