A HuggingFace LeRobot-compatible dataset containing 763 episodes of egocentric and exocentric videos from indoor environments with corresponding AR pose data.

Dataset Summary

Total Episodes: 763 (Egocentric: 243, Exocentric: 520)
Total Frames: 321,178
Frame Rate: 30.00 FPS
Codebase Version: v2.0

Tasks

The dataset contains two distinct viewpoint tasks:

Task 0 (Egocentric): First-person view videos - 243 episodes (31.8%)
Task 1 (Exocentric): Third-person view videos - 520 episodes (68.2%)

Each episode's data includes a task_index field to distinguish between egocentric and exocentric videos.

Environments

Videos were collected from 4 different indoor rooms:

Room	Episodes	Egocentric	Exocentric
Bedroom	39	14	25
Sandwich	54	41	13
Laundry	436	156	280
Bathroom	234	32	202

Dataset Structure

{
  "episode_index": int,
  "frame_index": int,
  "timestamp": float,
  "observation.state": List[float],  # 7D state (placeholder)
  "action": List[float],             # 7D pose: [quat_x,y,z,w, pos_x,y,z]
  "next.reward": float,
  "next.done": bool,
  "next.success": bool,
  "task_index": int,                 # 0=egocentric, 1=exocentric
  "index": int
}

Action Space

The action field contains 7-dimensional AR pose vectors:

Quaternion (4 values): Camera rotation as [x, y, z, w]
Position (3 values): Camera translation as [x, y, z]

This data comes from smartphone AR tracking during video recording.

Usage

Load with HuggingFace Datasets

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("YOUR_USERNAME/seesawvideos-lerobot")

# Access frames
print(dataset['train'][0])

Filter by Task (Egocentric vs Exocentric)

# Filter egocentric only (task_index = 0)
egocentric = dataset['train'].filter(lambda x: x['task_index'] == 0)

# Filter exocentric only (task_index = 1)
exocentric = dataset['train'].filter(lambda x: x['task_index'] == 1)

Using Episode Metadata

The meta/episodes.csv file contains detailed metadata for filtering:

import pandas as pd

# Load episode metadata
episodes = pd.read_csv("meta/episodes.csv")

# Filter by task and room
ego_laundry = episodes[
    (episodes['task_label'] == 'egocentric') &
    (episodes['room'] == 'Laundry')
]

print(f"Found {len(ego_laundry)} egocentric Laundry episodes")

Using Provided Utilities

from utils import SeesawDatasetFilter

dataset = SeesawDatasetFilter(".")

# Get egocentric episodes
ego_episodes = dataset.get_episodes(task="egocentric")

# Filter by room and duration
bathroom_short = dataset.get_episodes(
    room="Bathroom",
    max_duration=10.0
)

# Get video paths
ego_videos = dataset.get_video_paths(task="egocentric")

Files Included

data/chunk-000/*.parquet - Frame-level data for all episodes
videos/chunk-000/observation.image/*.mp4 - Video files
meta/info.json - Dataset metadata and configuration
meta/episodes.csv - Per-episode metadata with labels and room info
meta/splits.json - Pre-computed train/val/test splits (70/15/15)
utils.py - Helper functions for filtering and loading
example_usage.py - Complete usage examples
FILTERING_GUIDE.md - Quick reference for filtering by task

Statistics

Duration by Task

Egocentric: 12.5 ± 11.2 seconds (range: 1.7 - 82.2s)
Exocentric: 14.7 ± 22.7 seconds (range: 1.5 - 387.6s)

Frames by Task

Egocentric: 375 ± 335 frames (range: 51 - 2,467)
Exocentric: 443 ± 680 frames (range: 46 - 11,629)

Citation

If you use this dataset, please cite:

@dataset{seesawvideos_lerobot_2026,
  title={SeesawVideos LeRobot Dataset},
  author={Your Name},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/datasets/YOUR_USERNAME/seesawvideos-lerobot}}
}

License

MIT License

Additional Resources

LeRobot Documentation
Dataset Repository
See FILTERING_GUIDE.md for detailed filtering instructions
See example_usage.py for complete code examples

Downloads last month: 5