tennyyyin/playworld_human_demo_data

PlayWorld Human Demonstration Data

Human teleoperation demonstrations for robotic manipulation tasks.

Dataset Description

This dataset contains human demonstrations collected via teleoperation for robotic manipulation tasks. Each episode includes:

• Multi-view RGB videos (3 cameras: exterior_1, exterior_2, wrist)
• Pre-encoded video latents (PyTorch .pt files)
• Robot state trajectories (7-DOF: x, y, z, roll, pitch, yaw, gripper)
• Task descriptions and success labels

Dataset Structure

{version}/
├── annotation/
│   ├── train/
│   │   └── {episode_id}.json  # Episode metadata and trajectories
│   └── val/
│       └── {episode_id}.json
├── videos/
│   ├── train/
│   │   └── {episode_id}/
│   │       ├── 0.mp4  # exterior_1 camera
│   │       ├── 1.mp4  # exterior_2 camera
│   │       └── 2.mp4  # wrist camera
│   └── val/
├── latent_videos/
│   ├── train/
│   │   └── {episode_id}/
│   │       ├── 0.pt  # encoded features from camera 0
│   │       ├── 1.pt  # encoded features from camera 1
│   │       └── 2.pt  # encoded features from camera 2
│   └── val/
└── metainfo/
    ├── stat.json
    ├── train_sample.json
    └── val_sample.json

Versions

• v0_final: 290 train + 32 val episodes
• v1_final: 302 train + 33 val episodes

Statistics

• Total Episodes: 657
• Video Resolution: 320x192 (based on typical DROID format)
• Frame Rate: 5 Hz (stored), 15 Hz (raw proprioception)
• State Space: 7-dimensional (6D end-effector pose + 1D gripper)

Tasks

Primary task: "Move object into or out of container (ex: drawer, clothes hamper, plate, trashcan, washer)"

Citation

If you use this dataset, please cite the original data collection work.

License

MIT License