pranavsaroha/so100-synth-dataset

So100 Synthetic Robot Dataset

This dataset contains synthetic robot pose estimation data for the LeRobot So100 arm.

Dataset Structure

synth_dataset/
├── data/
│   └── chunk-000/
│       └── episode_000000.parquet  # Robot joint positions and metadata
├── videos/
│   └── chunk-000/
│       └── observation.images.main/
│           └── episode_000000.mp4  # RGB video frames
└── meta/
    ├── episodes_stats.jsonl        # Episode statistics
    ├── episodes.jsonl              # Episode metadata
    ├── info.json                   # Dataset information
    └── tasks.jsonl                 # Task definitions

Dataset Information

• Robot Type: LeRobot So100 arm
• Number of Frames: 348
• Video Format: MP4 (24 FPS)
• Data Format: Parquet files containing joint positions and metadata
• Source: Synthetic video with inferred motor positions from trained pose estimation model

Usage

This dataset can be used for:

• Robot pose estimation training and evaluation
• Joint angle prediction from RGB images
• Synthetic-to-real transfer learning
• Robot manipulation research

Generation Process

1. Input: Synthetic RGB video of So100 arm movements
2. Model: Trained HoRoPose model (DepthNet + Full Network)
3. Inference: Joint angle predictions for each frame
4. Dataset Creation: Combined video frames with predicted joint positions

Citation

If you use this dataset in your research, please cite the original HoRoPose paper and this dataset.

License

[Add your license information here]