README.md

metadata

license: cc

Dexdata format SO-101 dataset

Dataset Structure

so101_pick_cube
   ├── videos
   │    ├── so101_20260630_103330_filtered
   │    │    ├── file-000.mp4_top.mp4
   │    │    └── file-000.mp4_wrist.mp4
   │    └── ...
   └── jsonl
        ├── episode_00000.jsonl
        ├── episode_00001.jsonl
        └── ...

Task Description

This dataset contains robot manipulation demonstrations for the task:

• Task: Pick the cube and place it in the plate
• Object variations: Different colored cubes (e.g., orange)

Data Format

Each episode file in jsonl/ contains trajectory data with:

• prompt: Task instruction
• state: Robot state (6D vector)
• action: Robot action (6D vector)
• images_1: Top camera view reference
• images_2: Wrist camera view reference
• extra: Additional metadata (subtask, timestamp, episode_index, cube_color)

Citation

@misc{dm05,
    title  = {{DM0.5}: An Open-World Foundation Model for General-Purpose Embodied Intelligence},
    author = {{Dexmal Team}},
    month  = {July},
    year   = {2026},
    url    = {https://www.dexmal.com/blog/dm0.5/index_en.html}
}