---
license: mit
pretty_name: UMI (Universal Manipulation Interface) data
tags:
- robotics
---

# UMI (Universal Manipulation Interface) data

Demonstration data of UMI, collected by the authors: the raw GoPro recordings and the UMI SLAM pipeline session workspaces of the paper's tasks, plus the processed zarr datasets. Deduplicated: the authors' `*_mp4s.zip` packagings and the colleague-made transcodes are not repeated (see Layout).

- Code: https://github.com/real-stanford/universal_manipulation_interface
- Paper / project page: https://arxiv.org/abs/2402.10329

## Layout

`<task>/<name>.zarr.zip`: the processed UMI replay buffer of each task, the format the UMI training code reads (`data/camera0_rgb` 224x224 JPEG-XL frames, `data/robot0_eef_pos`, `data/robot0_eef_rot_axis_angle`, `data/robot0_gripper_width`, `meta/episode_ends`; the bimanual tasks also carry `camera1_*` / `robot1_*`). `cup_in_the_wild.zarr.zip` is the file published at https://real.stanford.edu/umi/data/zarr_datasets/. `cup_in_the_wild/mp4s/<location>/`, `cup_in_lab/mp4s/<date>/`, `dynamic_tossing/mp4s/<date>/`: the raw GoPro recordings (`demo_*.mp4`, `mapping*.mp4`, `gripper_calibration_*.mp4`), i.e. the `raw_videos/` input of the UMI SLAM pipeline. `bimanual_cloth_folding/sessions/<session>/` and `bimanual_dish_washing/sessions/<session>/`: the UMI SLAM pipeline workspace of each bimanual recording session: `demos/<demo>/` with `raw_video.mp4`, `imu_data.json`, `tag_detection.pkl`, `slam_mask.png`, `camera_trajectory.csv` (when SLAM succeeded) and the stage logs, `demos/mapping*/` with `map_atlas.osa` and `tx_slam_tag.json`, `demos/gripper_calibration_*/` with `gripper_range.json`, `calibration/` (some sessions) and `dataset_plan.pkl`. The pipeline's `raw_videos/` directory is omitted: it held only symlinks to `demos/*/raw_video.mp4`. `bimanual_cloth_folding/open_source_slam_reruns/<session>_480/`: the SLAM re-runs of sessions 20231228_cheng and 20231228_zhenjia made for the open-source release (same videos, omitted here; different `camera_trajectory.csv`, `map_atlas.osa`, `dataset_plan.pkl`). The authors' open-source packagings `bimanual_cloth_folding_mp4s.zip` (sessions 20231228_cheng + 20240103_zhenjia) and `bimanual_dish_washing_mp4s.zip` (session 20240109_zhenjia) were byte copies of those sessions' `demos/*/raw_video.mp4` and are not repeated. Every zarr store is one `*.zarr.zip`: an uncompressed zip of the store's files, readable in place with `zarr.open(zarr.storage.ZipStore(path, mode="r"))` or unpacked with `unzip -d <name>.zarr <name>.zarr.zip`.

## Contents

| path | files | GB |
|---|---|---|
| `bimanual_cloth_folding` | 1 | 14.71 |
| `bimanual_cloth_folding/open_source_slam_reruns` | 2225 | 1.08 |
| `bimanual_cloth_folding/sessions` | 11164 | 185.84 |
| `bimanual_dish_washing` | 1 | 22.84 |
| `bimanual_dish_washing/sessions` | 9591 | 218.99 |
| `cup_in_lab` | 1 | 4.49 |
| `cup_in_lab/mp4s` | 350 | 30.23 |
| `cup_in_the_wild` | 1 | 17.99 |
| `cup_in_the_wild/mp4s` | 2932 | 221.43 |
| `dynamic_tossing` | 1 | 8.78 |
| `dynamic_tossing/mp4s` | 298 | 47.39 |
| **total** | 26565 | 773.8 |

## License

MIT, the license of the project's GitHub repository (https://github.com/real-stanford/universal_manipulation_interface).