DecisionFacts/Physical-AI-UMI-transfer-objects-from-cups

Gripper camera Left

Gripper camera Right

Physical AI – UMI Bottle Cap Opening & Closing (Left Gripper)

A hand-held demonstration dataset of transferring objects from cups, recorded with the Trossen Robotics TRUMI gripper, an implementation of the Universal Manipulation Interface (UMI). The operator picks objects out of cups and transfers them to a new location using the TRUMI grippers. Raw GoPro video was processed with the TRUMI dataset generation pipeline (visual-inertial ORB-SLAM3 + ArUco gripper-width tracking) into training-ready MCAP and Zarr datasets.

• Curated by: DecisionFacts Inc
• Capture hardware: 2 × TRUMI hand-held grippers, each with a GoPro 13 camera (Max Lens Mod fisheye)
• License: CC BY 4.0
• Total size: ~5.3 GB

Dataset Summary

Item	Value
Task	Pick objects out of cups and transfer them
Acting gripper	Both (bimanual)
Grippers / cameras	2 (serials C3535424688550, C3535424692702)
Episodes in packaged dataset	1 (both grippers time-synchronized)
Demonstration clip length	~37 s (4,419 frames @ 119.88 fps)
Video recording	2704 × 2028 @ 119.88 fps, fisheye
SLAM processing rate	59.94 fps (slam_frame_stride = 2), images resized to 960 × 720
IMU rate	200 Hz (GoPro GPMF accelerometer + gyroscope)
SLAM mode	ORB-SLAM3 Monocular-Inertial, relocalized against a pre-built session map
Output formats	MCAP (per-episode), Zarr replay buffer (dataset.zarr.zip)

Note: This is a small pilot dataset (one demonstration session). It is well suited for inspecting the TRUMI/UMI data format, testing pipelines, and prototyping, but is not large enough on its own to train a robust policy. Other tasks recorded in the same session are available at Physical-AI-UMI-bottle-opening-with-left-gripper and Physical-AI-UMI-bottle-opening-with-right-gripper.

Repository Structure

.
├── README.md                     # this dataset card
├── dataset.zarr.zip              # Zarr replay buffer (all episodes, JpegXL-compressed images) – 105 MB
├── dataset_plan.pkl              # pipeline plan: which demos/cameras/grippers form each episode
├── mcap/
│   └── episode_000000.mcap       # time-aligned poses, gripper width, JPEG images, raw IMU – 4.59 GB
├── demos/                        # intermediate per-video pipeline outputs – 594 MB
│   ├── mapping_C3535424688550_2026.09.24_16.40.51.199542_mapping.MP4/
│   ├── gripper_calibration_C3535424688550_2026.09.24_16.42.56.549767_GX010047.MP4/
│   ├── gripper_calibration_C3535424692702_2026.09.24_16.43.42.695867_GX010028.MP4/
│   ├── demo_C3535424688550_2026.09.24_17.24.41.085133_GX010053.MP4/
│   └── demo_C3535424692702_2026.09.24_17.24.36.005058_GX010033.MP4/
└── raw_videos/
    └── gripper_calibration/

Directory names follow the TRUMI convention _.

Contents of each demos/demo_* directory

File	Pipeline stage	Description
raw_video.mp4	00	Original GoPro recording for this demonstration
imu_data.json	01	IMU telemetry (accelerometer, gyroscope) extracted from the GoPro GPMF stream
slam_mask.png	03	Mask hiding the gripper fingers/body from SLAM feature extraction
camera_trajectory.csv	03	Per-frame 6-DoF camera pose from ORB-SLAM3 localization
slam_stdout.txt, slam_stderr.txt	03	ORB-SLAM3 logs (camera intrinsics, frame rates, tracking status)
tag_detection.pkl	04	Per-frame ArUco tag detections used to compute gripper width
detect_aruco_stdout.txt, detect_aruco_stderr.txt	04	ArUco detection logs

The mapping_* directory contains the session map built in stage 02 (used to localize every demo in a shared world frame), and the gripper_calibration_* directories contain the per-gripper open/close range calibration recordings used in stage 05.

Data Fields

Both packaged formats store, per time step and per gripper:

• End-effector position – 3D position (meters) in the SLAM map / ArUco-tag-aligned frame
• End-effector rotation – axis-angle (3D)
• Gripper width – calibrated finger opening (meters), derived from ArUco markers on the fingers
• Demo start / end pose – the episode's first and last end-effector pose
• Camera image – wrist-mounted fisheye RGB frame

MCAP additionally contains the raw IMU samples (accelerometer + gyroscope).

Zarr (dataset.zarr.zip)

Zarr (dataset.zarr.zip)

A single flat, NumPy-backed replay buffer (the UMI ReplayBuffer layout) with JpegXL-compressed images. Arrays live under data/ and are prefixed per gripper (robot0_…, robot1_…) and per camera (camera0_…, camera1_…); episode boundaries are stored in meta/episode_ends. In the TRUMI pipeline, camera/gripper index 0 is assigned to the right gripper and index 1 to the left gripper. Run the snippet below to print the exact array names and shapes.

MCAP (mcap/episode_000000.mcap)

One self-describing, typed MCAP file per episode containing time-aligned robot state, JPEG-compressed camera images, and IMU telemetry. It can be opened directly in Foxglove for visual inspection.

Usage

Download

pip install -U huggingface_hub

# Everything (~5.3 GB)
hf download DecisionFacts/Physical-AI-UMI-transfer-objects-from-cups \
  --repo-type dataset --local-dir ./transfer_objects_cups

# Only the compact Zarr training file (~105 MB)
hf download DecisionFacts/Physical-AI-UMI-transfer-objects-from-cups \
  dataset.zarr.zip --repo-type dataset --local-dir ./transfer_objects_cups

Load the Zarr replay buffer

# pip install "zarr<3" imagecodecs numpy
import zarr
from imagecodecs.numcodecs import register_codecs

register_codecs()  # enables the JpegXL codec used for images

store = zarr.ZipStore("transfer_objects_cups/dataset.zarr.zip", mode="r")
root = zarr.open(store, mode="r")
print(root.tree())                       # list all arrays, shapes, and dtypes

episode_ends = root["meta/episode_ends"][:]
print("episodes:", len(episode_ends))

for key in root["data"].keys():
    print(key, root["data"][key].shape)

Read the MCAP episode

python

# pip install mcap
from mcap.reader import make_reader

with open("transfer_objects_cups/mcap/episode_000000.mcap", "rb") as f:
    reader = make_reader(f)
    summary = reader.get_summary()
    for channel in summary.channels.values():
        schema = summary.schemas[channel.schema_id]
        print(f"{channel.topic:40s} {schema.name} ({channel.message_encoding})")

Using the TRUMI repository:

# 3D plot of the gripper trajectory
uv run python scripts/visualize_trajectory.py \
  -t demos/demo_C3535424688550_2026.09.24_17.20.14.185167_GX010048.MP4/camera_trajectory.csv

# Annotated video of ArUco detections
uv run python scripts/visualize_aruco_video.py \
  -i demos/demo_C3535424688550_2026.09.24_17.20.14.185167_GX010048.MP4 \
  -ci <path/to/intrinsics.json> -sfs 2 -o aruco_overlay.mp4

Dataset Creation

Collection

Following the TRUMI data collection procedure, the session was recorded in this order:

• Mapping video (16:40) – a slow scan of the workspace with one gripper camera to build the SLAM map.
• Gripper calibration videos (16:42–16:43) – one per gripper, fully opening and closing the fingers to record the width range.
• Demonstration (17:24) – both grippers recording simultaneously while the operator picks objects out of cups and transfers them.

The mapping and gripper calibration recordings are shared with the left-gripper dataset from the same session, so both datasets are expressed in the same SLAM world frame.

The raw videos were processed with the TRUMI pipeline (scripts/dataset_generation_pipeline.py, --slam_frame_stride 2), which runs eight stages:

Stage	Script	Purpose
00	00_process_videos.py	Organize raw GoPro MP4s into the demos/ structure
01	01_extract_gopro_imu.py	Extract IMU telemetry to imu_data.json
02	02_create_map.py	Build the ORB-SLAM3 map atlas from the mapping video
03	03_batch_slam.py	Localize every demo video against the map
04	04_detect_aruco.py	Detect and localize ArUco tags
05	05_run_calibrations.py	SLAM-tag and gripper-range calibration
06	06_generate_dataset_plan.py	Match cameras/grippers into episodes → dataset_plan.pkl
07	07_generate_mcap_dataset.py / 07_generate_zarr_dataset.py	Package into MCAP and Zarr

SLAM runs on every second frame (≈60 fps) so that at least three 200 Hz IMU samples fall between consecutive SLAM frames (≈3.34 IMU samples/frame here). Camera intrinsics (Kannala-Brandt fisheye model) are recorded in each slam_stdout.txt.

Intended Uses

• Imitation learning / behavior cloning (e.g. Diffusion Policy) for bottle-cap twisting
• Prototyping and validating UMI/TRUMI data loaders and training pipelines
• Studying visual-inertial SLAM trajectories from hand-held manipulation devices
• Teaching and demonstrating the UMI data format

Limitations

• SLAM-derived poses: end-effector poses come from monocular-inertial SLAM and may contain drift or short tracking gaps; they are not motion-capture ground truth.
• Embodiment gap: data comes from a hand-held gripper, not a robot. Deploying a policy requires a robot whose gripper and camera mounting match the TRUMI geometry.

Citation

If you use this dataset, please cite it together with UMI:

@misc{decisionfacts2026umibottle,
  title        = {Physical AI -- UMI Bottle Cap Opening and Closing with Left Gripper (TRUMI)},
  author       = {{DecisionFacts Inc}},
  credits      = {Deepak Mehta, Sreeram B Unni, Prabhu Raghav, Sriram Gopalan},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/DecisionFacts/Physical-AI-UMI-bottle-opening-with-left-gripper}}
}

@inproceedings{chi2024universal,
  title     = {Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots},
  author    = {Chi, Cheng and Xu, Zhenjia and Pan, Chuer and Cousineau, Eric and Burchfiel, Benjamin and
               Feng, Siyuan and Tedrake, Russ and Song, Shuran},
  booktitle = {Proceedings of Robotics: Science and Systems (RSS)},
  year      = {2024}
}

@article{campos2021orbslam3,
  title   = {ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM},
  author  = {Campos, Carlos and Elvira, Richard and G{\'o}mez Rodr{\'i}guez, Juan J. and
             Montiel, Jos{\'e} M. M. and Tard{\'o}s, Juan D.},
  journal = {IEEE Transactions on Robotics},
  volume  = {37},
  number  = {6},
  pages   = {1874--1890},
  year    = {2021}
}

Acknowledgements

Capture hardware and processing pipeline by Trossen Robotics (TRUMI), building on the Universal Manipulation Interface from Stanford, Columbia, and Toyota Research Institute, and on ORB-SLAM3.

Contact

To discuss access to the full catalog or a custom collection scope, contact info@decisionfacts.io.