MicroAGI01: Egocentric Manipulation Dataset

License: See maginoresell

MicroAGI01 is an egocentric RGB-D dataset of human household manipulation with full pose annotations. 676 recordings spanning 137 task types across 14 activity categories.

What's Included Per Recording

• RGB + depth streams
• Camera pose (6DoF)
• Hand poses (3D landmarks)
• Task segmentation with text annotations

Quick Facts

Recordings	676 mcaps (283 cut, 393 uncut)
Task types	137
Container	.mcap
Previews	1 sample .mp4 file

Folder Structure

MicroAGI01/
├── uncut_mcaps/          # Full-length recordings, ≥80% hands validity
├── cut_mcaps/            # Shorter semantic chunks, ≥95% hands validity
├── task_mapping.csv      # Task labels per recording
├── microagi01viewerfoxglove.json
└── LICENSE

Start with uncut_mcaps — full-length recordings with all annotations included.

cut_mcaps contains shorter, semantically-complete segments with stricter hand tracking validity.

Task Categories

Kitchen: kitchen_cooking, kitchen_prep, kitchen_dishes, kitchen_organization, kitchen_dining, kitchen_general

Cleaning: cleaning_general, cleaning_floor

Laundry: laundry

Organization: general_organization, general_household

Rooms: bedroom, bathroom, living_room

Topic Structure

Overview

Meta      /meta
Camera
          /tf_static
          /camera/color/image, /camera/color/info (+ /camera/color/health)
          /camera/depth/image, /camera/depth/info, /camera/depth/unit_of_depth_in_mm
SLAM      /tf/camera (+ .../health, .../state)
Hands     /tf/hands, /hands/left, /hands/right (+ .../health)
IMU       /imu/accel/sample, /imu/gyro/sample
Task      /task (includes task_title)

Descriptions (of relevant topics)

/meta: Information about the mcap, the operator, ... (operator_height_in_m, metadata for general task description)
/tf_static: Any static transforms (Includes transforms between camera, imu, depth and color)
/camera/.../image: JPEG@90 image for color, PNG for depth
/camera/.../info: Parameters for sensor (especially intrinsics)
/camera/depth/unit_of_depth_in_mm: Defines the depth unit conversion. Currently set to 1, meaning the raw pixel values in the depth image are measured directly in millimeters (e.g., a pixel value of 1000 equals 1 meter)
/camera/color/health: Signals bad images which are e.g. too dark, blurry, ...
/tf/camera: Pose of camera (Only valid if a msg on .../health exists with the same timestamp and valid == true, otherwise they should be ignored. Poses are only coherent to poses in the same block of valid poses.)
/tf/camera/health: Signals regions which successful tracking
/tf/hands: Pose of left and right wrist
/hands/...: Positions of Hand keypoints (In wrist frame)
/hands/.../health: Signals whether to trust the hands position or not
/imu/.../sample: Raw imu samples
/task: Description of the current task (includes task_title)

TF-Tree (Across all tf (static) topics)

TF_TREE (RightHanded Coordinate Systems):
world (On the ground; z is up, gravity aligned)
    camera (Center of camera; z is up, x is front)
        # Camera data
        depth (Reference for the depth image; x to the right, y is down)
            accel (Reference for the accel)
            gyro (Reference for the gyro)
            color (Reference for the color image; x to the right, y is down)

        left_wrist (x is in direction from pinky to thumb, z is in direction of arm)
        right_wrist (x is in direction from pinky to thumb, z is in direction of arm)

Download

Everything:

huggingface-cli download MicroAGI-Labs/MicroAGI01 --repo-type dataset --local-dir ./MicroAGI01

Single file:

huggingface-cli download MicroAGI-Labs/MicroAGI01 uncut_mcaps/open-source-06.mcap --repo-type dataset --local-dir ./

Viewing

We use Foxglove. A layout template is included in the repo:

1. Open Foxglove
2. Layout → Import layout → select microagi01viewerfoxglove.json
3. Load any .mcap file

This sets up the 3D view, camera feed, hand validity state transitions, and task annotations panel.

Extracting protobuf

We use our github repo. A script is included in the repo.

Intended Uses

• Policy and skill learning (robotics / VLA)
• Action detection and segmentation
• Hand/pose estimation and grasp analysis
• World-model pre/post training

Attribution

This work uses the MicroAGI01 dataset (MicroAGI, Inc. 2026).

Contact

Questions: info@micro-agi.com

Custom data or derived signals: data@micro-agi.com