April Glove is shared for research and evaluation under the April Glove Data Access Terms (LICENSE.md). Requests are reviewed by hand; please tell us who you are and what you plan to use the data for.

April Robotics

April Glove

50 hours of egocentric human bimanual manipulation with mocap based 3D hand poses and 122-taxel tactile sensing per hand, synchronized with head stereo and wrist video.

20 s of an ironing take: the head camera with both hands' mocap fused keypoints projected into it, both wrist cameras, and below them a 3D mesh of each hand and what the gloves' 122 pressure sensors per hand feel, all on one clock.

Continuous recordings of skilled manual work, from about 10 minutes to over 3 hours each, seen from the worker's point of view and measured on their hands. Every recording puts on one clock:

what the person sees: a head-mounted stereo camera and a camera on each wrist;
what the hands do: 21 keypoints per hand, placed in the room by the head and wrist poses;
what the hands feel: 122 pressure sensors per glove, and the gloves' IMUs;
what the task is: subtask segments in plain language.

All of it is production work recorded during regular shifts at real workplaces, above all a garment manufacturer: ironing, folding, sewing and finishing garments; and at other workplaces, sorting eggs and stocking shelves.

At a glance

Recordings	Hours	Per recording
41	50	10–197 min

Stream	Rate	Detail
Head stereo camera	30 fps	1920×1080, H.265
Wrist cameras	30 fps	1120×1080, H.264
Glove pressure	≈120 Hz	122 sensors per hand, raw counts
Glove IMUs	fingers ≈60 Hz, cuff ≈120 Hz	raw, in each sensor's frame
Head IMU	≈470 Hz	raw, in the sensor's frame
Poses	head ≈30 Hz, wrists ≈90 Hz, keypoints ≈60 Hz	fused by April Robotics' pipeline
Labels	per segment, on change	subtasks, health of every stream

Tasks

Workplace	Tasks	Recordings	Hours
a garment manufacturer	ironing and folding garments, ironing reflective tape onto safety vests, sewing, cutting materials and packing boxes	35	45.5
other workplaces	sorting and packing eggs, stocking shelves	6	4.4

Getting started

Repository layout

LICENSE.md                            the access terms
index.parquet                         one row per recording; the table the viewer shows
clean_intervals.parquet               one row per clean stretch of a recording
assets/pressure_layout.json           where each pressure taxel sits on the hand
recordings/<session-id>/
  recording.mcap                      every time series of the recording
  visualization.mcap                  hands and pressure drawn for Foxglove (see Viewing)
  metadata.json                       everything else about it
  thumbnail.jpg                       head camera 0 at four points of the recording
assets/sample.mp4                     the clip above
assets/foxglove_layout.json           a Foxglove layout for both files
assets/april-logo.png                 the April Robotics logo above

index.parquet lists each recording with its title, a thumbnail (head camera 0 at 10, 35, 60 and 85% of the recording, which the viewer shows inline), its duration in minutes, size in GB, quality, what it lacks that most have and path. Its numbers are rounded for reading; metadata.json has them exact. Filter the index in the viewer, or download it (about 1.5 MB; clean_intervals.parquet is about 100 KB), then download only the recordings you need.

Loading a recording

from huggingface_hub import hf_hub_download
from mcap.reader import make_reader
from mcap_protobuf.decoder import DecoderFactory  # pip install huggingface_hub mcap mcap-protobuf-support scipy numpy
from scipy.spatial.transform import Rotation
import numpy as np

session = "510a9026-4d11-4b9c-b632-c65b95a4b068"
# once access is granted: hf auth login
path = hf_hub_download(
    "aprilrobotics/april-glove", f"recordings/{session}/recording.mcap", repo_type="dataset"
)


def matrix(tf):
    T = np.eye(4)
    T[:3, :3] = Rotation.from_quat([tf.rotation.x, tf.rotation.y, tf.rotation.z, tf.rotation.w]).as_matrix()
    T[:3, 3] = [tf.translation.x, tf.translation.y, tf.translation.z]
    return T


latest = {}  # child frame -> (parent frame, T_parent_child), the newest of each


def to_root(frame):
    T = np.eye(4)
    while frame in latest:
        frame, T_parent_child = latest[frame]
        T = T_parent_child @ T
    return T, frame


with open(path, "rb") as f:
    reader = make_reader(f, decoder_factories=[DecoderFactory()])
    for _, channel, message, proto in reader.iter_decoded_messages(
        topics=["/tf", "/hand/left/keypoints"], log_time_order=True
    ):
        if channel.topic == "/tf":
            # a cuff's pose is held while the cuff is out of view: how long to trust it is up to you
            latest[proto.child_frame_id] = (proto.parent_frame_id, matrix(proto))
            continue
        keypoints = np.array([[p.x, p.y, p.z] for p in proto.points])  # (21, 3), metres
        T_root_wrist, root = to_root(proto.reference_frame)  # left_fused_wrist
        in_world = keypoints @ T_root_wrist[:3, :3].T + T_root_wrist[:3, 3]

Messages are stored in the order the pipeline wrote them, not sorted by time across topics; sort by log_time if you need one timeline, or read one topic at a time.

Viewing

Foxglove shows a recording with its hands, pressure and subtasks:

Download recording.mcap and visualization.mcap of a recording, and assets/foxglove_layout.json.
Open both MCAP files together in Foxglove (Open local file, select both).
Import the layout (Layouts → Import from file).

The layout shows the head camera with the current subtask, the hands in 3D beside it, both wrist cameras, the pressure on each hand, and the health timeline. visualization.mcap holds only pictures for the viewer: the hand skeletons (/vis/hands/fused/{left,right}), the pressure drawn on a hand (/vis/pressure/{left,right}/hand), the subtask as a caption (/vis/annotation/subtask) and frames to view the 3D scene from the head (/vis/tf). Use recording.mcap for numbers.

Reference

Hand keypoints: the 21 points and their order

21 points per hand, in metres, in the wrist frame <side>_fused_wrist; the wrist itself (index 20) is the origin.

Index	Joint	Index	Joint	Index	Joint
0	thumb_tip	7	index_mcp	14	ring_pip
1	thumb_ip	8	middle_tip	15	ring_mcp
2	thumb_mcp	9	middle_dip	16	pinky_tip
3	thumb_cmc	10	middle_pip	17	pinky_dip
4	index_tip	11	middle_mcp	18	pinky_pip
5	index_dip	12	ring_tip	19	pinky_mcp
6	index_pip	13	ring_dip	20	wrist

metadata.json: what each recording's metadata holds

Key	Content
`session_id`	the recording's id
`setting`	`production`: recorded during regular work
`recording`	file, size, first and last `log_time`, duration, the root of the transform tree and the frame pairs it connects, per topic its schema, message count and bytes
`visualization`	the same for `visualization.mcap`, or `null` without one
`sensors`	the pressure's units, grid and taxel names per hand; the IMUs' units, values and finger order; whether the cuff IMU mountings are validated
`health`	the quality columns of the index, exact, and the clean intervals
`annotation`	a title and a summary of the recording, generated with the subtask segments

All recordings are at format_version 9; each MCAP carries an april/dataset metadata record with the recording's session id and that version.

Downloads last month: 38

Total file size:

652 GB