20 three-minute first-person clips (60 minutes) of people doing real work, recorded on a head-mounted phone looking down at the hands. Every clip comes with the phone's motion sensors (IMU), aligned to the video frame by frame, a per-frame camera log, per-frame hand keypoints, and dense annotations: a per-hand action timeline (2,261 segments), the tasks in each clip with their outcomes, key events and object state changes. Recorded by five Weply contributors on 23–24 October 2026.

Category	Clips	Minutes	What's in it
Construction	7	21	Interior renovation (plastering, electrical rough-in: running cable, stripping and connecting wire); exterior insulation on scaffolding (cutting, buttering and fixing foam boards)
Carpentry, cabinetmaking or woodwork	8	24	Woodworking shop (sliding table saw, planer, marking out cabinet panels); hand carving (hatchet, pull saw, gouge and mallet, hook knife, power sanding)
Livestock handling	5	15	Leading and holding a cow, tending and feeding calves in their stalls, preparing feed, carrying water, hauling manure

Auto services, egg collection and commercial cleaning are being recorded and will follow as a second wave.

Browsing on Hugging Face

The dataset viewer has two configurations:

previews (shown first): the 20 clips at 720p with everything drawn on: hand keypoints (green right, orange left), the annotated action of each hand, the current task, and the gyroscope over the surrounding 10 seconds. For looking only; use clips for anything else.
clips: the clean clips, one row per clip, with its activity, setting, quality figures, the annotated tasks and the per-hand action timeline (tasks, segments).

from datasets import load_dataset
clips = load_dataset("weply-labs/windmill_ai_egocentric_sample", "clips", split="train")        # clean clips + metadata, tasks and segments
previews = load_dataset("weply-labs/windmill_ai_egocentric_sample", "previews", split="train")  # annotated previews

Sensor files and full annotations are plain CSV and JSON next to the videos (sensors/<clip>/, annotations/).

Clips

Clip	Activity	Setting	Gloves
construction_01	Plastering a wall around electrical boxes with a trowel	indoor, apartment renovation	no
construction_02	Plastering a wall corner beside a door frame with a trowel	indoor, apartment renovation	no
construction_03	Electrical rough-in: running cables into wall-mounted boxes	indoor, apartment renovation	no
construction_04	Electrical rough-in: stripping and connecting wires at wall boxes	indoor, apartment renovation	no
construction_05	Exterior insulation: measuring, fitting and levelling foam boards	outdoor, house facade on scaffolding	yes
construction_06	Exterior insulation: buttering foam boards with adhesive and fixing them to the wall	outdoor, house facade on scaffolding	yes
construction_07	Exterior insulation: buttering foam boards with adhesive	outdoor, house facade on scaffolding	yes
woodwork_01	Cutting live-edge boards on a sliding table saw	indoor, woodworking shop	no
woodwork_02	Machining lumber and marking out cabinet panels with a square	indoor, woodworking shop	no
woodwork_03	Planing boards and cutting them on a sliding table saw	indoor, woodworking shop	no
woodwork_04	Sorting offcuts, marking pieces and cutting strips on a table saw	indoor, woodworking shop	no
woodwork_05	Carving a wooden piece with a gouge, mallet and knife	outdoor carving workshop	yes
woodwork_06	Sawing a block with a pull saw, then hewing it with a hatchet	outdoor carving workshop	yes
woodwork_07	Hollowing a wooden cup with a hook knife	outdoor carving workshop	yes
woodwork_08	Drilling and sanding a wooden cup with power tools	outdoor carving workshop	yes
livestock_01	Preparing mash feed in buckets	farm shed	yes
livestock_02	Hauling manure in a cart and emptying it onto the heap	farmyard	yes
livestock_03	Holding a cow on a lead while it drinks and grazes, then leading it into the barn	farmyard and barn	yes
livestock_04	Tending and feeding calves in their stalls	barn	yes
livestock_05	Filling pails at an outdoor tank and watering cattle in the barn	farmyard and barn	yes

Layout

README.md
preview.jpg                      one frame from each clip
previews/<clip>.mp4              the clip at 720p with hands, actions, task and gyroscope drawn on (for viewing)
manifest.csv, manifest.jsonl     one row per clip: activity, setting, device, camera settings, quality and sync figures
clips/<category>/<clip>.mp4      the video
sensors/<clip>/frames.csv        one row per video frame: host-clock time, exposure, ISO, camera intrinsics
sensors/<clip>/imu.csv           fused device motion at 100 Hz: attitude, rotation rate, user acceleration, gravity
sensors/<clip>/imu_raw.csv       raw accelerometer and gyroscope at 100 Hz each
sensors/<clip>/hands.csv         hand keypoints every frame: up to two hands, 21 joints each, left/right
annotations/segments.jsonl       per-hand action timeline, all clips, one row per segment
annotations/tasks.jsonl          the tasks in each clip, with outcome and success predicate
annotations/<clip>.json          the full labels for one clip

Video

H.264 High, 1920x1080, about 30 fps, 20 Mbit/s, 180 s per clip, no audio, no capture metadata (no location, no device names). Cut from the phone's original HEVC recording and re-encoded once; against the decoded original it measures SSIM 0.975 / PSNR 44 dB. Frame timing is the phone's own and slightly variable (most frames 33.3 ms apart), so read each frame's time from the file or from frames.csv, not from the frame number.

Camera: the iPhone's ultra-wide lens (about 106° horizontal field of view), focus locked, exposure capped at 10 ms to limit motion blur, lens distortion correction on, no stabilisation. Phones: iPhone 12, iPhone 12 Pro Max, iPhone 13 Pro.

Sensors and synchronisation

All three CSVs share one clock: the phone's host clock (t_host_s in frames.csv, timestamp_s in the IMU files), which is the clock the phone stamped both camera frames and IMU samples on. t_clip_s in each file is the same time counted from the clip's first frame. Lines starting with # describe the columns.

frames.csv: every video frame of the clip, in order, with its host-clock time. The time sits near the end of the frame's exposure; the instant the image shows is t_host_s - exposure_s/2 + c, where c is a camera-to-IMU constant of a few milliseconds per phone model (not calibrated for these models).
imu.csv: CoreMotion fused device motion. Attitude quaternion (gravity-aligned, yaw free-running), bias-corrected rotation rate (rad/s), user acceleration with gravity removed (g), gravity direction (g). Axes are the phone's (portrait: x right, y up, z out of the screen).
imu_raw.csv: uncorrected accelerometer (g, including gravity, sensor = a) and gyroscope (rad/s, sensor = g), same axes. The two are not co-sampled: sort per sensor and interpolate.

Each file covers the clip plus half a second either side. Resample from the timestamps, not the row number (IMU rows are about 9.95 ms apart).

How well it lines up, per clip, is in the manifest. Every one of the 108,000 frames is matched to the phone's own frame log (frames_timestamped_pct = 100, residual under 0.86 ms, the mp4's timestamp rounding). As an independent check, image motion was correlated against the gyroscope in each minute of every clip: the best match sits 0–14 ms apart (video_imu_lag_ms_median), which is the camera-to-IMU constant and the bias hand motion puts on that method, and nothing near one frame (33 ms).

Hand keypoints

sensors/<clip>/hands.csv holds 2D hand keypoints for every video frame, from Google's MediaPipe Hand Landmarker (Apache 2.0, float16 model), run on the clips after recording. Per frame there is one row per detected hand (up to two) with handedness (the wearer's left or right hand) and 21 joints (wrist and four joints per finger) as x, y in the 1920x1080 frame (0–1, origin top-left) and z, depth relative to the wrist. A frame with no hand found has one row with hand = -1. frame and t_clip_s match frames.csv, so keypoints line up with the IMU on the same clock.

MediaPipe works on bare hands and misses most gloved ones, so coverage depends on the clip (hands_keypoints_pct and two_hands_keypoints_pct in the manifest):

Clips	Hands	Frames with keypoints
construction_01–04, woodwork_01–04	bare	60–95% (both hands in up to 67%)
woodwork_05	white gloves	27%
construction_05–07, woodwork_06–08, livestock_01–05	gloves	0–7%

Keypoints are model output: left/right is occasionally swapped, and a blurred frame can very rarely yield a false hand. handedness_score gives the classifier's confidence.

Quality figures in the manifest

Measured by the phone over every frame of the clip:

hands_visible_pct: share of frames with hands in view, from the phone's hand signal (hand-pose detector, or arms reaching into the frame). It undercounts gloved hands: on the livestock clips it reads 25–60%, while the annotation (annotation_hands_visible_pct, judged from the images) finds hands visible in 92–99% of the time.
hand_pose_detector_pct: the pose detector alone, which misses most gloves.
camera_angle_ok_pct: frames with the camera 45–80° below the horizon, the target for hand work.
too_dark_pct, brightness_median: frames the phone judged too dark, and median brightness (0–255).

The four carving clips (woodwork_05 to 08) come from an earlier app version that did not log hands, angle or brightness; those fields are empty there. Their video, IMU and frame log are complete.

Annotations

Generated automatically by a vision-language model (GPT-6.1 Sol, high reasoning) from frames sampled every 0.5 s, so boundaries are accurate to about half a second. The model also had each clip's frame timestamps and IMU on the phone's clock. The labels are model output, not human-verified.

segments.jsonl: per-hand action timeline. One row per segment: clip_id, start_s, end_s, arm (left, right or both hands), action (a verb phrase), object, destination, spatial_relation, contribution (advancing, wasteful or idle), progress (0 to 1), hands_visible, hands_wearing (bare, or the gloves worn), notes. 2,261 segments, about 38 per minute.
tasks.jsonl: the distinct tasks in each clip (84 in all), with objects, outcome, success_predicate and completed_at_s. Clips are cut from longer recordings, so a task running at a clip's edge can read unclear.
<clip>.json: everything for one clip: scene objects, timeline, tasks, key events (310), object state changes (176), a scene graph over time, operator mistakes and recoveries, data issues, a one-line summary, alternative phrasings of the instruction, and a short performance review.

Times are seconds from the clip's first frame, the same as t_clip_s in the sensor files.

Contributors

Contributors are identified only by a code (contributor_01 to 05). They recorded with the Weply app and accepted Weply's Terms of Use and Privacy Policy before recording.

Access

A private evaluation sample, shared with invited reviewers. For licensing terms, more hours and the categories still to come, contact Weply Labs.

Downloads last month: 10,474

Total file size:

11.1 GB