Access is reviewed manually by vfrog, usually within 1–2 business days. Approval gives you three 10-minute segments with stereo depth and SLAM (Raw format) and one full demo session in Raw, LeRobot and MCAP formats, plus the schema. The full dataset is licensed commercially.

The demo is provided for evaluation only. You may not redistribute it, use it to train models you ship, or try to identify, contact or track anyone who appears in it. Use is governed by the vfrog Data Licence, version 1.0 (LICENSE.md).

vfrog

vfrog · Egocentric Data Card
Collected, processed and licensed by vfrog

Azimov: Egocentric Dataset

800 hours of head-mounted, first-person recordings of people doing real facility work in commercial buildings: maintenance, electrical and ceiling work, cleaning, painting and plastering, and operating floor machines. The recordings come from parking garages, shopping malls, retail stores, offices, restrooms, commercial kitchens and plant rooms.

Every session is captured on a 6-camera stereo headset. Each one includes 26-joint articulated hand tracking for both hands, 6-DoF head pose and ~1 kHz IMU, plus a per-second action label with the objects involved and a machine-readable QC report.

The data is built for Physical AI: imitation learning, VLA and world-model pre-training, hand-object interaction, action segmentation and egocentric VIO/SLAM.

Access: this repository is gated. Approved users get three blurred 10-minute segments with stereo depth and SLAM (v2, Raw format) and one complete, blurred demo session (v1, about 24 minutes) in all three delivery formats, plus the full schema. The full dataset is licensed commercially and delivered from cloud storage. See Access and licensing.

Demo vs. full dataset: the hours, label and environment figures on this card describe the full collection (statistics as of 2026-11-13). The hosted demo data is three v2 segments of finishing work in parking garages and one v1 session of a worker replacing a ceiling light fixture in a shopping-mall corridor. It does not represent the task mix.

Release: v2.0.0 (2026-11-11) adds three segments with stereo depth and SLAM under v2/ · v1.0.0 (2026-10-09) is kept under v1/ and as the tag v1.0.0 · schema vfrog.ego.artifacts/5 · statistics as of 2026-11-13

New in v2: stereo depth and SLAM

Version 2 adds three 10-minute segments of finishing work in parking garages: painting a curb, plastering a wall and painting a wall with a roller. Each one has the camera, hand, head-pose and IMU streams of a v1 session, plus two new layers:

Stereo depth for every RGB frame: a metric depth map of the left RGB eye, estimated from the rectified RGB stereo pair (depth/).
SLAM: 6-DoF head trajectories in metres with a gravity-aligned z axis, both hands placed in the same frame, gripper-style hand actions, standing positions, per-second quality and, where available, a sparse map (slam/).



`curb-painting` · RGB frame 8,054	`wall-plastering` · RGB frame 4,754	`roller-painting` · RGB frame 11,624

Left RGB eye from the blurred video, with the hand-tracking overlay as rendered in each segment's handtracking/rgb_hands_preview_anon.mp4. Both hands are tracked in all three frames.

Segment	Task	Duration	RGB frames	Hand-tracking rate	Depth frames	SLAM covered	Pose source, seconds (depth / ORB-SLAM3 / headset / none)	Hands placed in the room (L / R)	Size
`curb-painting`	painting a curb	10.1 min	18,160	29.7 / s	18,160 (0 blank)	98.3% in 3 trajectories	480 / 0 / 116 / 10	97% / 98%	2.12 GB (51 files)
`wall-plastering`	plastering a wall	10.1 min	18,244	29.5 / s	18,244 (0 blank)	54.3% in 8 trajectories	200 / 0 / 130 / 278	48% / 52%	2.19 GB (57 files)
`roller-painting`	painting a wall with a roller	10.2 min	18,280	24.3 / s	18,280 (1 blank)	78.5% in 5 trajectories	100 / 20 / 359 / 131	78% / 78%	2.07 GB (54 files)

SLAM covered: the share of seconds with a pose in one of the merged trajectories (slam/trajectory_mNN.tum).
Pose source: seconds by the source chosen for that moment. depth is ORB-SLAM3 RGB-D + IMU on the stereo depth, ORB-SLAM3 is stereo-inertial on the tracking cameras, headset is the headset's own tracker (not vfrog SLAM) and none is lost, dark or not worn.
Hands placed in the room: the share of tracked hand samples that fall inside a trajectory, from slam/handpaths.json.



`depth/depth_preview_anon.mp4`: the left eye beside its depth map, coloured from 0.25 m (red) to 4 m (purple)	`slam/slam_sidebyside.mp4`: the left eye, the top-down trajectory, head height and the per-second pose source

Both stills are from curb-painting at t = 4:30.

The v1 demo session at a glance

The v1 demo, azimov-demo, is one continuous 24-minute recording of a worker replacing a ceiling light fixture in a shopping-mall corridor. The worker carries and climbs a ladder, unscrews the cover and fixture with a screwdriver, swaps the bulb and reassembles the fixture.



`hold` · light cover, screwdriver	`take` · light cover, screwdriver	`twist_release` · screwdriver, screw, light fixture

Left RGB eye from the blurred demo video. The overlays are the session's own hand_2d.csv keypoints for that frame (blue left hand, orange right hand), drawn without smoothing. The captions are the action_labels.csv row for that second.

Duration	RGB frames	Hands detected	Hand joints	Action classes	Raw size
24.0 min (1,442.5 s, 13 spans)	43,191 at 30 fps	98.5% of frames	26 × 2, 2D + 3D	29 in 1,440 labels	4.9 GB (23 files)

Stream	Rate	Format
Stereo RGB video	30 fps	HEVC, 2 × 2328 × 1748
Stereo grayscale tracking	~60 fps	HEVC, 2 × 640 × 480
Stereo grayscale ctrl	~60 fps	HEVC, 2 × 640 × 480
Head pose (6-DoF)	30 Hz	CSV, m + quaternion
Accelerometer / gyroscope	1,014 Hz	CSV, m/s² · rad/s
Articulated hand tracking	15.6 Hz	CSV, 26 joints × 2
2D / 3D hand keypoints	per RGB frame	CSV, normalised px · mm
Action labels	1 Hz	CSV, 29 classes
Calibration, sync, QC	—	JSON / CSV

Dataset details


Total duration	791.8 h recorded, 600.1 h fully processed (synced, QC'd, action-indexed), as of 2026-11-13
Sessions	1,393 (1,043 fully processed). Mean 34 min, median 26 min, longest 2 h 36 min
Perspective	Egocentric, head-mounted, hands-free (controllers not used)
Cameras per session	6: RGB stereo pair, tracking stereo pair, ctrl stereo pair
RGB	2 × 2328 × 1748 fisheye, 30 Hz, HEVC, ~6 cm baseline
Tracking / ctrl	2 × 640 × 480 each, ~60 Hz, HEVC
Hands	26 joints × 2 hands (OpenXR layout), world frame, median ~17 Hz native (6–27 Hz). 2D and 3D keypoints are derived per RGB frame
Head pose	6-DoF, ~30 Hz, timestamp-matched to RGB frames
IMU	Accelerometer and gyroscope, ~1,008 Hz, with per-device calibration
Annotations	Per-second action label + free-text objects + confidence. 598.8 h labelled (76% of recorded time). English
Action vocabulary	115 labels in use. 38 of them are curated fine-grained manipulation labels
Hand visibility	98.5% of time with at least one tracked hand (duration-weighted, 1,055 sessions)
Headsets	5 devices (EgoSense E6)
Wearers	10 (as of 2026-10-09)
Collection	Uploaded 2026-08-17 to 2026-11-13, Qatar. Mostly night-shift work
Size	6.22 TB raw as uploaded, plus processed derivatives (~10.5 GB per recorded hour in total)
Depth and SLAM	Stereo depth maps and SLAM trajectories ship with the three v2 segments. Every session includes calibrated RGB stereo, so you can compute depth yourself
Audio / body pose	Not included. Audio is never retained
Ongoing collection	~230 h in the last 30 days (5 headsets). 10,000 h projected over the next 3 months

size_categories in the metadata counts sessions (1,393), not frames or hours.

How it compares

	This dataset	Ego4D	EgoDex	Egocentric-10K
Hours	800 (growing ~230 h/month)	3,670	829	10,000
Setting	Commercial facility work	Daily life	Tabletop manipulation	Factory work
Cameras	6 (RGB stereo + 2 mono stereo pairs)	Mostly 1 RGB	1 RGB	1 RGB fisheye
Articulated hands	26 joints × 2, every frame	Subset only	25 joints (ARKit)	—
Head pose / IMU	6-DoF / ~1 kHz	Subset only	6-DoF / —	—
Dense action labels	Per second, 100% of processed time	Subset only	Per task, natural language	—
Commercial licence	Yes (paid)	Research only	Research only	Yes (Apache-2.0)

What sets this dataset apart is dense multimodal coverage of real, unscripted work. Every processed hour has synchronised stereo video, both hands, head pose, IMU and a label. Most comparable sets offer only some of these, and only for part of the data.

Offerings

Tier	Contents	Typical use
Raw	All 6 camera streams, per-frame metadata, calibration, head pose, articulated hands, IMU, QC report	VIO/SLAM, stereo depth, custom annotation
Annotated	Raw + per-second action labels and objects, derived 2D/3D hand keypoints, motion signals, spans	Imitation learning, action segmentation, VLA pre-training
Curated subsets / clips	Clips cut frame-exactly by action label or object search, filtered by QC thresholds, to your hour target	Targeted fine-tuning, evaluation sets

Licences are non-exclusive by default. Exclusive licences for subsets are available. Contact us for pricing.

Dataset structure

The repository has one folder per release. v1/ holds the demo session azimov-demo (24 minutes of replacing a ceiling light fixture) in all three delivery formats, organised format first, so you can download only the format you use. v2/ holds the three v2 segments in the Raw format, with depth and SLAM. Everything is cut from the same blurred source.

vfrogAI/azimov/
├── README.md
├── LICENSE.md
├── v1/
│   ├── manifests/
│   │   ├── sessions.jsonl             # one record per session (see "Manifests")
│   │   ├── files.csv                  # every file in v1/ with its size, rows / frames / messages
│   │   └── checksums.sha256           # SHA-256 of every file in v1/
│   ├── raw/                           # Format 1: MP4 + JSON + CSV
│   │   └── azimov-demo/
│   │       ├── rgb.mp4                # RGB stereo, side-by-side (2 × 2328 × 1748 → 4656 × 1748), 30 Hz, HEVC
│   │       ├── ctrl.mp4               # ctrl stereo, side-by-side (2 × 640 × 480 → 1280 × 480), ~60 Hz, HEVC
│   │       ├── tracking.mp4           # tracking stereo, side-by-side (1280 × 480), ~60 Hz, HEVC
│   │       ├── rgb_metainfo.csv       # per-frame timestamps, exposure, gain
│   │       ├── ctrl_metainfo.csv
│   │       ├── tracking_metainfo.csv
│   │       ├── camera_params_rgb.json # factory intrinsics + extrinsics, left/right
│   │       ├── camera_params_ctrl.json
│   │       ├── camera_params_tracking.json
│   │       ├── rgb_stereo_calibration.json  # refined Kannala-Brandt RGB stereo calibration
│   │       ├── imu_calibration.json   # IMU bias, scale, non-orthogonality, time alignment, noise
│   │       ├── head_pose.csv          # 6-DoF head pose
│   │       ├── hand_tracking.csv      # 26-joint articulated hands, both hands
│   │       ├── accel.csv              # accelerometer
│   │       ├── gyro.csv               # gyroscope
│   │       ├── action_labels.csv      # one action label per second, with objects
│   │       ├── hand_2d.csv            # 26 2D keypoints per hand per RGB frame
│   │       ├── hand_3d.csv            # 26 3D keypoints per hand per RGB frame (mm, head frame)
│   │       ├── frame_signals.csv      # per-frame motion and hand-activity signals
│   │       ├── spans.csv              # ~120 s segmentation
│   │       ├── sync_anchor.json       # session clock anchor
│   │       ├── sync_manifest.csv      # RGB frame → device / wall clock
│   │       └── qc_report.json         # per-stream QC
│   ├── lerobot/                       # Format 2: LeRobot v3.0 dataset root
│   │   ├── meta/
│   │   │   ├── info.json              # features, fps, paths
│   │   │   ├── stats.json
│   │   │   ├── tasks.parquet
│   │   │   ├── episodes/chunk-000/file-000.parquet
│   │   │   ├── calibration.json       # camera intrinsics scaled to the exported resolution
│   │   │   ├── qc.parquet             # per-episode hand / head-pose coverage
│   │   │   └── provenance.json
│   │   ├── data/chunk-000/file-000.parquet
│   │   └── videos/
│   │       ├── observation.images.head_left/chunk-000/file-000.mp4
│   │       └── observation.images.head_right/chunk-000/file-000.mp4
│   └── mcap/                          # Format 3: MCAP
│       └── azimov-demo.mcap
└── v2/
    ├── manifests/                     # sessions.jsonl, files.csv, checksums.sha256 for v2/
    └── raw/
        ├── curb-painting/
        ├── wall-plastering/
        └── roller-painting/
            ├── rgb.mp4, ctrl.mp4, tracking.mp4 and their *_metainfo.csv
            ├── camera_params_*.json, imu_calibration.json
            ├── head_pose.csv, hand_tracking.csv, accel.csv, gyro.csv
            ├── segment.json           # how the segment was cut from its recording
            ├── depth/
            │   ├── rgb_depth.mp4      # metric depth per RGB frame, 10-bit HEVC (a data store, not for viewing)
            │   ├── rgb_depth.json     # decoding formula, rectification, model and QA
            │   └── depth_preview_anon.mp4         # preview: left eye beside its colour depth map
            ├── handtracking/
            │   └── rgb_hands_preview_anon.mp4     # preview: RGB with the hand-tracking overlay
            └── slam/
                ├── README.md          # how this segment's SLAM was made, and its result
                ├── trajectory_mNN.tum # merged trajectories: use these
                ├── trajectory_{d,s,h,c}NN*.tum    # per-source pieces
                ├── merge.json, quality.json, light.csv, status.json, slam_run.json
                ├── handpaths.parquet/json, actions.parquet/json, stances.parquet/json
                ├── map_sNN_pMM.ply    # sparse map, where an ORB-SLAM3 tracking-camera piece exists
                └── slam_sidebyside.mp4            # preview of the SLAM result

Format 1: Raw (MP4 + JSON + CSV)

Every stream exactly as recorded, plus calibration, derived keypoints and QC, with one folder per session. Each file is documented field by field under Data fields.

Format 2: LeRobot v3.0

A dataset that lerobot.datasets.LeRobotDataset loads directly. Each session is one episode at 30 fps on the RGB clock.

Feature	Shape	Contents
`observation.images.head_left` / `head_right`	video 480 × 640 × 3	The two RGB eyes, split from the stereo pair, fisheye kept, AV1
`observation.state.head_pose`	7	World `x y z qx qy qz qw`
`observation.state.hand_left` / `hand_right`	182	26 OpenXR joints × (`x y z qx qy qz qw`), world frame
`observation.state.hand_left_21` / `hand_right_21`	63	21 MANO-order keypoints, head frame
`observation.state.wrist_left` / `wrist_right`	7	Wrist pose, head frame
`observation.state.hand_valid`	2	Left / right hand tracked this frame
`observation.imu`	6	Mean accel xyz and gyro xyz over the frame interval
`observation.capture_time_ns`	1	True UTC capture time of the frame
`action`	140	Next-frame wrist poses + 21 keypoints for both hands, in the current head frame
`action_valid`	2	Both frames tracked, per hand
`language_persistent`	—	Action-label runs as timestamped `subtask` entries

The episode task is the session's task, e.g. replace a ceiling light fixture. This format does not include the ctrl and tracking cameras; use Raw or MCAP for those.

Format 3: MCAP

One file per session, which opens in Foxglove and works with ROS 2 / MCAP tooling. Every message is stamped on the session's UTC clock (mid_exposure_utc_ns). Video packets are the original HEVC streams, copied without re-encoding.

Topic	Schema	Contents
`/rgb/video`, `/ctrl/video`, `/tracking/video`	`foxglove.CompressedVideo` (h265)	Side-by-side stereo, left eye in the left half
`/rgb/left/calibration`, `/rgb/right/calibration`	`foxglove.CameraCalibration`	Refined Kannala-Brandt intrinsics per RGB eye
`/tf`, `/tf_static`	`foxglove.FrameTransforms`	`world → head` per frame; `head → {rgb,ctrl,tracking}_{left,right}` extrinsics
`/head/pose`	`foxglove.PoseInFrame`	6-DoF head pose, world frame
`/hands/{left,right}/joints`	`foxglove.PosesInFrame`	26 joint poses, world frame, when the hand is tracked
`/hands/{left,right}/keypoints_2d`	`foxglove.ImageAnnotations`	26 keypoints in left-RGB-eye pixels, for overlay on `/rgb/video`
`/hands/{left,right}/keypoints_3d`	JSON	26 keypoints, mm, head frame
`/imu/accel`, `/imu/gyro`	JSON	`x, y, z` in m/s² / rad/s, ~1,008 Hz
`/action_label`	JSON	`label`, `objects`, `confidence`, once per second
`/signals`	JSON	Per-frame motion and hand-activity signals

The file also carries a session metadata record, plus these attachments: qc_report.json, imu_calibration.json, rgb_stereo_calibration.json, camera_params_*.json and sync_anchor.json.

Manifests

Each release folder has its own manifests/ (v1/manifests/, v2/manifests/), and the paths in it are relative to that folder. sessions.jsonl has one JSON record per session:

Field	Description
`session_id`	Public session name, e.g. `azimov-demo`
`date`	Recording day (UTC, from `mid_exposure_utc_ns`)
`schema`	Artifact schema version, e.g. `vfrog.ego.artifacts/5`
`duration_s`	From `qc_report.json`
`rgb_frames`	From `sync_anchor.json`
`device_uid`	Headset identifier (from `imu_calibration.json`)
`modalities`	Flags, e.g. `{"rgb": true, "ctrl": true, "tracking": true, "hands": true, "head_pose": true, "imu": true, "action_labels": true, "depth": false, "audio": false}`
`qc`	`sync_quality`, `hand_active_ratio`, per-stream `status` and `gaps`
`paths`, `size_bytes`	Location and size of the session in each format (`raw`, `lerobot`, `mcap`)
`provenance_id`	Provenance Record / Copy Identifier, filled in at delivery

v2 records have no schema or qc, because the segments have no QC report yet. They add slam.covered_share and slam.hands_in_room, and their modalities mark depth and slam. Their duration_s and rgb_frames come from rgb_metainfo.csv.

files.csv lists every file in its release folder with its size and its count: rows for CSV and Parquet tables (header excluded), frames for MP4 videos, and messages for MCAP files.

checksums.sha256 uses the standard sha256sum format, so you can check a download from inside a release folder, for example cd v1 && sha256sum -c manifests/checksums.sha256.

File inventory (v1 demo)

Counts come from v1/manifests/files.csv, which is generated from the files themselves. Video frames are counted from the stream's packets, not from container metadata.

Folder	Files	Size	Video frames	Table rows	MCAP messages
`v1/raw/azimov-demo/` (Raw)	23	4.90 GB	215,115	3,366,194	—
`v1/lerobot/` (LeRobot)	10	929.9 MB	86,382	43,194	—
`v1/mcap/` (MCAP)	1	4.83 GB	—	—	3,373,313

Raw: v1/raw/azimov-demo/

File	Size	Count	Note
`accel.csv`	66.6 MB	1,469,094 rows
`action_labels.csv`	49.7 KB	1,440 rows	one per second
`camera_params_ctrl.json`	1.0 KB	—
`camera_params_rgb.json`	1.0 KB	—
`camera_params_tracking.json`	1.0 KB	—
`ctrl.mp4`	673.3 MB	85,984 frames
`ctrl_metainfo.csv`	6.3 MB	85,984 rows
`frame_signals.csv`	4.7 MB	43,406 rows	one per RGB frame, plus 215 trailing rows after the last frame
`gyro.csv`	72.9 MB	1,469,089 rows
`hand_2d.csv`	11.2 MB	29,580 rows	one per visible hand per RGB frame
`hand_3d.csv`	11.0 MB	29,648 rows	one per visible hand per RGB frame
`hand_tracking.csv`	88.0 MB	22,518 rows	QC keeps 22,517 after dropping 1 clock-rollback row
`head_pose.csv`	3.6 MB	43,100 rows
`imu_calibration.json`	1.2 KB	—
`qc_report.json`	2.9 KB	—
`rgb.mp4`	3.20 GB	43,191 frames
`rgb_metainfo.csv`	3.2 MB	43,191 rows
`rgb_stereo_calibration.json`	1.8 KB	—
`spans.csv`	1.3 KB	13 rows
`sync_anchor.json`	260 B	—
`sync_manifest.csv`	2.0 MB	43,191 rows
`tracking.mp4`	748.0 MB	85,940 frames
`tracking_metainfo.csv`	6.4 MB	85,940 rows

LeRobot: v1/lerobot/

File	Size	Count	Note
`data/chunk-000/file-000.parquet`	50.1 MB	43,191 rows	one per frame
`meta/calibration.json`	6.5 KB	—
`meta/episodes/chunk-000/file-000.parquet`	6.8 KB	1 row	one per episode
`meta/info.json`	28.6 KB	—
`meta/provenance.json`	621 B	—
`meta/qc.parquet`	2.5 KB	1 row
`meta/stats.json`	191.4 KB	—
`meta/tasks.parquet`	2.1 KB	1 row
`videos/observation.images.head_left/chunk-000/file-000.mp4`	438.9 MB	43,191 frames
`videos/observation.images.head_right/chunk-000/file-000.mp4`	440.6 MB	43,191 frames

MCAP: v1/mcap/

File	Size	Count	Note
`azimov-demo.mcap`	4.83 GB	3,373,313 messages

File inventory (v2 segments)

Counts come from v2/manifests/files.csv, which is generated from the files themselves. Video frames are counted from the stream's packets, not from container metadata.

Folder	Files	Size	Video frames	Table rows	MCAP messages
`v2/raw/curb-painting/` (curb-painting)	51	2.12 GB	164,042	1,429,884	—
`v2/raw/roller-painting/` (roller-painting)	54	2.07 GB	164,838	1,395,031	—
`v2/raw/wall-plastering/` (wall-plastering)	57	2.19 GB	164,659	1,396,759	—

v2/raw/curb-painting/ (curb-painting)

File	Size	Count	Note
`accel.csv`	27.7 MB	618,394 rows
`camera_params_ctrl.json`	1.0 KB	—
`camera_params_rgb.json`	1.1 KB	—
`camera_params_tracking.json`	1.0 KB	—
`ctrl.mp4`	309.0 MB	36,561 frames
`ctrl_metainfo.csv`	2.8 MB	36,561 rows
`depth/depth_preview_anon.mp4`	144.1 MB	18,160 frames	preview video for viewing, not data
`depth/rgb_depth.json`	2.7 KB	—
`depth/rgb_depth.mp4`	90.3 MB	18,160 frames	one depth map per RGB frame (10-bit inverse depth, not for viewing)
`gyro.csv`	31.1 MB	618,394 rows
`hand_tracking.csv`	85.9 MB	17,968 rows	one per RGB frame with tracking, over the RGB window ± 0.5 s
`handtracking/rgb_hands_preview_anon.mp4`	321.0 MB	18,160 frames	preview video for viewing, not data
`head_pose.csv`	1.6 MB	18,409 rows
`imu_calibration.json`	1.2 KB	—
`rgb.mp4`	640.4 MB	18,160 frames
`rgb_metainfo.csv`	1.4 MB	18,160 rows
`segment.json`	1.7 KB	—
`slam/actions.json`	631 B	—
`slam/actions.parquet`	3.7 MB	32,301 rows	one per hand per frame with a gripper state
`slam/handpaths.json`	557 B	—
`slam/handpaths.parquet`	21.3 MB	32,339 rows	one per tracked hand sample placed in the room
`slam/light.csv`	12.8 KB	612 rows
`slam/merge.json`	8.3 KB	—
`slam/quality.json`	23.3 KB	—
`slam/README.md`	2.6 KB	—
`slam/slam_run.json`	1.9 KB	—
`slam/slam_sidebyside.mp4`	122.1 MB	18,160 frames	preview video for viewing, not data
`slam/stances.json`	395 B	—
`slam/stances.parquet`	9.0 KB	55 rows
`slam/status.json`	1.6 KB	—
`slam/trajectory_c00.sources.csv`	77 B	3 rows
`slam/trajectory_c00.tum`	414.7 KB	—
`slam/trajectory_c01.sources.csv`	79 B	3 rows
`slam/trajectory_c01.tum`	311.4 KB	—
`slam/trajectory_c02.sources.csv`	99 B	4 rows
`slam/trajectory_c02.tum`	208.9 KB	—
`slam/trajectory_d00.tum`	114.8 KB	—
`slam/trajectory_d01.tum`	803.1 KB	—
`slam/trajectory_d02.tum`	393.9 KB	—
`slam/trajectory_d03.tum`	72.0 KB	—
`slam/trajectory_h00.tum`	165.1 KB	—
`slam/trajectory_h01.tum`	82.9 KB	—
`slam/trajectory_h02.tum`	248.4 KB	—
`slam/trajectory_h03.tum`	247.5 KB	—
`slam/trajectory_h04.tum`	163.7 KB	—
`slam/trajectory_h05.tum`	246.3 KB	—
`slam/trajectory_m00.tum`	129.8 KB	—
`slam/trajectory_m01.tum`	835.0 KB	—
`slam/trajectory_m02.tum`	597.7 KB	—
`tracking.mp4`	308.4 MB	36,681 frames
`tracking_metainfo.csv`	2.8 MB	36,681 rows

v2/raw/roller-painting/ (roller-painting)

File	Size	Count	Note
`accel.csv`	27.8 MB	619,865 rows
`camera_params_ctrl.json`	1.0 KB	—
`camera_params_rgb.json`	1.1 KB	—
`camera_params_tracking.json`	1.1 KB	—
`ctrl.mp4`	299.9 MB	36,810 frames
`ctrl_metainfo.csv`	2.8 MB	36,810 rows
`depth/depth_preview_anon.mp4`	182.7 MB	18,280 frames	preview video for viewing, not data
`depth/rgb_depth.json`	3.0 KB	—
`depth/rgb_depth.mp4`	96.8 MB	18,280 frames	one depth map per RGB frame (10-bit inverse depth, not for viewing)
`gyro.csv`	30.5 MB	619,865 rows
`hand_tracking.csv`	56.6 MB	14,813 rows	one per RGB frame with tracking, over the RGB window ± 0.5 s
`handtracking/rgb_hands_preview_anon.mp4`	282.7 MB	18,280 frames	preview video for viewing, not data
`head_pose.csv`	1.5 MB	18,441 rows
`imu_calibration.json`	1.2 KB	—
`rgb.mp4`	645.0 MB	18,280 frames
`rgb_metainfo.csv`	1.4 MB	18,280 rows
`segment.json`	1.7 KB	—
`slam/actions.json`	632 B	—
`slam/actions.parquet`	1.7 MB	14,783 rows	one per hand per frame with a gripper state
`slam/handpaths.json`	606 B	—
`slam/handpaths.parquet`	9.5 MB	14,880 rows	one per tracked hand sample placed in the room
`slam/light.csv`	12.7 KB	611 rows
`slam/map_s00_p00.ply`	60.3 KB	—
`slam/merge.json`	7.7 KB	—
`slam/quality.json`	24.1 KB	—
`slam/README.md`	2.6 KB	—
`slam/slam_run.json`	1.9 KB	—
`slam/slam_sidebyside.mp4`	127.3 MB	18,280 frames	preview video for viewing, not data
`slam/stances.json`	419 B	—
`slam/stances.parquet`	8.6 KB	48 rows
`slam/status.json`	1.6 KB	—
`slam/trajectory_c00.sources.csv`	49 B	2 rows
`slam/trajectory_c00.tum`	212.7 KB	—
`slam/trajectory_c01.sources.csv`	37 B	1 row
`slam/trajectory_c01.tum`	51.6 KB	—
`slam/trajectory_c02.sources.csv`	39 B	1 row
`slam/trajectory_c02.tum`	52.2 KB	—
`slam/trajectory_c03.sources.csv`	79 B	3 rows
`slam/trajectory_c03.tum`	206.9 KB	—
`slam/trajectory_d00.tum`	54.6 KB	—
`slam/trajectory_d01.tum`	145.5 KB	—
`slam/trajectory_d02.tum`	114.7 KB	—
`slam/trajectory_h00.tum`	664.4 KB	—
`slam/trajectory_h01.tum`	165.8 KB	—
`slam/trajectory_h02.tum`	331.9 KB	—
`slam/trajectory_h03.tum`	83.2 KB	—
`slam/trajectory_m00.tum`	131.1 KB	—
`slam/trajectory_m01.tum`	614.1 KB	—
`slam/trajectory_m02.tum`	132.5 KB	—
`slam/trajectory_m03.tum`	365.1 KB	—
`slam/trajectory_m04.tum`	80.2 KB	—
`slam/trajectory_s00_p00.tum`	193.0 KB	—
`tracking.mp4`	300.1 MB	36,628 frames
`tracking_metainfo.csv`	2.8 MB	36,628 rows

v2/raw/wall-plastering/ (wall-plastering)

File	Size	Count	Note
`accel.csv`	27.6 MB	618,381 rows
`camera_params_ctrl.json`	1.0 KB	—
`camera_params_rgb.json`	1.1 KB	—
`camera_params_tracking.json`	1.0 KB	—
`ctrl.mp4`	309.1 MB	36,724 frames
`ctrl_metainfo.csv`	2.8 MB	36,724 rows
`depth/depth_preview_anon.mp4`	188.7 MB	18,244 frames	preview video for viewing, not data
`depth/rgb_depth.json`	2.7 KB	—
`depth/rgb_depth.mp4`	135.5 MB	18,244 frames	one depth map per RGB frame (10-bit inverse depth, not for viewing)
`gyro.csv`	30.7 MB	618,381 rows
`hand_tracking.csv`	81.9 MB	17,934 rows	one per RGB frame with tracking, over the RGB window ± 0.5 s
`handtracking/rgb_hands_preview_anon.mp4`	312.2 MB	18,244 frames	preview video for viewing, not data
`head_pose.csv`	1.6 MB	18,405 rows
`imu_calibration.json`	1.2 KB	—
`rgb.mp4`	646.7 MB	18,244 frames
`rgb_metainfo.csv`	1.4 MB	18,244 rows
`segment.json`	1.7 KB	—
`slam/actions.json`	631 B	—
`slam/actions.parquet`	1.8 MB	15,689 rows	one per hand per frame with a gripper state
`slam/handpaths.json`	682 B	—
`slam/handpaths.parquet`	10.2 MB	15,631 rows	one per tracked hand sample placed in the room
`slam/light.csv`	13.0 KB	614 rows
`slam/merge.json`	8.0 KB	—
`slam/quality.json`	23.9 KB	—
`slam/README.md`	2.6 KB	—
`slam/slam_run.json`	1.9 KB	—
`slam/slam_sidebyside.mp4`	120.4 MB	18,244 frames	preview video for viewing, not data
`slam/stances.json`	511 B	—
`slam/stances.parquet`	7.2 KB	31 rows
`slam/status.json`	1.6 KB	—
`slam/trajectory_c00.sources.csv`	74 B	3 rows
`slam/trajectory_c00.tum`	407.7 KB	—
`slam/trajectory_c01.sources.csv`	79 B	3 rows
`slam/trajectory_c01.tum`	255.6 KB	—
`slam/trajectory_c02.sources.csv`	59 B	2 rows
`slam/trajectory_c02.tum`	203.4 KB	—
`slam/trajectory_c03.sources.csv`	59 B	2 rows
`slam/trajectory_c03.tum`	103.2 KB	—
`slam/trajectory_d00.tum`	55.6 KB	—
`slam/trajectory_d01.tum`	179.3 KB	—
`slam/trajectory_d02.tum`	64.6 KB	—
`slam/trajectory_d03.tum`	120.9 KB	—
`slam/trajectory_d04.tum`	142.2 KB	—
`slam/trajectory_d05.tum`	92.4 KB	—
`slam/trajectory_h00.tum`	81.8 KB	—
`slam/trajectory_h01.tum`	245.8 KB	—
`slam/trajectory_h02.tum`	161.4 KB	—
`slam/trajectory_m00.tum`	234.7 KB	—
`slam/trajectory_m01.tum`	26.7 KB	—
`slam/trajectory_m02.tum`	26.5 KB	—
`slam/trajectory_m03.tum`	21.7 KB	—
`slam/trajectory_m04.tum`	104.4 KB	—
`slam/trajectory_m05.tum`	287.2 KB	—
`slam/trajectory_m06.tum`	132.6 KB	—
`slam/trajectory_m07.tum`	26.7 KB	—
`tracking.mp4`	309.6 MB	36,715 frames
`tracking_metainfo.csv`	2.8 MB	36,715 rows

Data fields

Pose schema summary

Component	Shape per sample	Type	Frame	Rotation order
Head position	(3,)	float	World (metres, per-session origin)	—
Head orientation	(4,)	float	World	x, y, z, w
Hand joint positions (L/R)	(26, 3)	float	World (same as head)	—
Hand joint orientations (L/R)	(26, 4)	float	World	x, y, z, w
Hand joint radii (L/R)	(26,)	float	—	—
Hand 3D keypoints (L/R)	(26, 3)	int, mm	Head (`+x right, +y up, -z forward`)	—
Hand 2D keypoints (L/R)	(26, 2)	float, normalised	Left RGB image	—
Camera extrinsics (`camera_params_*.json`)	position (3,), rotation (4,)	float	Camera → head	w, x, y, z (inferred, see `rgb_stereo_calibration.json → quality.factory_quaternion_order_inferred`)

Rotation orders differ: head and hand poses use x, y, z, w, but factory camera extrinsics use w, x, y, z.

21-joint compatibility: the 26-joint OpenXR hand is a superset of the common 21-joint (MANO-style) layout. To get 21 joints, keep indices [1, 2, 3, 4, 5, 7, 8, 9, 10, 12, 13, 14, 15, 17, 18, 19, 20, 22, 23, 24, 25]. This drops PALM and the index, middle, ring and little METACARPAL joints.

Video and per-frame camera metadata

Each MP4 holds a stereo pair side by side, with the left eye in the left half. Frame N of each MP4 is the row with frame_index = N in the matching *_metainfo.csv.

frame_index,frame_id,pts_us,exposure_start_utc_ns,exposure_duration_ns,gain,mid_exposure_utc_ns
0,8653,0,1788485854906968827,8888888,776,1788485854911413271

Column	Meaning
`frame_index`	0-based index into the MP4
`frame_id`	Device frame counter
`pts_us`	Presentation timestamp, µs from first frame
`exposure_start_utc_ns`	Exposure start, Unix epoch ns
`exposure_duration_ns`	Exposure time, ns (varies per frame)
`gain`	Sensor gain (varies per frame)
`mid_exposure_utc_ns`	Mid-exposure, Unix epoch ns. This is the reference clock

Use the *_metainfo.csv row count as the frame count. Container-level nb_frames metadata is not reliable for every stream.

Camera calibration

camera_params_{rgb,ctrl,tracking}.json each hold a cameras array with left and right entries. Each entry has:
- width and height
- intrinsics: focalX, focalY, centerX, centerY and radialDistortion. radialDistortion has 8 coefficients; RGB uses only the first 4, and they are all zero for ctrl and tracking.
- extrinsics: position [3] and rotation quaternion [4], relative to the device.
rgb_stereo_calibration.json is a refined RGB stereo calibration, and the file to use for stereo rectification and depth.
- It uses the Kannala-Brandt fisheye model (fx, fy, cx, cy, k[4] per eye) and the OpenCV convention X_right = R_left_to_right · X_left + T_left_to_right_m.
- The baseline is about 6.3 cm.
- It includes held-out epipolar error statistics. In the demo session the median is 0.77 px, against 22.4 px for factory extrinsics used as is.

Head pose: `head_pose.csv`

timestamp_ns,pos_x,pos_y,pos_z,quat_x,quat_y,quat_z,quat_w
1788485854844759052,45.2283,-1.20498,51.8738,-0.118051,0.98657,-0.0508983,0.100764

Head pose is sampled at about 30 Hz, and timestamp_ns equals RGB mid_exposure_utc_ns. Position is in the device's world tracking frame (metres, arbitrary origin per session). Orientation is an x, y, z, w quaternion.

Articulated hand tracking: `hand_tracking.csv`

524 columns: frame_number, timestamp, left_active, right_active, then 26 joints × 10 fields for the left hand, then the same for the right hand (left_joint0_id … right_joint25_orientation_w).
Per-joint fields: id, name, radius, pos_x/y/z, orientation_x/y/z/w.
Joint order follows the OpenXR 26-joint hand: PALM, WRIST, THUMB_{METACARPAL,PROXIMAL,DISTAL,TIP}, then {METACARPAL,PROXIMAL,INTERMEDIATE,DISTAL,TIP} for INDEX, MIDDLE, RING and LITTLE.
The native rate varies by session: median 16.5 Hz, 5th–95th percentile 5.9–26.6 Hz (sessions to 2026-10-01). Every timestamp equals an RGB mid_exposure_utc_ns, and positions are in the head-pose world frame.
When a hand is not tracked, *_active = 0, joint ids are -1 and numeric fields are 0.
These are on-device tracking estimates, not motion-capture ground truth.

Derived hand keypoints: `hand_2d.csv`, `hand_3d.csv`

# hand_2d: t_us,frame,hand,x0,y0,...,x25,y25
# hand_3d: t_us,frame,hand,x0,y0,z0,...,x25,y25,z25

These files have one row per visible hand per RGB frame. t_us equals RGB pts_us, and frame equals RGB frame_index.

hand_2d: normalised coordinates on the left RGB eye image. A few points can fall slightly outside [0, 1].
hand_3d: integer mm in the head frame, axes +x right, +y up, -z forward. Latency compensation is 4 frames (see qc_report.json → hand_3d).

IMU: `accel.csv`, `gyro.csv`, `imu_calibration.json`

timestamp_ns,x,y,z

The IMU runs at about 1,008 Hz, with Unix-epoch ns timestamps. Accelerometer values are in m/s² including gravity, and gyroscope values are in rad/s.

imu_calibration.json gives:

bias, scale factor and non-orthogonality
noise and bias-walk standard deviations
time-alignment offsets in seconds, for imu_to_pose and for each camera (rgb-left, rgb-right, trackingA/B, ctrl-trackingA/B)

Action labels: `action_labels.csv`

second,start_s,end_s,action_label,objects,confidence
0,0.0,1.0,walk,,
1232,1232.0,1233.0,twist_fasten,light bulb;socket,0.95

Column	Meaning
`second`	Integer second from the first RGB frame
`start_s`, `end_s`	Interval in seconds relative to the first RGB frame
`action_label`	Verb-style action label from a controlled vocabulary (see Statistics)
`objects`	Semicolon-separated objects involved, free text, empty if none
`confidence`	Model-reported 0–1 score, averaged over each run of consecutive same-label seconds. Empty when not reported (63% of label runs)

Labels are machine-generated by vision-language models (see Annotations). The vocabulary has three tiers:

structural labels: idle, no_action, other, occluded
general verbs: walk, hold, adjust, inspect, …
38 curated fine-grained manipulation labels: wipe_surface, twist_fasten, insert_align, paint_coat, drill_hole, …

The vocabulary grows from evidence: a label proposed by the model in 3 or more sessions is promoted into it.

Motion signals, sync and segmentation

File	Contents
`frame_signals.csv`	`frame_index, device_ns, finger_speed, grasp_aperture, gyro_energy, hand_activity, head_speed, motion_energy`. Hand signals are empty when no hand is tracked
`sync_anchor.json`	First and last device and wall clock, `frame_hz`, `frames`, `quality`
`sync_manifest.csv`	`frame_index, device_ns, wall_ns` per RGB frame
`spans.csv`	`seq, t0_ns, t1_ns, t0_wall_ns, t1_wall_ns, start_s, end_s`, spans of about 120 s

Quality control: `qc_report.json`

The report covers four streams: accel, gyro, hand_tracking and head_pose. For each it gives status, rows, rate_hz, covered_fraction, gaps, rollback_drops, unplaceable_drops and slip_runs.

It also gives hand_active_ratio, hand-geometry conventions, span statistics and sync.quality. Filter on these fields to select clean sessions.

v2 segments

Each v2/raw/<segment>/ folder has the camera, metadata, calibration, head-pose, hand and IMU files of a v1 session, cut from a longer recording, plus segment.json, depth/, slam/ and the preview videos. Where v2 differs from v1:

Cut. Videos are stream copies cut at each camera's keyframes, with no re-encoding. ctrl and tracking start a few seconds before the RGB. segment.json gives the frame ranges, the UTC window and per-camera start offsets (offsets_ms). The tables cover the RGB window ± 0.5 s. Align on mid_exposure_utc_ns, never by frame index.
hand_tracking.csv has the same 524 columns, but one row per RGB frame with tracking: timestamp equals that frame's mid_exposure_utc_ns and frame_number its index in rgb.mp4. Rows in the ± 0.5 s margin have a negative frame_number or one past the last frame.
Not included: controller_poses.csv (controllers are not used) and rgb_stereo_calibration.json. depth/rgb_depth.json carries the rectification the depth was computed in.

Stereo depth: `depth/rgb_depth.mp4`, `rgb_depth.json`

Estimated, not measured. Depth comes from FoundationStereo (NVIDIA, 24 iterations, TensorRT fp16) run on the rectified, blurred RGB stereo pair. There is one map per RGB frame, 640 × 480, for the left eye in its rectified frame: rectified.K, rectified.R_left (left camera → rectified) and rectified.baseline_m (about 6.3 cm).
Encoding. The 10-bit HEVC luma v stores inverse depth: 1/z = 1/zmax + (v − 32) / 991 · (1/zmin − 1/zmax), with zmin 0.25 m and zmax 20 m. v < 16 means no depth. Decode with the parameters in encoding; the quick start has a snippet. The video is a data store, not meant for viewing; depth_preview_anon.mp4 is the one to watch.
Masking. Frames whose depth collapses are stored as no depth (collapsed_frames). Pixels that flicker against both neighbouring frames are blanked too (spike_masked_share of pixels).
QA. alignment_qa checks the rectification on held-out frames, and model_qa checks disparity against sparse SIFT matches. Neither is a ground-truth depth comparison.

SLAM: `slam/`

Each segment's slam/README.md says how it was produced and what it covers. In short:

trajectory_mNN.tum: use these. There is one merged trajectory per continuous stretch, with the best source per moment, in TUM format (t_s x y z qx qy qz qw). t_s is UTC epoch seconds, on the same clock as mid_exposure_utc_ns. The pose is of the IMU/body frame, in metres, levelled so that z points up. Every file has its own origin, and trajectories are not linked to each other.
Sources, in the order the cascade runs them:
1. ORB-SLAM3 RGB-D + IMU on the blurred left RGB eye and the stereo depth
2. ORB-SLAM3 stereo-inertial on the tracking cameras, for the gaps
3. the headset's own pose (head_pose.csv), where it passes an accelerometer check
4. cuVSLAM camera-only fill
merge.json records the source chosen for each 10-second cell. quality.json gives the status of every second (seconds.intervals) and the checks. light.csv counts trackable corners per second.
Per-source pieces: trajectory_dNN.tum (depth), trajectory_sNN_pMM.tum (tracking cameras), trajectory_hNN.tum (headset) and trajectory_cNN.tum with .sources.csv (cuVSLAM chains). map_sNN_pMM.ply is a sparse point map, only where a tracking-camera piece exists.
Hands in the room. handpaths.parquet has the 26 joints of each tracked hand in the trajectory frame named by frame, with t_ns on the RGB clock. actions.parquet turns each hand into a gripper: observation.state is x y z qx qy qz qw gripper at the frame and action the same at the next frame of the same action segment (segment), with the aperture in aperture_m. stances.parquet lists periods of standing in one place (within 0.2 m for at least 3 s) with body position and heading. The .json next to each file summarises it.
Run records. slam_run.json holds the settings, inputs, summary and compute. status.json is the job status; its handpaths figures are from a first pass, so use handpaths.json.

Time base

Reference clock: RGB mid_exposure_utc_ns, which equals head_pose.timestamp_ns and hand_tracking.timestamp.
pts_us, t_us and *_s fields are relative to the first RGB frame.
The IMU is not frame-aligned. Use merge_asof or interpolation on timestamp_ns, and apply the per-camera offsets in imu_calibration.json.
ctrl and tracking streams can start up to about ±0.6 s from RGB. Align them on their own mid_exposure_utc_ns, never by frame index.

Statistics

Metric	Full dataset	v1 demo session `azimov-demo`
Duration	791.8 h (600.1 h fully processed)	1,442.5 s (24 min 3 s)
Sessions	1,393 (1,043 fully processed)	1
Session length	mean 34 min · median 26 min · max 156 min	—
RGB frames (30 Hz)	≈ 85.5 M	43,191
Hand-tracking rate	median 16.5 Hz (5th–95th pct 5.9–26.6 Hz)¹	15.6 Hz (22,518 rows; 22,517 after QC)
IMU rate	~1,008 Hz¹	1,014 Hz (1,469,094 samples)
Hand active ratio	0.985 (duration-weighted, 1,055 sessions)	0.9846
Labelled time	598.8 h (76%) in 312,175 label runs	1,440 s in 323 label runs
Labels in use	115 (38 curated manipulation)	29
Distinct object strings	3,847 (free text, not a class set)	30 (46 combinations)
Sync quality `ok`	1,069 / 1,078 sessions with a sync check	✓
Hand tracking, head pose and sync all `ok`	1,039 sessions · 607.1 h	✓
Size	6.22 TB raw + processed derivatives	Raw 4.90 GB · LeRobot 0.93 GB · MCAP 4.83 GB

¹ Computed from the per-session QC reports of the 764 sessions available on 2026-10-01; the database figures above are as of 2026-11-13.

Hours by headset

Device letters match the first release of this card.

Headset	Sessions	Hours
Device A	222	163.8
Device B	246	166.9
Device C	393	153.4
Device D	269	158.2
Device E	262	149.4

Where the time goes

Shares of the 598.8 labelled hours. Categories follow the vocabulary: curated labels are the 38 fine-grained manipulation labels, other hand use is every other label flagged as manipulation, and the rest is locomotion and observation.

Category	Share	Examples
Curated fine-grained manipulation	26.6%	`wipe_surface`, `twist_fasten`, `paint_coat`, `take`, `press_control`, `roll_coat`, `insert_align`, `cut_divide`, `trowel_spread`, `twist_release`
Other hand use and tool handling	36.8%	`hold`, `adjust`, `drive_machine`, `carry`, `mop_floor`, `push_wheeled`, `wash`, `use_phone`, `take_photo`, `tighten_screw`, `scrape_wall`
Locomotion and observation	24.2%	`walk`, `inspect`, `watch_coworker`, `look_around`, `look_around_in_dark`, `climb_ladder`
Structural (`idle`, `no_action`, `other`, `occluded`)	12.3%	—

Top 25 labels by hours

Label	Hours	Sessions	Label	Hours	Sessions
`walk`	77.0	970	`mop_floor`	11.3	134
`idle`	53.5	953	`look_around`	11.2	266
`hold`	41.4	891	`take`	10.6	893
`adjust`	34.3	840	`press_control`	9.5	613
`roll_coat`	25.1	121	`push_wheeled`	8.3	356
`inspect`	23.7	817	`scrape_wall`	8.2	144
`carry`	18.7	803	`occluded`	8.2	545
`drive_machine`	17.6	100	`load_tool`	7.7	172
`wipe_surface`	15.9	497	`insert_align`	7.3	482
`trowel_spread`	14.1	79	`cut_divide`	7.2	234
`twist_fasten`	13.7	382	`look_around_in_dark`	7.2	190
`paint_coat`	11.7	86	`wash`	6.6	173
`watch_coworker`	11.5	222

The full table of 115 labels, with hours and session counts, ships with the review package.

Most frequent objects (sessions): phone/smartphone, floor, ladder, door, tool, cart, cable, glove, cloth, wall, ceiling, screwdriver, panel, bucket, pipe, door handle, screw, wire, hose, box, ceiling panel, switch, flashlight, mop, elevator button, control panel.

Environments

Each session carries one environment category, assigned by the labelling model from the footage. It is not site metadata.

Environment	Sessions	Hours
Parking garage	273	141.7
Circulation (corridors, lobbies, lifts)	135	78.4
Plant room (electrical, HVAC)	142	72.1
Restroom	90	52.8
Office	46	38.8
Food and hospitality	57	36.0
Outdoor	59	31.4
Ceiling plenum	71	27.5
Retail	42	27.5
Leisure and fitness	31	23.6
Other	61	37.8
Not categorised	386	224.3

Inside the v1 demo session

Hand tracking

Hands come in three forms, all in the same 26-joint OpenXR order: device-native articulated tracking (hand_tracking.csv, 22,518 rows at 15.6 Hz, metres, world frame), 2D keypoints on the left RGB eye (hand_2d.csv, 29,580 rows: 12,387 left, 17,193 right) and 3D keypoints in the head frame (hand_3d.csv, 29,648 rows: 12,401 left, 17,247 right).

Both hands in 3D, head frame, frame 25,365

Both hands from hand_3d.csv at RGB frame 25,365 (twist_release, t = 14:05), in the head frame. The larger dot marks the wrist.

Share of hand_tracking.csv samples with each hand active, per minute. Over the session the right hand is active in 77% of samples and the left in 55% (at least one hand: 98.5%). The left hand is often out of view during overhead screwdriver work.

Action annotations

action_labels.csv has one row per second (1,440 rows), with an action class, the objects involved (556 rows) and a model-reported confidence (889 rows).

The 29 classes are grouped into four families for this chart only. The families are not part of the vocabulary.

Seconds labelled per action class

Anonymisation in the demo



A bystander's face blurred in the RGB stream (frame 24,943, t = 13:51)	The grayscale wide-FOV ctrl pair, side by side (`ctrl.mp4`, 640 × 480 per eye)

Faces are blurred in this demo. Shop signage is not, so store names in the mall are readable.

Data samples

Expand any block to see real rows from the demo. Most windows start at RGB frame 25,365 (t = 14:05), the twist_release moment in the 3D plot above.

hand_3d.csv: RGB frames 25,360–25,374, joints 1 (wrist), 5 (thumb tip), 10 (index tip) of 78 coordinate columns, mm, head frame

t_us	frame	hand	x1	y1	z1	x5	y5	z5	x10	y10	z10
845175676	25360	left	-166	70	-290	-111	174	-282	-112	167	-276
845175676	25360	right	-47	-20	-257	-92	90	-270	-99	95	-282
845209004	25361	left	-165	69	-289	-111	173	-283	-111	166	-276
845209004	25361	right	-49	-21	-257	-93	90	-271	-101	98	-284
845242331	25362	left	-164	68	-289	-110	172	-284	-111	166	-277
845242331	25362	right	-50	-21	-257	-93	90	-271	-101	102	-286
845275658	25363	left	-153	63	-283	-103	166	-292	-97	170	-286
845275658	25363	right	-50	-21	-258	-94	90	-271	-100	106	-289
845308985	25364	left	-148	61	-279	-100	162	-294	-91	173	-293
845308985	25364	right	-51	-20	-259	-95	91	-271	-100	109	-290
845342312	25365	left	-150	63	-280	-102	164	-295	-93	174	-295
845342312	25365	right	-51	-20	-259	-95	91	-271	-101	108	-290
845375639	25366	left	-150	63	-281	-102	166	-300	-91	173	-304
845375639	25366	right	-51	-20	-259	-96	91	-270	-101	108	-290
845408966	25367	left	-150	69	-286	-102	175	-310	-89	174	-315
845408966	25367	right	-54	-20	-261	-101	91	-269	-104	108	-290
845442293	25368	right	-54	-20	-261	-102	92	-268	-105	108	-290
845475621	25369	right	-55	-20	-261	-102	92	-269	-101	111	-296
845508948	25370	right	-55	-20	-263	-114	88	-278	-100	109	-316
845542275	25371	right	-57	-21	-264	-105	91	-283	-96	112	-323
845575602	25372	right	-58	-21	-265	-105	90	-284	-95	113	-325
845608929	25373	right	-58	-21	-266	-105	91	-285	-94	113	-327
845642256	25374	right	-56	-21	-277	-109	90	-288	-102	112	-331

hand_2d.csv: same window, joints 0, 1, 5, 10 of 52 coordinate columns, normalised to the left RGB eye

t_us	frame	hand	x0	y0	x1	y1	x5	y5	x10	y10
845175676	25360	left	0.3846	0.3684	0.3676	0.4247	0.4358	0.2517	0.4324	0.2583
845175676	25360	right	0.5201	0.5138	0.5207	0.5915	0.4541	0.3781	0.4479	0.3745
845209004	25361	left	0.3851	0.3703	0.3683	0.4268	0.436	0.2543	0.4332	0.2601
845209004	25361	right	0.5175	0.5146	0.5171	0.5925	0.4523	0.3786	0.446	0.3695
845242331	25362	left	0.3856	0.3709	0.3687	0.4276	0.4366	0.2564	0.4337	0.2605
845242331	25362	right	0.5171	0.5149	0.5159	0.593	0.4522	0.3786	0.4463	0.3637
845275658	25363	left	0.3955	0.3778	0.3781	0.4341	0.4463	0.2697	0.4531	0.2586
845275658	25363	right	0.5169	0.5144	0.5153	0.5922	0.4517	0.378	0.4474	0.3585
845308985	25364	left	0.4002	0.3806	0.3827	0.4365	0.4508	0.2758	0.4617	0.26
845308985	25364	right	0.5165	0.5134	0.5147	0.5911	0.4508	0.3768	0.4478	0.3552
845342312	25365	left	0.398	0.379	0.3809	0.4344	0.4487	0.275	0.4593	0.2601
845342312	25365	right	0.5162	0.5133	0.5145	0.591	0.4498	0.3763	0.4475	0.3557
845375639	25366	left	0.3992	0.3807	0.3816	0.4334	0.4497	0.2752	0.463	0.2668
845375639	25366	right	0.5163	0.5134	0.5147	0.5909	0.4489	0.3761	0.4475	0.3565
845408966	25367	left	0.402	0.3772	0.3841	0.4245	0.4515	0.2695	0.4673	0.2721
845408966	25367	right	0.5116	0.5127	0.51	0.5899	0.4419	0.3751	0.4431	0.3568
845442293	25368	right	0.5107	0.5123	0.5092	0.5895	0.4405	0.3745	0.4421	0.3571
845475621	25369	right	0.5114	0.5129	0.5086	0.5893	0.4398	0.3741	0.448	0.3536
845508948	25370	right	0.511	0.5186	0.5077	0.5897	0.4257	0.387	0.4542	0.3662
845542275	25371	right	0.509	0.5192	0.505	0.5901	0.4403	0.3829	0.4594	0.3634

hand_tracking.csv: wrist (joint 1) and index tip (joint 10) positions, metres, world frame, 10 rows with both hands tracked (16 of 524 columns)

frame_number	timestamp	left_active	right_active	left_joint1_pos_x	left_joint1_pos_y	left_joint1_pos_z	left_joint10_pos_x	left_joint10_pos_y	left_joint10_pos_z	right_joint1_pos_x	right_joint1_pos_y	right_joint1_pos_z	right_joint10_pos_x	right_joint10_pos_y	right_joint10_pos_z
25351	1788486699920453792	1	1	108.547493	3.516947	35.397533	108.577919	3.618667	35.464241	108.744545	3.484156	35.45779	108.671356	3.603851	35.437977
25352	1788486699953780875	1	1	108.558136	3.530879	35.385803	108.587997	3.625873	35.453716	108.733902	3.491326	35.449711	108.661118	3.611229	35.431568
25353	1788486699987108011	1	1	108.575867	3.559894	35.365509	108.609703	3.638509	35.433273	108.727547	3.496884	35.441666	108.649635	3.611587	35.436569
25354	1788486700020435094	1	1	108.582886	3.571405	35.364529	108.61998	3.64894	35.433231	108.720879	3.501462	35.438229	108.645729	3.612324	35.428677
25355	1788486700053762229	1	1	108.585922	3.576835	35.364159	108.62664	3.656284	35.432068	108.708946	3.509086	35.433197	108.641411	3.614136	35.418736
25356	1788486700087089365	1	1	108.59816	3.600505	35.362007	108.644859	3.686777	35.423637	108.711266	3.507752	35.434082	108.644928	3.613889	35.421722
25357	1788486700120416448	1	1	108.601173	3.609275	35.369324	108.638206	3.695004	35.434399	108.710762	3.506396	35.433743	108.647644	3.615098	35.422474
25358	1788486700153743584	1	1	108.602493	3.6129	35.372681	108.630409	3.695807	35.443066	108.708672	3.505731	35.432247	108.650337	3.617271	35.423374
25359	1788486700187070719	1	1	108.60228	3.611184	35.3708	108.622757	3.689281	35.44788	108.70504	3.505884	35.429577	108.652901	3.620103	35.424572
25360	1788486700220397802	1	1	108.603088	3.609834	35.372005	108.623672	3.688085	35.448509	108.702553	3.506662	35.426651	108.650948	3.624632	35.423435

action_labels.csv: seconds 1,204–1,233, the bulb swap

second	start_s	end_s	action_label	objects	confidence
1204	1204.0	1205.0	idle	—	0.7
1205	1205.0	1206.0	climb_ladder	ladder	0.7
1206	1206.0	1207.0	twist_release	screw;screwdriver;light fixture	0.9
1207	1207.0	1208.0	remove	light cover;light fixture	0.9
1208	1208.0	1209.0	hold	light cover;screwdriver	0.867
1209	1209.0	1210.0	hold	light cover;screwdriver	0.867
1210	1210.0	1211.0	hold	light cover;screwdriver	0.867
1211	1211.0	1212.0	reorient_in_hand	light cover	0.8
1212	1212.0	1213.0	hold	light cover;screwdriver	0.8
1213	1213.0	1214.0	reorient_in_hand	light cover	0.85
1214	1214.0	1215.0	hold	light cover;screwdriver	0.85
1215	1215.0	1216.0	descend_ladder	ladder	0.7
1216	1216.0	1217.0	climb_ladder	ladder	0.7
1217	1217.0	1218.0	climb_ladder	ladder	0.7
1218	1218.0	1219.0	idle	—	0.7
1219	1219.0	1220.0	idle	—	0.7
1220	1220.0	1221.0	idle	—	0.7
1221	1221.0	1222.0	take	light bulb	0.9
1222	1222.0	1223.0	put	light bulb	0.85
1223	1223.0	1224.0	reorient_in_hand	light cover	0.85
1224	1224.0	1225.0	reorient_in_hand	light cover	0.85
1225	1225.0	1226.0	put	light cover	0.8
1226	1226.0	1227.0	put	light cover	0.8
1227	1227.0	1228.0	take	light bulb	0.9
1228	1228.0	1229.0	hold	light bulb;light cover;screwdriver	0.85
1229	1229.0	1230.0	put	screwdriver	0.8
1230	1230.0	1231.0	climb_ladder	ladder	0.8
1231	1231.0	1232.0	insert_align	light bulb;socket	0.9
1232	1232.0	1233.0	twist_fasten	light bulb;socket	0.95
1233	1233.0	1234.0	descend_ladder	ladder	0.7

head_pose.csv, accel.csv, gyro.csv: 10 rows each from RGB frame 25,365

timestamp_ns	pos_x	pos_y	pos_z	quat_x	quat_y	quat_z	quat_w
1788486700253724938	108.607	3.41276	35.6507	0.218953	-0.280225	0.0352711	0.933965
1788486700287052073	108.606	3.41217	35.6506	0.219767	-0.281043	0.0359044	0.933503
1788486700320379157	108.606	3.41159	35.6505	0.220369	-0.282161	0.0362706	0.93301
1788486700353706292	108.606	3.4112	35.6505	0.220693	-0.283203	0.0369608	0.93259
1788486700387033427	108.606	3.41055	35.6508	0.221487	-0.283765	0.0367717	0.932239
1788486700420360511	108.607	3.40987	35.6513	0.22313	-0.284827	0.0356278	0.931567
1788486700453687646	108.608	3.40928	35.6517	0.224499	-0.286585	0.0355521	0.930702
1788486700487014729	108.608	3.40904	35.6518	0.225119	-0.287924	0.0357847	0.930129
1788486700520341865	108.608	3.40862	35.6518	0.22527	-0.28883	0.0359152	0.929807
1788486700553668948	108.608	3.40831	35.6515	0.225944	-0.289226	0.0363286	0.929504

timestamp_ns	x	y	z
1788486700253882219	-0.633566	8.82383	-4.37062
1788486700254868156	-0.621595	8.86094	-4.38139
1788486700255854094	-0.604835	8.88728	-4.38139
1788486700256840031	-0.585682	8.88847	-4.38139
1788486700257825969	-0.572513	8.87171	-4.35386
1788486700258811906	-0.558148	8.83101	-4.37421
1788486700259797844	-0.552163	8.7951	-4.36343
1788486700260783781	-0.562937	8.78074	-4.37899
1788486700261769719	-0.576105	8.78552	-4.39216
1788486700262755656	-0.588076	8.80228	-4.41132

timestamp_ns	x	y	z
1788486700253882219	0.0746995	-0.0323565	0.0347532
1788486700254868156	0.0744332	-0.0334217	0.0358185
1788486700255854094	0.0739005	-0.033688	0.0368837
1788486700256840031	0.0733679	-0.0339543	0.0366174
1788486700257825969	0.072569	-0.0342206	0.03715
1788486700258811906	0.0717701	-0.0350195	0.0379489
1788486700259797844	0.0715038	-0.0374163	0.0382152
1788486700260783781	0.0701722	-0.0384815	0.0384815
1788486700261769719	0.069107	-0.0390142	0.0390142
1788486700262755656	0.0669765	-0.0403457	0.0392805

rgb_metainfo.csv and sync_manifest.csv: 10 rows from frame 25,365

frame_index	frame_id	pts_us	exposure_start_utc_ns	exposure_duration_ns	gain	mid_exposure_utc_ns
25365	34018	845342312	1788486700249280494	8888888	273	1788486700253724938
25366	34019	845375639	1788486700282607629	8888888	273	1788486700287052073
25367	34020	845408966	1788486700315934713	8888888	273	1788486700320379157
25368	34021	845442293	1788486700349261848	8888888	277	1788486700353706292
25369	34022	845475621	1788486700382588983	8888888	277	1788486700387033427
25370	34023	845508948	1788486700415916067	8888888	277	1788486700420360511
25371	34024	845542275	1788486700449243202	8888888	277	1788486700453687646
25372	34025	845575602	1788486700482570285	8888888	277	1788486700487014729
25373	34026	845608929	1788486700515897421	8888888	277	1788486700520341865
25374	34027	845642256	1788486700549224504	8888888	281	1788486700553668948

frame_index	device_ns	wall_ns
25365	1788486700253724938	1788486700253724928
25366	1788486700287052073	1788486700287052032
25367	1788486700320379157	1788486700320379136
25368	1788486700353706292	1788486700353706240
25369	1788486700387033427	1788486700387033344
25370	1788486700420360511	1788486700420360448
25371	1788486700453687646	1788486700453687552
25372	1788486700487014729	1788486700487014656
25373	1788486700520341865	1788486700520341760
25374	1788486700553668948	1788486700553668864

spans.csv: all spans

seq	t0_ns	t1_ns	t0_wall_ns	t1_wall_ns	start_s	end_s
0	1788485854944740407	1788485974922360042	1788485854944740352	1788485974922360064	0.033	120.011
1	1788485974922360042	1788486094899979677	1788485974922360064	1788486094899979776	120.011	239.989
2	1788486094899979677	1788486214877599313	1788486094899979776	1788486214877599232	239.989	359.966
3	1788486214877599313	1788486334855218896	1788486214877599232	1788486334855218944	359.966	479.944
4	1788486334855218896	1788486454832838532	1788486334855218944	1788486454832838400	479.944	599.921
5	1788486454832838532	1788486574810458167	1788486454832838400	1788486574810458112	599.921	719.899
6	1788486574810458167	1788486694788077802	1788486574810458112	1788486694788077824	719.899	839.877
7	1788486694788077802	1788486814765697438	1788486694788077824	1788486814765697536	839.877	959.854
8	1788486814765697438	1788486934743317073	1788486814765697536	1788486934743316992	959.854	1079.832
9	1788486934743317073	1788487054720936709	1788486934743316992	1788487054720936704	1079.832	1199.81
10	1788487054720936709	1788487174698556344	1788487054720936704	1788487174698556416	1199.81	1319.787
11	1788487174698556344	1788487296575386705	1788487174698556416	1788487296575386624	1319.787	1441.664
12	1788487296575386705	1788487303375386637	1788487296575386624	1788487303375386624	1441.664	1448.464

qc_report.json (slip runs summarised)

{
  "activity_ratio": 1.0,
  "duration_s": 1442.53,
  "hand_3d": {
    "axes": "+x right, +y up, -z forward",
    "frame": "head",
    "latency_frames": 4,
    "rows": 29648,
    "rows_dropped_on_unusable_head": 90,
    "time_base": "pts_us",
    "units": "mm"
  },
  "hand_active_ratio": 0.9846,
  "hand_geometry": {
    "axis_fit": 0.9992878861017902,
    "convention": "xyzw camera_to_device fisheye roll270 direct",
    "crop": "iw/2:ih:0:0",
    "eye": "left",
    "height": 1748,
    "latency_frames": 4,
    "stream": "rgb",
    "time_base": "pts_us",
    "width": 2328
  },
  "notes": [],
  "schema": "vfrog.ego.artifacts/5",
  "session_id": "azimov-demo",
  "spans": {
    "count": 13,
    "longest_s": 121.877,
    "median_s": 119.978,
    "seconds_total": 1448.431,
    "shortest_s": 6.8
  },
  "streams": {
    "accel": {
      "bridge_out_of_order": 0,
      "covered_fraction": 1.0041,
      "detail": null,
      "gaps": 0,
      "rate_hz": 1014.221,
      "rollback_drops": 0,
      "rows": 1469093,
      "status": "ok",
      "unplaceable_drops": 0,
      "slip_runs": "0 runs"
    },
    "gyro": {
      "bridge_out_of_order": 0,
      "covered_fraction": 1.0041,
      "detail": null,
      "gaps": 0,
      "rate_hz": 1014.221,
      "rollback_drops": 0,
      "rows": 1469089,
      "status": "ok",
      "unplaceable_drops": 0,
      "slip_runs": "0 runs"
    },
    "hand_tracking": {
      "bridge_out_of_order": 0,
      "covered_fraction": 0.9966,
      "detail": null,
      "gaps": 0,
      "rate_hz": 15.663,
      "rollback_drops": 1,
      "rows": 22517,
      "status": "ok",
      "unplaceable_drops": 0,
      "slip_runs": "12 runs"
    },
    "head_pose": {
      "bridge_out_of_order": 0,
      "covered_fraction": 0.996,
      "detail": null,
      "gaps": 0,
      "rate_hz": 29.997,
      "rollback_drops": 0,
      "rows": 43100,
      "status": "ok",
      "unplaceable_drops": 0,
      "slip_runs": "0 runs"
    }
  },
  "sync": {
    "first_wall_ns": 1788485854911413271,
    "frames": 43191,
    "last_wall_ns": 1788487296208720042,
    "quality": "ok"
  },
  "thumbnails": 12
}

imu_calibration.json

{
  "device_uid": "1752133326",
  "imu": {
    "imu_id": 0,
    "is_primary": true,
    "bias": {
      "accelerometer_mps2": [
        -0.003453006,
        0.068447895,
        -0.142576277
      ],
      "gyroscope_rads": [
        -0.00092585,
        0.01141167,
        -0.004215184
      ]
    },
    "scale_factor": {
      "accelerometer": [
        0.001388019,
        0.000703511,
        0.003950217
      ],
      "gyroscope": [
        0.005798217,
        0.004832713,
        0.001032803
      ]
    },
    "nonorthogonality": {
      "accelerometer": [
        -0.001107834,
        0.001083146,
        -0.00060965
      ],
      "gyroscope": [
        0.001343174,
        -0.005489959,
        0.000255489
      ]
    },
    "time_alignment_s": {
      "imu_to_pose": 0.004795488,
      "cameras": {
        "trackingA": 0.004795488,
        "trackingB": 0.004795488,
        "ctrl-trackingA": 0.004786782,
        "ctrl-trackingB": 0.004786782,
        "rgb-left": 0.002417459,
        "rgb-right": 0.002400031
      },
      "accel": 0.0
    }
  },
  "noise": {
    "accel_noise_std_mps2": [
      0.02,
      0.02,
      0.02
    ],
    "gyro_noise_std_rads": [
      0.0016,
      0.0016,
      0.0016
    ],
    "accel_bias_std_mps2": [
      0.050000001,
      0.050000001,
      0.050000001
    ],
    "gyro_bias_std_rads": [
      0.005,
      0.005,
      0.005
    ]
  }
}

camera_params_rgb.json (left eye)

{
  "group": "rgb",
  "cameras": [
    {
      "eye": "left",
      "width": 2328,
      "height": 1748,
      "intrinsics": {
        "focalX": 877.437744,
        "focalY": 877.437744,
        "centerX": 1166.628052,
        "centerY": 864.766052,
        "radialDistortion": [
          -0.118433,
          0.298468,
          -0.256264,
          0.073082,
          0.0,
          0.0,
          0.0,
          0.0
        ]
      },
      "extrinsics": {
        "position": [
          -0.050093,
          0.027196,
          -0.01393
        ],
        "rotation": [
          0.710602,
          0.70336,
          -0.012391,
          -0.013293
        ]
      }
    }
  ]
}

sync_anchor.json

{
  "first_device_ns": 1788485854911413271,
  "first_wall_ns": 1788485854911413271,
  "frame_hz": 30.0,
  "frames": 43191,
  "last_device_ns": 1788487296208720042,
  "last_wall_ns": 1788487296208720042,
  "quality": "ok",
  "schema": "vfrog.ego.artifacts/5"
}

Quick start

# v1 Raw: MP4 + JSON + CSV
hf download vfrogAI/azimov --repo-type dataset --include "v1/raw/azimov-demo/**" --local-dir ./azimov

# v1 Raw tables and calibration only, no video
hf download vfrogAI/azimov --repo-type dataset \
  --include "v1/raw/azimov-demo/*.csv" --include "v1/raw/azimov-demo/*.json" --include "v1/manifests/*" \
  --local-dir ./azimov

# v1 LeRobot v3.0
hf download vfrogAI/azimov --repo-type dataset --include "v1/lerobot/**" --local-dir ./azimov

# v1 MCAP
hf download vfrogAI/azimov --repo-type dataset --include "v1/mcap/azimov-demo.mcap" --local-dir ./azimov

# v2: one segment with depth and SLAM (about 2 GB)
hf download vfrogAI/azimov --repo-type dataset --include "v2/raw/curb-painting/**" --local-dir ./azimov

v2: depth and SLAM

import json
import subprocess

import numpy as np
import pandas as pd
from huggingface_hub import snapshot_download

seg = snapshot_download(
    repo_id="vfrogAI/azimov", repo_type="dataset", allow_patterns="v2/raw/curb-painting/**",
) + "/v2/raw/curb-painting"

# Merged SLAM trajectory: UTC seconds, metres, z up
traj = pd.read_csv(f"{seg}/slam/trajectory_m01.tum", sep=" ", comment="#",
                   names=["t_s", "x", "y", "z", "qx", "qy", "qz", "qw"])

# Gripper-style hand actions in the same frames
actions = pd.read_parquet(f"{seg}/slam/actions.parquet")

# Depth of RGB frame 1200, in metres (NaN = no depth); needs ffmpeg on PATH
meta = json.load(open(f"{seg}/depth/rgb_depth.json"))
enc, (w, h) = meta["encoding"], meta["rectified"]["size"]
raw = subprocess.run(
    ["ffmpeg", "-v", "error", "-i", f"{seg}/depth/rgb_depth.mp4", "-vf", r"select=eq(n\,1200)",
     "-frames:v", "1", "-f", "rawvideo", "-pix_fmt", "gray10le", "-"],
    capture_output=True, check=True,
).stdout
v = np.frombuffer(raw, np.uint16).reshape(h, w).astype(np.float32)
inv = 1 / enc["zmax_m"] + (v - enc["code_floor"]) / 991 * (1 / enc["zmin_m"] - 1 / enc["zmax_m"])
depth_m = np.where(v < 16, np.nan, 1 / inv)

v1 Raw

from huggingface_hub import snapshot_download
import pandas as pd

root = snapshot_download(
    repo_id="vfrogAI/azimov", repo_type="dataset", allow_patterns="v1/raw/azimov-demo/**",
) + "/v1/raw/azimov-demo"

rgb_meta = pd.read_csv(f"{root}/rgb_metainfo.csv")
head     = pd.read_csv(f"{root}/head_pose.csv")
actions  = pd.read_csv(f"{root}/action_labels.csv")
hand_3d  = pd.read_csv(f"{root}/hand_3d.csv")
accel    = pd.read_csv(f"{root}/accel.csv")

print(actions["action_label"].value_counts())

The tabular files also load with datasets:

from datasets import load_dataset
actions = load_dataset("vfrogAI/azimov", "demo_action_labels", split="train")

v1 LeRobot

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset
from lerobot.utils.collate import lerobot_collate_fn
from torch.utils.data import DataLoader

root = snapshot_download(repo_id="vfrogAI/azimov", repo_type="dataset", allow_patterns="v1/lerobot/**") + "/v1/lerobot"
ds = LeRobotDataset("vfrogAI/azimov", root=root)
item = ds[100]
print(item["observation.images.head_left"].shape, item["action"].shape, item["task"])

# The language columns are lists of dicts, which PyTorch's default collate rejects
loader = DataLoader(ds, batch_size=8, shuffle=True, collate_fn=lerobot_collate_fn)

v1 MCAP

Open v1/mcap/azimov-demo.mcap in Foxglove, or read it in Python:

from mcap.reader import make_reader

with open("azimov/v1/mcap/azimov-demo.mcap", "rb") as f:
    reader = make_reader(f)
    for schema, channel, message in reader.iter_messages(topics=["/action_label"]):
        print(message.log_time, message.data.decode())
        break

Align an RGB frame with head pose, hands, label and IMU (v1 Raw)

t0 = rgb_meta["mid_exposure_utc_ns"].iloc[0]
f = rgb_meta[["frame_index", "pts_us", "mid_exposure_utc_ns"]].copy()
f["second"] = ((f["mid_exposure_utc_ns"] - t0) // 1_000_000_000).astype(int)

f = f.merge(head, left_on="mid_exposure_utc_ns", right_on="timestamp_ns", how="left")
f = f.merge(actions[["second", "action_label", "objects"]], on="second", how="left")
f = f.merge(hand_3d[hand_3d["hand"] == "left"], left_on="frame_index", right_on="frame", how="left")

f = pd.merge_asof(
    f.sort_values("mid_exposure_utc_ns"),
    accel.sort_values("timestamp_ns").rename(columns={"timestamp_ns": "imu_ns", "x": "ax", "y": "ay", "z": "az"}),
    left_on="mid_exposure_utc_ns", right_on="imu_ns", direction="nearest",
)

Read one eye of an RGB frame

import cv2
cap = cv2.VideoCapture(f"{root}/rgb.mp4")      # HEVC: needs an OpenCV/FFmpeg build with H.265
cap.set(cv2.CAP_PROP_POS_FRAMES, 1200)
ok, frame = cap.read()
left_eye, right_eye = frame[:, : frame.shape[1] // 2], frame[:, frame.shape[1] // 2 :]

Dataset creation

Curation rationale

Robot-learning teams need large volumes of human demonstrations of real work, in real buildings and under real lighting, with the hands, head motion and timing captured precisely enough to learn from.

Phone and single-camera footage lacks the hands, pose and stereo that policy learning needs. Lab teleoperation lacks the variety and the unscripted behaviour. This dataset records professional facility workers during their normal shifts, with every sensor stream the headset provides kept and synchronised.

Source data: collection and processing

Capture. Workers wear a head-mounted 6-camera headset (EgoSense E6) through their normal work. Handheld controllers are not used.
Offload. Recordings are offloaded to cloud storage in the region where they were collected.
Processing pipeline.
- Ingest and probe the raw streams.
- Synchronise clocks (sync_anchor.json, sync_manifest.csv).
- Run per-stream QC.
- Project hand geometry into 2D/3D keypoints.
- Compute motion signals.
- Index actions per second.
- Output conforms to schema vfrog.ego.artifacts/5.
Depth and SLAM (v2 segments). Depth comes from FoundationStereo run on the rectified, blurred RGB stereo pair. SLAM runs a cascade: ORB-SLAM3 RGB-D + IMU on that depth, ORB-SLAM3 stereo-inertial on the tracking cameras, the headset's own pose where it passes an accelerometer check, and cuVSLAM camera-only fill. The hands are then placed in the resulting trajectories.
Selection. Sessions not yet processed, and those that fail probing (no video), clock anchoring or indexing, are excluded from the processed totals (350 of 1,393).
Audio is never retained.

Who are the source data producers?

The recordings were made by professional facility, maintenance and cleaning staff during their normal work. There are 10 wearers (as of 2026-10-09), all based in Qatar.

Annotations

Process. Action labels are produced by vision-language models.
- An index lane samples frames at 1 fps over activity spans proposed from telemetry.
- The model picks one label per second from the controlled vocabulary.
- It also names the objects involved and reports a confidence.
Annotators. Three models labelled the data, and each row records which model labelled it:

Model Share of labelled time

Qwen 3.8 Max 52%

Google Gemini 3.7 Flash 46%

MiniMax M3 2%
Human review. Labels are not human-reviewed.
Evaluation. The acceptance bar is: boundaries within ±1 s on ≥ 80% of matched segments, verb + object correct on ≥ 75%, and ≥ 70% of human segments matched at tIoU ≥ 0.5. Accuracy against human ground truth has not yet been measured. Results will be published with a later release.
Language: English.

Privacy and consent

Consent. Each of the 10 wearers gave written informed consent. The consent covers:

being recorded while working
commercial licensing of the recordings to third parties
use of the recordings to train and evaluate AI and robotics models

Recording on each site was authorised by the site owner or employer.

Wearers. The data has no names, employee IDs or other wearer identifiers. Wearers appear only from the first-person view: hands, forearms, gloves and occasional reflections. Voices are never recorded, because audio is deleted before upload.

Bystanders. Recordings take place in working commercial buildings, so co-workers and members of the public can appear on camera. So can documents, screens, phones and vehicle number plates. Bystanders are protected by blurring and by the licence's ban on identification.

Anonymisation. Every delivered video is blurred before it leaves vfrog: the RGB, ctrl and tracking streams, in every format. Raw unblurred video is never delivered.

Blurred: faces and phones. Legible text (documents, screens, number plates) is also targeted, but large signage is often missed: the shop signs in the demo are readable.
Method: detection runs on keyframes at 2 Hz using in-region (Qatar) detection services. Each box is held for ±15 frames and enlarged by 25%, and the blur is burned in on the fisheye image.
Detection is automated, so misses are possible. Report any you find to hello-azimov@vfrog.ai and a corrected file will be reissued.
Hand tracking, head pose, IMU and labels are computed on-device or from the original frames, so blurring does not affect them. v2 depth, SLAM and the preview videos are computed from the blurred video.

Data protection.

Recordings are offloaded to cloud storage in the region where they were collected, and anonymisation detection runs on in-region services.
The data contains no GPS or other location data. Building interiors may still be recognisable.

Your rights and ours. To ask for a recording to be removed, contact hello-azimov@vfrog.ai. Removals propagate to licensees under the licence.

Uses

Direct use

Pre-training and fine-tuning of VLA models, world models and robot policies from human demonstrations
Imitation learning and learning from video for manipulation
Hand pose estimation, hand-object interaction and grasp analysis
Temporal action recognition and segmentation
Egocentric visual-inertial odometry, SLAM and stereo depth
Benchmarking multi-rate sensor alignment

Out-of-scope and prohibited use

Under every licence type, the following are prohibited:

identifying, contacting, profiling, tracking or re-identifying people in the data
facial recognition, or building facial-image databases
biometric categorisation
emotion inference in the workplace
social scoring
worker surveillance or performance monitoring
weapons development

Bias, risks and limitations

Domain. The data comes from facility work in commercial buildings in one region (Qatar), recorded by five headsets. Domestic, kitchen-cooking, factory-assembly and outdoor tasks are rare or absent.
Lighting. Much of the work is night-shift, so a significant share of footage is dim. Labels such as look_around_in_dark, work_in_dark and walk_in_dark account for about 11 h (as of 2026-10-09).
Uneven task mix. Locomotion and general hand use dominate. Curated fine-grained manipulation is 26.6% of labelled time. Check the statistics before you sample.
Machine labels.
- Labels are VLM output, with no human review. Expect label noise, near-duplicate labels (for example dip_roller_into_paint / dip_paint_roller) and a coarse 1-second granularity.
- confidence is model-reported and empty for 63% of label runs.
- Objects are free text: phone and smartphone, or glove and gloves, appear as separate strings.
Environment metadata is model-written. The environment category is assigned by the labelling model (386 sessions have none, including sessions not yet labelled), and there are no structured site, task or wearer fields.
Hands.
- Hand tracking is an on-device estimate at a median 16.5 Hz (as low as ~6 Hz in some sessions), not mocap ground truth.
- One hand can be tracked much more often than the other. In the v1 demo session the right hand is active in 77% of hand-tracking samples and the left in 55%, and the left hand is often out of view during overhead screwdriver work.
- hand_2d uses the left eye only.
Factory RGB extrinsics in camera_params_rgb.json are not accurate enough for stereo matching. Use rgb_stereo_calibration.json.
Clocks.
- Headset wall clocks drifted by up to weeks, and they have no timezone. Session folder names (YYYYMMDD_HHMMSS) are device-local and must not be used as recording dates.
- Stream timestamps (*_utc_ns) are internally consistent within a session. Use them for alignment.
Units are explicit in the files only for hand_3d (mm) and IMU calibration (m/s², rad/s). Head-pose units (metres) are inferred from value ranges.
Partial sessions. Some processed sessions lack hand_3d (21 as of 2026-10-01), and 39 have at least one degraded or absent QC stream. Filter on qc_report.json (1,039 sessions / 607.1 h pass every check).
v2 depth is estimated. It comes from a stereo model, not a depth sensor. It covers 0.25–20 m, can smooth thin structures and object edges, and has been checked only against sparse feature matches, not against ground truth.
v2 SLAM coverage varies by segment (see the table under New in v2). Dim scenes, blank walls and fast head turns break tracking. Trajectories are not linked to each other, so each one has its own origin. Seconds whose source is headset use the headset's own tracker. ORB-SLAM3 is not deterministic on this data, so a rerun can differ by a few points of coverage.
v2 segments are Raw only. They have no action labels, 2D/3D keypoints, QC report or sync files yet.

Recommendations: filter on the QC report and the manifest before training. Treat labels as weak supervision. Normalise object strings. Hold out sessions by headset and by date, not at random, to avoid leakage between near-identical sessions.

Access and licensing

Step	What you get
1. Request access (form above)	The three v2 segments with depth and SLAM (Raw), the v1 demo session `azimov-demo` in all three formats (Raw, LeRobot, MCAP), plus the schema
2. Evaluation call / review package	Additional demo sessions across environments and tasks, the full label table, and a session index with per-session QC
3. Licence	Full dataset, a custom subset (by hours or QC), or clip orders cut by action label or object, delivered from cloud storage

Licence: vfrog Data Licence, version 1.0. Granted by vfrog, Inc. The full licence in LICENSE.md governs; this summary does not replace it.
- Non-exclusive, worldwide, perpetual and commercial. vfrog can license the same data to others. Exclusive terms are available by order.
- Permitted uses: commercial and non-commercial research; training, fine-tuning and evaluation of AI and machine-learning models; developing and operating robotics and embodied-AI systems; and commercialising the models and products you build. You own the models you train.
- Who may use it: your own legal entity, including employees and contractors working for you under confidentiality.
- Not permitted: redistributing, reselling or sublicensing the data; publishing it or reconstructable derived data; distributing a model that can reproduce it; and the uses listed under Out-of-scope and prohibited use.
- The demo in this repository is licensed for evaluation only (section 2.1): no redistribution, and no training of models you deploy or ship.
Provenance. Each delivered copy carries a Provenance Record and Copy Identifier. Compliance confirmation (section 6) and termination for unauthorised distribution (section 7) apply.
Regulatory support. This card and the per-session manifests and QC reports are supplied to support licensees' data-governance documentation, including EU AI Act Art. 10 and the Art. 53(1)(d) training-data summary.
Ongoing collection. You can subscribe to new hours as they are collected (currently about 230 h/month).

Contact: vfrog, Inc., hello-azimov@vfrog.ai

Citation

@dataset{vfrog_azimov_2026,
  title     = {Azimov: Egocentric Dataset},
  author    = {vfrog, Inc.},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/vfrogAI/azimov}
}

Dataset card authors and contact

vfrog, Inc. · hello-azimov@vfrog.ai

Downloads last month: 602

Total file size:

17 GB

Request access to Azimov