Short answer
A head-mounted capture specification fixes the device and mount, field of view, resolution, constant frame rate, shutter and exposure limits, time base, calibration routine, audio policy and side streams such as IMU, gaze or hand tracking. Set each value from how the footage will be used and from your humanoid's head camera, then confirm it in a pilot before full collection.
What this specification governs
This page covers the camera worn on a human demonstrator's head during commissioned recording. It does not decide whether first-person footage is the right view; that comparison is in egocentric vs exocentric workplace video. Cameras on the demonstrator's wrist are a separate capture choice, covered in wrist-mounted cameras on human demonstrators.
However well specified, a head camera records what a person saw, not what a robot did. It produces no joint states, commanded actions or forces.
Devices used in published egocentric work
Device choice decides which streams exist before any setting is chosen. Project Aria research glasses carry a rolling-shutter RGB camera with a 110° horizontal field of view, two global-shutter monochrome SLAM cameras at 150°, two eye-tracking cameras with infrared illumination, two IMUs and a seven-microphone array, among other sensors (Engel et al., 2023); EgoMimic recorded its human demonstrations on these glasses (Kareer et al., 2024). EgoDex recorded 829 hours of egocentric video with paired 3D hand and finger tracking on a mixed-reality headset, across 194 tabletop tasks (Hoque et al., 2025). HOT3D used two headset types and released multi-view RGB and monochrome video, eye gaze and point clouds, with hand annotations in UmeTrack and MANO formats (Banerjee et al., 2024). EPIC-KITCHENS-100 used head-mounted cameras in 45 kitchens (Damen et al., 2020).
The practical lesson: a consumer action camera gives you video and perhaps IMU. A research headset can add head pose, gaze and native hand tracking, which are far easier to record at capture time than to estimate afterwards.
Illustrative specification
| Parameter | Illustrative setting | Set from | Why it matters |
|---|---|---|---|
| Mount | Glasses-style or headband; logged per session | PPE, comfort, shift length | Changes height, pitch and stability |
| Field of view | Wide enough to keep both hands in frame at bench distance | Task reach; robot head-camera lens | Narrow lenses lose hands; very wide lenses distort edges |
| Resolution | A fingertip spans several pixels at working distance | Hand-tracking needs, storage | Detail for hand pose and small parts |
| Frame rate | Constant; for example 30 fps, higher for fast tasks | Task speed; robot camera rate | Variable frame rate breaks timing models |
| Shutter and exposure | Exposure time capped; shutter type recorded | Lighting, pace | Long exposures smear hands; rolling shutter skews fast motion |
| Stabilization and lens correction | Off; raw frames plus calibration | Downstream geometry | Electronic stabilization changes the effective intrinsics every frame |
| Time base | Per-frame timestamps on a shared clock | Number of devices | Aligns IMU, audio, other cameras and logs |
| Side streams | IMU on; gaze and hand tracking per use case | Device; privacy review | Head motion, attention, hand pose |
| Audio | Off, or on under a written policy | Narration plans, consent | Speech identifies people and may be regulated |
# Illustrative only — example values a buyer replaces; not an available dataset
capture_spec: head_cam_v0.3
device:
model: recorded per unit (unit_id e.g. hcam_07)
firmware: pinned for the whole program
mount: glasses_frame | headband # logged per session
eye_to_lens_offset_mm: measured per demonstrator fit
video:
resolution: 1920x1080 minimum # example
frame_rate: 30 fps constant # example; raise for fast tasks
exposure_max_ms: buyer sets from blur pilot
shutter: rolling | global # recorded, not assumed
stabilization: off
in_camera_distortion_correction: off
codec: H.264 or H.265, high bitrate, no variable frame rate
streams:
imu: on, timestamped on the video clock
slam_pose: on if device supports
gaze: off unless approved per use case
hand_tracking: on if device supports; native format kept
audio: off by default
time:
timestamps: per frame, monotonic
cross_device: shared clock or slate at start and end
calibration:
intrinsics: per unit, per lens setting; repeated if the unit is serviced
camera_imu: spatial and temporal offset per unit
session_card: shown to camera at start (session_id, site_017, operator_A12)
Parameters only you can set
- Field of view and mount angle relative to your robot. Record the offset from the demonstrator's eye to the lens and compare it with your humanoid's head camera; the correction method is described in matching demonstration camera viewpoints to a humanoid's sensors.
- Frame rate against task speed. Fast pick-and-place at production pace needs a higher rate or shorter exposure than slow assembly.
- Gaze capture. Useful for attention and intent, but it adds sensitive data; weigh it using eye-gaze data in task demonstrations.
- Hand tracking. If you plan co-training or retargeting, prefer devices that log hand poses natively.
- Audio. Decide early whether narration is wanted, because it changes consent and redaction.
Calibration and time synchronization
Calibrate every unit, not every model. Intrinsics and lens distortion can be estimated from a planar pattern viewed at a few orientations (Zhang, 2000). Camera-to-IMU spatial and temporal offsets and rolling-shutter intrinsics can be estimated with a toolbox such as Kalibr, which writes a PDF report, a text summary and YAML (ETH Zurich ASL). Ask for those reports in the delivery.
Networked devices can share time through the Precision Time Protocol, which targets sub-microsecond synchronization (IEEE, 2019). Head-worn devices are often wireless and battery powered, so add a visual and audio slate at the start and end of each session and measure the residual offset. Delivery formats for calibration files are set in calibration requirements, and the verification method in sensor synchronization requirements.
Demonstrator protocol
- Fit check before each session: camera level, both hands visible in the preview at working posture.
- Ask demonstrators to work normally. Do not tell them to stare at their hands; it changes the behavior you are trying to record.
- Keep caps, hoods and face shields clear of the lens, or choose a mount that works with the required PPE.
- Plan session length around battery and heat limits, and log every interruption.
- Show a session card with synthetic IDs to the camera at the start.
Acceptance tests
| Test | Method | Example pass criterion (buyer sets) |
|---|---|---|
| Constant frame rate | Read container timestamps | No variable-rate segments; dropped frames below a set fraction |
| Timestamp integrity | Compare video and IMU timestamps | Monotonic; no gaps above threshold |
| Calibration quality | Reprojection error from the calibration sequence | Below threshold for every unit |
| Hand visibility | Hand detector on a random sample | Both hands visible for the share set per task |
| Blur and exposure | Sharpness score and clipped-pixel share on sampled frames | Within limits |
| Stabilization off | Metadata check and a static-scene test | No cropping or warping |
The full delivery checklist is on acceptance criteria for demonstration video.
Limitations
- A human eye line and head motion differ from a humanoid's camera height and neck dynamics, so even good footage needs viewpoint correction.
- The wearer's own arms and tools occlude the hands; head turns move them out of frame.
- Many devices do not fit comfortably over safety glasses or under hard hats, and weight and heat limit session length.
- Gaze and hand-tracking streams raise privacy questions beyond ordinary video; see biometric privacy and workplace video. This is not legal advice.
Questions to ask collection providers
- Which device model and firmware will be used, and will they stay fixed for the whole program?
- Can electronic stabilization and in-camera lens correction be turned off?
- Which side streams are recorded (IMU, head pose, gaze, hand tracking), and in what format?
- How is each unit calibrated, how often, and will the calibration reports be delivered?
- How are clocks aligned across head cameras and any fixed cameras?
- What are the battery and heat limits per session, and how are breaks logged?
- How will the device be worn with the PPE the site requires?
- What hand-visibility rate did the pilot achieve for each task?
Discuss your requirements
If you have a draft capture specification, share it with SourceX along with the tasks and permitted uses. SourceX reviews the requirement and follows up about potential sources; the site also describes a physical-world path for new recordings of hands-on work, where each opportunity has its own requirements. A request does not guarantee a match. The workplace task demonstration hub and the humanoid training data guide cover related capture and sourcing topics, and the catalog entry for first-person video of skilled manual work describes the related asset type.
Describe the workplace demonstration data your robotics project needs
Sources
- Project Aria: A New Tool for Egocentric Multi-Modal AI Research — sensor suite of research glasses used for egocentric capture.
- EgoMimic: Scaling Imitation Learning via Egocentric Video — demonstrations recorded on research glasses for co-training.
- EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video — headset video with paired 3D hand tracking.
- HOT3D: Hand and Object Tracking in 3D from Egocentric Multi-View Videos — headset recordings with gaze, point clouds and hand-pose formats.
- Rescaling Egocentric Vision — head-mounted kitchen recordings.
- A Flexible New Technique for Camera Calibration — planar-pattern intrinsic calibration.
- Kalibr — camera-IMU and rolling-shutter calibration with report outputs.
- IEEE 1588-2019 - IEEE Standard for a Precision Clock Synchronization Protocol for Networked Measurement and Control Systems — networked clock synchronization.