Short answer
A wrist camera on a human demonstrator records what a robot's wrist camera would see: the hand, the object and the contact zone, up close and moving with the hand. It is most usable with a fixed mount, known lens, tracked pose and a gripper-like interface. It does not record forces, and a bare hand differs from a robot end-effector.
Why wrist views are a separate capture decision
Many manipulation policies take a wrist-camera image alongside one or more external views. The wrist view shows the gripper's immediate surroundings, which makes grasp alignment and contact visible even when the arm occludes the scene from a fixed camera. A human demonstration recorded from the wrist can therefore resemble the observations a policy sees at run time, if the capture is designed for that.
Head-mounted capture is covered in the head-mounted camera capture specification; this page covers what changes when the camera rides on the wrist or on a handheld tool. The choice is made at recording time and cannot be reversed from archive footage, so it belongs in the specification you send to a collection team.
Two ways to put a camera on the wrist
| Approach | What is recorded | Strength | Main limitation |
|---|---|---|---|
| Camera on the bare forearm or wrist | Natural hand motion seen from above the wrist | Demonstrator works as usual | Hand shape differs from a gripper; the view varies with wrist rotation |
| Handheld gripper device with camera | Gripper-like interface with the camera in robot-like position | Observation resembles robot wrist view; actions defined by device pose | Constrained, less natural motion; the device is not a humanoid hand |
The Universal Manipulation Interface is a well-documented example of the second approach. It uses a hand-held gripper with a wrist-mounted GoPro with a fisheye lens as its only sensor, plus side mirrors, and recovers motion through IMU-aided SLAM; policies trained on its data were deployed across multiple robot platforms, with about 70% success in out-of-distribution tests (Chi et al., 2024). Note the design: the device sets the action space, so there is no human hand to retarget. For a humanoid with dexterous hands, that trade-off may not fit; see retargeting human hand demonstrations.
For comparison, robot-side wrist views look like DROID's rig: a wrist-mounted ZED Mini with two exterior ZED 2 stereo cameras, and calibration, depth and language instructions per episode (Khazatsky et al., 2024). That is a robot dataset recorded by teleoperation; a human wrist-camera capture tries to match the wrist view, not the robot data around it.
What wrist-camera human data can and cannot support
Can support
- Visual policy inputs that resemble a robot wrist camera, when mount and lens are matched.
- Contact-zone detail: grasp approach, object alignment, insertion, tool engagement.
- Co-training with robot data when hand motion is tracked. EgoMimic, using egocentric human demonstrations, reports 34-228% improvement over robot-only data and that an extra hour of human hand data was more valuable than an extra hour of robot data (Kareer et al., 2024). That study used head-worn glasses, not wrist cameras, so treat it as evidence for human hand data in general.
- Step segmentation and hand-object interaction labels, where the hand stays in frame; tracked hands feed 3D hand pose data.
Cannot support
- Contact forces or tactile state. Even instrumented systems often lack them: DexCap's portable mocap rig has no force sensing and a roughly 40-minute session limit from power (Wang et al., 2024).
- The scene context. A wrist view alone gives little global layout, so pair it with a head or external camera.
- Robot actions, unless the device pose defines them (handheld gripper) or you retarget tracked hands.
- A match to a humanoid's own wrist cameras if the humanoid's mounting height, field of view and finger occlusion differ. Check against matching camera viewpoint to humanoid sensors.
Illustrative capture specification
# Illustrative only - parameters are examples the buyer must set
capture:
wrist_camera:
mount: rigid strap or device bracket; offset from wrist joint recorded in mm
lens: fisheye or wide, intrinsics calibrated per unit (checkerboard, before each day)
resolution_fps: 1080p, constant 60 fps
exposure: fixed shutter cap to limit motion blur; auto-ISO allowed
companion_views:
head_camera: yes (scene context)
external_camera: optional, fixed
pose_tracking:
method: IMU-aided SLAM or marker-based tracking
output: 6-DoF camera pose per frame, timestamped
synchronization:
shared_clock: all cameras and pose streams; sync pulse at start and end of take
gripper_device: optional (handheld gripper; aperture logged per frame)
labels:
per_take: task_id, object_set, operator_A12, site_017, outcome
per_step: start/end, step_name
hand_visibility: per-second flag
acceptance:
hand_in_wrist_view: buyer-set share of frames during graded steps
pose_dropout: buyer-set maximum gap length
sync_error: buyer-set maximum offset between streams
excluded: forces, tactile, robot joint states
Every number shown is an example, not a recommendation. Frame rate and exposure depend on task speed, and tracking method depends on the lighting and the clutter on site.
Parameters that vary by architecture
- Gripper-like device or bare hand. A parallel-jaw robot suggests a handheld gripper interface; a dexterous humanoid hand suggests tracked bare hands with a separate retargeting plan.
- Pose source. Visual SLAM, IMU, optical markers or hand-tracking gloves each fail differently; ask for the dropout rate.
- Lens and field of view. Match the robot's wrist camera, or deliberately record wider and crop later.
- Number of wrists. Bimanual tasks need two cameras and one shared clock; see bimanual task demonstrations. Combining the wrist view with head and fixed cameras is a rig question covered in multi-view synchronized task recording; the first- versus third-person tradeoff is in egocentric vs exocentric workplace video.
- Glove use. Workplace gloves change hand appearance and tracking; see gloved hands and PPE in workplace video.
Acceptance tests to run on a pilot
- Replay a take with the pose overlay and check drift against a known fixed object.
- Re-run camera calibration at the end of the day; compare intrinsics to the start-of-day values.
- Verify that hands stay in the wrist view during contact steps.
- Confirm a sync event is visible in every stream at the start and end of each take.
- Check motion blur on the fastest steps by sampling frames.
- Confirm that demonstrators are not identifiable in clips unless consent says so.
Scope these in a pilot collection before committing to volume.
Permissions and provenance
A wrist camera records the hand and surroundings, so it captures bystanders, screens and products as well. Ask the site what is in the field of view, who consented, and whether demonstrators are employees. Commissioned collection can design consent in, as covered in recruiting and briefing demonstrators. Blurring does not resolve permission questions by itself.
Questions to ask collection teams
- Which mount and lens are you proposing, and how do they compare with our robot's wrist camera?
- How is camera pose recovered, and what is the dropout rate on a sample?
- How are streams synchronized, and what offset have you measured?
- How do you handle gloves, tools that block the lens, and fast motion?
- Can you deliver calibration files per unit and per day?
- Which tasks have you captured with a wrist view before, and can you share a sample clip?
- What is the demonstrator's gripper or hand interface, and how is it logged?
Discuss your requirements
If you need wrist-view demonstrations for a defined task, describe the camera, tracking and task requirements to SourceX. If no existing asset fits, SourceX can scope a commissioned collection of new recordings; SourceX's role is sourcing, rights review, licensing and delivery coordination, with specialist capture partners engaged case by case. A request does not guarantee a match. See the humanoid training data guide and the workplace task demonstration hub; the catalog entry on first-person video of skilled manual work describes related footage.
Describe the workplace demonstration data your robotics project needs
Sources
- Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots — handheld gripper with wrist camera and SLAM tracking.
- DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset — robot-side wrist plus exterior camera rig and per-episode calibration.
- EgoMimic: Scaling Imitation Learning via Egocentric Video — co-training human hand data with robot data.
- DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation — portable mocap limits, power and no force sensing.