Skip to content

Resources

Head and wrist cameras: what each view adds

Plan complementary egocentric viewpoints without mistaking moving head/wrist cameras for a fixed stereo pair.

Published 2026-10-27 · 6 min read

People moving through a city crossing in the evening.
Illustrative multi-camera capture layoutTwo head cameras form a stereo pair only when their geometry and timing are validated. Separate wrist cameras add views. Each stream links to the same episode, timestamps and calibration references.Head leftcamera 01Head rightcamera 02Wrist leftcamera 03Wrist rightcamera 04Validated stereo geometry + time relationshipOne episode · distinct streams · timing and calibration references
Planning example, not a Fiund rig or available dataset. Two cameras alone do not establish calibrated stereo.
Planning comparison — verify against the actual capture package
View or relationshipWhat to inspect
HeadScene context, head turns and visibility of both hands
Left/right wristLocal action visibility and changing occlusion
Head to wristDynamic relationship; tracking if a shared frame is required
All streamsCommon episode IDs, timing evidence and missing-view policy

A head camera follows the wearer’s head, while a wrist camera moves with the hand or attached device. They can provide complementary context, but neither guarantees that a grasp, contact point or target object remains visible throughout a task.

Design around the action you need to observe. Ask which view will show the approach, manipulation and result, then test the arrangement with realistic movement rather than judging a stationary demonstration.

Head views provide context, with their own occlusions

Forested mountain ridges receding into lavender evening mist.

A head view may include the scene, both hands and the destination of an action. The operator can look away, an arm can block the view and a close manipulation can fall outside the camera’s useful coverage. Document actual placement rather than using head-mounted as a completeness claim.

Run representative actions while inspecting the recording. Look at reach, turn, crouch and close-work phases. Check the entire sequence, not only its final frame. A comfortable, safe mount and clear site permissions are part of planning before longer collection.

Wrist views follow local interaction

A wrist view can provide a closer observation near manipulation, but fingers, a tool or the grasped object can obscure the contact itself. The useful viewing direction changes with wrist rotation. Record left/right roles, attachment and the view’s limitations.

The UMI project is a specific demonstration framework using handheld grippers and an associated policy interface. A wrist video alone should not be relabeled UMI data or a robot-action demonstration. Say whether trajectories, gripper state or other signals exist, and how they were obtained.

Moving cameras need changing relationships

A rigid stereo pair has a fixed relative camera relationship while the rig remains intact. A head and a separately moving wrist do not generally share that property. One initial extrinsic calibration cannot describe every later head-to-wrist pose.

If a buyer needs views expressed in one coordinate frame, state the tracking method, timestamps, transform convention and uncertainty. If that information is unavailable, offer synchronized multiple views as such rather than claiming a calibrated fixed-baseline pair.

A producer reviewing conversation recordings and audio tracks.

Plan synchronization and stream recovery

Create stable IDs for each camera and session. Describe capture clocks, synchronization or alignment method, dropped-frame handling and measured drift. Keep original timestamps and explain any corrected timeline. Two exports of equal length do not prove that they align.

Agree what happens when one wrist view fails or disappears during an otherwise useful session. A project may accept a reduced-view episode, reject it or request recapture. Write this rule before scaling so recorded hours are not mistaken for accepted multi-view hours.

Build a view-coverage pilot

Select actions that stress the setup: occluded grasps, crossed hands, reaching outside the table and transitions between locations. These are suggested test cases, not a universal acceptance standard. Review the actual use with the buyer.

Deliver paired clips, a camera-role manifest, timing notes and available tracking information. Count unique task duration separately from head and wrist stream-hours. Fiund assesses existing material or collection capability; a submission is not an equipment recommendation or open paid assignment.

Sources

← All resources

Frequently asked questions

Are head plus two wrists a stereo dataset?

Not automatically. They are multiple viewpoints. A stereo claim needs the relevant geometric and timing evidence for the intended pair.

Does a wrist camera provide robot actions?

Not by itself. Identify any trajectory, gripper or action signals and their source separately.

Related resources