What Is a Wrist Camera in Robot Learning?

A head camera captures the wide egocentric view of a task — the environment, the target object, and the approach — but as the hand closes in on an object, the head camera’s view of the actual contact point is frequently occluded by the hand and arm. A wrist camera solves exactly that problem: it stays close to the object through the final phase of a reach and grasp, capturing contact detail the head view has already lost. Most manipulation policies need both viewpoints together, not one instead of the other.

Firsthand captures two wrist cameras, left and right, alongside its head-mounted global-shutter 4K camera. All three views — head and both wrists — are hardware-triggered onto the same clock as the depth and IMU streams, and the head↔wrist↔depth transforms are solved against a shared calibration target and expressed in the head-camera frame. That means wrist footage is not just temporally aligned with the head view; it is geometrically related to it, so a policy can reason about where the wrist camera is relative to the head camera at every timestamp.

What is a wrist camera?

A wrist camera is a camera mounted at or near the wrist that moves with the hand, keeping the grasp point in frame during fine manipulation rather than relying solely on a head-mounted or third-person view.

Why do robot learning datasets need a wrist camera?

As a hand approaches and grasps an object, a head-mounted camera’s view of the contact point is frequently blocked by the hand and arm. A wrist camera stays close to the object through that final phase, capturing the contact detail the head view has already lost.

Is a wrist camera a replacement for a head camera?

No. A head camera provides the wide egocentric view of the environment, the target object, and the approach; a wrist camera provides a close, occlusion-resistant view of the grasp point. Most manipulation policies need both viewpoints together, not one instead of the other.

How is wrist camera data different from head camera data?

A head camera covers the whole task and environment from a fixed vantage point near the head. A wrist camera moves with the hand and only covers what is near it, so it keeps contact detail in frame during fine manipulation but loses scene context outside its narrow field of view.

How are wrist cameras synchronized with other streams?

Firsthand hardware-triggers its wrist cameras onto the same clock as the head camera, depth, and IMU streams, rather than aligning separately recorded footage after the fact.

How are wrist cameras calibrated relative to the head camera?

The head↔wrist↔depth transforms are solved against a shared calibration target and expressed in the head-camera frame, so the wrist view is geometrically related to the head view, not just temporally aligned with it.

How many wrist cameras does a robot learning dataset typically need?

Firsthand captures two wrist cameras, one for each hand, alongside its head-mounted camera, since most manipulation tasks involve both hands.

When should a custom data collection include wrist cameras?

Any collection involving fine manipulation, in-hand object handling, or grasp precision benefits from wrist cameras, since the head camera alone loses the contact point exactly when the hand closes in on the object.