The pinch itself must not shift the gaze point
Aliases: commit disturbs aim · pinch steals gaze · eye-hand coupling shift
What it is
The eyes are already on the target. The fingers close. In that same frame the gaze jumps — to the fingers, or half a degree off with a nod. The pinch itself must not shift the gaze point: the commit channel is rewriting the pointing channel. The split is undone by its own execution.
This is not Midas (“looked at, therefore fired”). The target was right and the commit was intentional. What fails is that commit moved the aim, and the system recorded the point after the move.
Why it happens
A pinch is a distal action of the arm that still hits the eye along several paths. People glance at their own hand before confirming the fingers will actually meet — a voluntary saccade that leaves the target. Even if they refuse that glance, a pinch often rides with a small head or shoulder motion; the eye tracker’s frame on the headset moves with it, and the visual axis rotates a little in the world. Further down, the force itself sends a small whole-body displacement into the head, while the tracker is still using last frame’s calibration.
So the gaze sample at commit is not the aiming sample. If the system reads gaze on the pinch’s rising edge, it may catch a mid-saccade toward the fingers, or a neighbour key walked onto by head motion. It is the same coupling as a mouse click that dips the whole hand and takes the cursor with it; here the disturbed pointer is the eye. The finer the pointing, the more lethal the same shift: half a degree is nothing on a large button, and the wrong label in a tight row.
Studying it
Align the time series of pinch onset and gaze direction, and ask whether gaze leaves the target inside the commit window.
Independent variables: whether the pinch requires the hand in view, target visual angle and spacing, whether gaze is latched before the pinch, hand in the lap versus raised in front of the face. Dependent variables: angular error of gaze from target centre at commit, number of saccades toward the hand, neighbour-key error rate.
Contrast hiding the hand out of view with parking it near the target to separate “look at the hand” from “head motion.” A latch condition (gaze from 50–100 ms before pinch onset, not gaze on the rising edge) that cuts neighbour errors confirms the shift lives in the commit action, not in the aiming stage.
Where it stops holding
On large, widely spaced targets, a half-degree shift still lands in the right hit zone and the problem does not show. Good head-motion compensation in the eye tracker will ease the shoulder path; the look-at-hand path remains. A controller button instead of a pinch has shorter finger travel and less need to look at the hand, so shift shrinks — that is not evidence that pinch is already safe. Rapid repeated selection makes people look at the hand more often to keep the beat, and shift rate rises through the session. Participants instructed “stare at the target, don’t look at your hands” will understate field shift; nobody runs that instruction in real use.
Applying it
- Latch the gazed target before the pinch’s rising edge. Commit the latched value, not live gaze during the pinch.
- Make the pinch completable out of view, and do not teach “watch your own hand” as the correct move.
- Tight small targets should not assume gaze will hold to half a degree through a pinch. Widen spacing, or refuse new gaze jumps after latch.
- How to check: put eye tracking and pinch on one timeline. From 100 ms before commit to 50 ms after, every gaze departure from the target is a shift. If that curve still decides what was hit after latching, the latch is not in effect.