Bystanders cannot predict the wearer's next movement

What it is

Someone walking past cannot see the virtual cup the wearer is reaching for, and so cannot give way the way they would for a hot mug. Kinematic intent is readable in ordinary collision avoidance from gaze, preparatory posture, and a target both people can see. The headset splits those signals from the room. A failed prediction costs a collision, not an odd-looking gesture.

Why it happens

People dodge others on a few hundred milliseconds of trajectory preview: eyes go where the body will go, the shoulder turns, the arm extends toward a target both can see. The bystander reads that and steps off early. The wearer's gaze is pinned to a virtual object; head and hand point at nothing in the room — to the bystander, a sudden swing toward an empty table. Joint attention is absent, the preview window closes, and avoidance becomes a reaction, usually too late.

Preparatory motion itself lies. To tap a virtual board the person may sidestep, lean back, then throw a hand up, none of which maps onto reaching for a real object. A bystander inverting from real objects guesses the wrong direction and amplitude. In a shared corridor the unworn party carries the whole prediction load and is denied the target the prediction needs.

Studying it

A wearer reaches or swings at a virtual target while an unworn bystander walks past at the side or from behind. Compare whether the bystander can see the virtual target (shared see-through versus none). Log when the bystander starts to dodge and whether contact happens.

Independent variables: bystander can see the wearer's target, approach direction (front / side / behind), motion amplitude. Dependent variables: bystander dodge latency, collision or near-miss, error when the bystander points at “where the next move goes.”

Ask the bystander, before the motion starts, to point at where it will go. A wrong point is unread intent; you do not have to wait for a hit.

Where it stops holding

Two people in headsets sharing a world frame, each able to see the other's hands and targets, can recover some preview — that is collaboration, not this leaf's bystander condition. If the bystander stays outside the tracked volume and never enters the sweep, failed prediction does not become a collision. Children and pets do not read adult shoulders, and they cannot read a virtual target; the prediction window is near zero. Behind glass, or at a distance, collision risk drops and this safety issue leaves.

Applying it

• When someone may walk into the operating volume, shrink large reaches and give the bystander a preparatory shape readable from the room: hands to the chest first, not a swing out of peripheral vision.
• Do not park targets that require a sudden sidestep or lean-back in a doorway or corridor, where the next move is unreadable in the room.
• If the system can detect a second person inside an arm's length, pause motions that leave the body envelope. Do not rely on that person yielding.
• How to check: have someone without a headset walk past a player from the side and guess “where does the next move go.” Wrong guesses that also graze mean the intent channel is closed to the bystander — shrink or pause the motion, rather than asking the bystander to adapt.

Related

• Same group: N2.13.1 Large gestures hit furniture and people the wearer cannot see · N2.13.2 With vision occluded, estimates of one's own position drift · N2.13.4 Banning dangerous actions in the content is more effective than after-the-fact warnings
• Nearby: N4.04 Room Scale and Safety Boundaries · N5.06 Bystanders
• Search terms: kinematic intent · action anticipation · collision avoidance