Latency from head motion to photon update must be extremely low

What it is

The head has already turned to a new heading, the photons are still painting the old wall, and the world has slipped relative to the skull. The interval from the start of head motion to the matching photons reaching the eye is motion-to-photon latency (MTP). In a headset it has to be extremely low — industry often parks the acceptable ceiling near twenty milliseconds, and stricter rotational tasks push the target into the low teens. Fifty milliseconds of click feedback on a desktop can still pass as “responsive enough.” The same order of magnitude on a head turn and the room is already swimming.

“Extremely low” is relative to motion the vestibule and proprioception have already reported. The headset is maintaining a stationary world, not a pointer that follows the hand.

Why it happens

On a head turn the semicircular canals report angular velocity almost at once, and neck proprioception arrives with it. If vision is late, the two bodies of evidence for “I turned” disagree: the vestibule says the world should have swept the retina in the opposite direction, and the actual sweep arrives a beat later. People treat the world as a stable reference, so the delay is read as the world moving, not as the head moving. Once the stable reference loosens, stance and spatial orientation have to be re-estimated.

A headset pastes the field onto the face; there is no monitor bezel to serve as an excuse that “this is a screen.” Latency therefore has nowhere to hide: every degree of head motion has to be honoured across the whole retina at once. Pointer delay can be patched by predicting a click; world delay has no discrete “click,” it is a continuous spatial promise. So the same millisecond count that passes in a GUI does not pass on a head turn.

Studying it

Inject controlled delay: sample head pose on time, hold back rendering or display by a known number of milliseconds, and have people judge whether the world slides during a head turn, or run a pointing / alignment task. The Adelstein–Allison–Jerald line measures detection and operational thresholds, not mean frame time.

Independent variables: injected MTP in milliseconds, head angular velocity, density of near structure in the scene. Dependent variables: whether “the world is sliding” can be detected, pointing error, subjective “stuck to the head / dragging behind.”

Measurement has to run from real head motion to real photons, with a high-speed camera on the headset image and on markers on the head. Reading only engine frame cost drops display persistence and scan time, and the reported number comes out low.

Where it stops holding

Seated in front of a fixed picture, with the head almost still, MTP is not sampled, and even high latency can “look fine”; that task cannot sign off a headset. The same delay is harder to detect on a slow smooth turn than on a quick nod, so acceptance should use the fast fragments that actually occur, not only the slow booth turn. Video see-through stacks a camera delay that optical see-through does not have; “extremely low” is not the same budget on the two machines. Prediction (extrapolating from current angular velocity) can push effective latency down for a short time, but it depends on motion being predictable; a sudden stop or a reverse adds the wrong increment, so prediction is not a durable substitute for low latency. Thresholds move with person and task; twenty milliseconds is not a physiological constant, it is a commonly used engineering ceiling.

Applying it

• Write MTP as a hard budget for the headset experience, not as a quality item among visual effects. However refined the content, if the world drags on a head turn the budget was not kept.
• Accept with quick nods and sharp left–right turns, not with slow exhibition pans. A scene with sharp near edges exposes delay better than a skybox.
• Do not substitute engine frame time for MTP. The acceptance number comes from motion-to-photon footage or from a platform MTP counter, and the source goes in the report.
• How to check: three short nods on the target device, asked of someone who did not build the project: does the world wobble with the nod. If they can point to “the wall is chasing my head,” latency is not in the “extremely low” band — push the milliseconds down before talking about other image quality.

Related

• Same group: N1.05.2 Suprathreshold latency directly causes discomfort · N1.05.3 This latency cannot be covered up with animation
• Nearby: N1.11 Render Latency and Motion-to-Photon Time · N1.12 Refresh Rate and Persistence
• Search terms: motion-to-photon latency · head-motion lag · latency detection threshold