Identify the cue FloWM uses to infer motion after a single sighting

Identify the cue that FloWM uses to infer a block’s direction from a single sighting during long periods without observations.

Background

The study compares FloWM’s learned grid memory with a Kalman-tracker node memory in Dynamic Block World, where blocks may remain unobserved during a 70-step prediction phase. The tracker cannot infer velocity from a block observed only once and therefore performs poorly on such cases.

FloWM nevertheless predicts the direction of many blocks sighted only once: among answers landing on one of the four possible direction cells, 89% select the true direction, compared with 25% expected from guessing. The authors report that the underlying cue or signal enabling this behavior has not been determined.