The Event Camera Reacted After Four Centimeters

REACT is a fully spiking state-space model that updates on each raw event rather than collecting events into frames or bins. On the EvTTC time-to-collision benchmark it reached 9.59% relative error with 4.6 milliseconds of end-to-end latency, within 0.15 percentage points of the best learned comparison while using no target prior. At the dataset’s mean approach speed, that latency corresponds to four centimeters of vehicle travel rather than one meter for the fastest competing learned method. INT8 quantization reduced estimated energy per 32,768 events from 18.5 to 2.8 millijoules.