Sensor fusion is the practice of combining several different kinds of sensor so that a self-driving system can perceive its surroundings more reliably than any one sensor allows. The three workhorses are cameras, which see color, text, and lane markings; lidar, which measures precise 3D distance and shape; and radar, which detects velocity and works through rain, fog, and glare. Each has weaknesses the others cover, and the fusion approach is to let them vote together.
Waymo’s design philosophy states the idea plainly: no single sensor type is enough, so it engineers complementary sensors that “work as a unified system.” Its fifth-generation hardware paired a 360-degree lidar with perimeter lidars, long-range and peripheral cameras, and “one of the world’s first imaging radar” systems able to flag objects that are “moving, barely moving, or stopped.” A pedestrian half-seen by a camera at dusk may be cleanly ranged by lidar; a stopped truck that confuses radar may be obvious to vision. Public datasets like nuScenes, which ships 6 cameras, 5 radars, and a lidar, exist precisely so researchers can develop and test fusion methods.
The counter-position, taken by camera-first companies, is that fusion adds cost and complexity and that vision alone, with enough learning, can suffice. So sensor fusion sits at the heart of the field’s central economic and safety tradeoff.
For a general reader, sensor fusion captures a basic engineering principle that extends well beyond cars: when no single measurement is trustworthy on its own, combining independent, differently-flawed sources is how you build a system that is reliable as a whole.