LIDAR, Radar and Camera-Only: The Three Sensor Philosophies

Three sensor philosophies compete for space on a self-driving car's roof and bumpers, and none of them has won outright.

Lidar is the most expensive of the three by a wide margin. It fires laser pulses and times how long each one takes to bounce back, which gives a direct, physical measurement of distance rather than an inference. The output is a 3D point cloud: a dense scatter of coordinates describing exactly where every surface around the vehicle sits in space. That directness is the appeal, and the cost is a large part of why it took years for lidar hardware to come down to a price commercial fleets could absorb.

Radar sits at the opposite end of the cost curve. It's comparatively cheap, and it keeps working in conditions that degrade the other two sensors, rain and fog especially, because radio waves punch through weather that scatters light. The tradeoff is resolution. A radar return tells you something is out there and roughly how fast it's moving, but not much about its shape, which is a poor substitute for a point cloud when the question is exactly where an object's edges are.

Camera-only approaches skip a dedicated distance sensor entirely and ask software to do the work instead: infer depth, classify objects, estimate how far away a shape in a 2D image actually is. This shifts cost out of the hardware bill and into the software problem, which is harder to solve and harder to verify from the outside than a physical distance measurement. Mobileye, a chip and perception supplier that sits in the D20 index, has built much of its business case around exactly this bet on vision-first perception.

The disagreement isn't only about price; it's about what happens when a system gets something wrong. A lidar-based error tends to be a measurement error, a return that was occluded or misclassified, but the underlying distance was still measured directly rather than inferred. A camera-only error can be a purer inference failure: the system judged a shape to be well beyond where it actually sat, with nothing in the pipeline that measured distance rather than estimated it from pixels. That distinction matters more to a safety case than it does to a spec sheet, since a measurement error and an inference error tend to get investigated, and weighed, differently after the fact even when the on-road outcome looks identical.

Cost has not stood still either. Lidar hardware has grown cheaper as production volumes rise and mechanical, spinning designs give way to solid-state units with fewer moving parts to manufacture and calibrate. That shift narrows the price gap without settling the underlying argument, because the disagreement was never purely about what a single unit costs on a bill of materials. A company that has already built years of perception software around inferring depth from images has little institutional incentive to bolt on a sensor it spent that same time arguing it doesn't need, regardless of how far the sticker price has fallen since.

In practice, almost no commercial operator picks one lane and stays in it. Most combine two or three sensor types and let each one's weaknesses get covered by the others, a process covered in more detail in the explanation of how sensor fusion actually works. Whether any single philosophy eventually pulls ahead as volumes rise, or the blended approach simply becomes the permanent default, is still an open question.