Perception sees the world; localization answers where the car sits in it — to within ~10 cm, ten times a second. The whole industry splits on one decision: is the map already known (a pre-built HD map) or built live from sensors right now? This guide shows both, the sensors that feed them, and the real algorithms Waymo, Tesla and Mobileye run today — each with a step-by-step player.
This is the defining question. Most robotaxis localize against a centimetre-accurate map built weeks earlier; the live sensors mainly answer "where on that known map am I?" The newer "mapless" camp throws the prior map away and reconstructs a lightweight map every frame. Almost everyone actually lives on a spectrum between the two.
An HD map — a centimetre-accurate 3D point cloud plus vectorized lanes, signs and traffic lights — is built ahead of time by survey vehicles. On the road, the car matches live sensor data against this prior to pin down its pose. Localization, not mapping, is the live job.
No heavy prior. The car runs SLAM-style online mapping: it builds a fresh, lightweight vector map of lanes and road edges from cameras (sometimes + radar) every frame, and localizes within that. Maps may be crowd-sourced from the fleet and merged in the cloud.
No single sensor is enough; each covers another's blind spot. Flip them on and off to see what the car "sees." This is why fusion exists — and why the Tesla-vs-Waymo debate is really an argument about which of these you can drop.
Read it as two questions per row: does it need a prior map or work live, and what sensor drives it. The top rows are today's robotaxi workhorses; the bottom rows are where the field is heading.
| Method | Map | Primary sensor | What it does | Typical accuracy | Who uses it |
|---|
There is no single winner; the frontier is a spectrum, and the industry is quietly converging toward the middle.
Every currently-operating driverless robotaxi (Waymo, Baidu, Pony, WeRide) uses a pre-built HD map plus LiDAR scan-matching (NDT / GICP / reflectivity) fused with GNSS+IMU. It is the only stack with a proven driverless safety record as of 2026.
Tesla's vision-only FSD and Mobileye's REM crowd-mapping reduce or remove the heavy prior, trading some accuracy for the ability to expand anywhere cheaply. The clear trajectory is toward learned online maps (MapTR-style) and ultimately world models where the "map" lives inside a neural net.
The honest 2026 summary: pre-built HD maps are treated as a prior, not a crutch — even Waymo says its cars drive on live sensors and can handle a map that's gone stale. Meanwhile the mapless camp keeps borrowing classic SLAM ideas (online vector maps are just lightweight SLAM). The two ends are meeting in the middle: a light prior + strong online mapping + tight multi-sensor fusion.