Textbook of AI

The Waymo Driver is the autonomous-driving system developed by Waymo (originally Google's self-driving project, spun out as an Alphabet subsidiary in 2016) and the only Level 4 robotaxi service operating commercially at meaningful scale. As of 2026 Waymo One offers fully driverless rides in Phoenix, San Francisco, Los Angeles and Austin, and has logged tens of millions of rider-only autonomous miles on public roads.

Architecturally Waymo follows a multi-sensor, modular philosophy in deliberate contrast to Tesla's vision-only approach. The current 5th-generation hardware suite combines five lidars (one 360° rooftop, four short-range corner units), 29 cameras providing surround coverage at multiple resolutions and dynamic ranges, and six imaging radars. Sensor fusion is performed in a unified 3D representation rather than at decision level, and the redundancy across modalities is treated as a safety property: a sun-blinded camera, a rain-attenuated lidar return and a snow-occluded radar can still produce a reliable joint percept.

The software stack is broken into recognisable modules: perception (object detection, tracking, semantic segmentation across modalities); behaviour prediction (trajectory forecasting for every agent, typically multi-modal with associated probabilities); planning (selecting a trajectory subject to comfort, legality and safety constraints); and control (low-level actuation). Each module is a learned neural network , Waymo has published extensively on VectorNet for HD-map encoding, MultiPath for multi-modal trajectory prediction, and MotionLM which casts trajectory prediction as language modelling over discretised motion tokens, but the interfaces between modules are explicit, interpretable and amenable to formal verification, in deliberate contrast to end-to-end policies.

Waymo invests heavily in simulation and structured offline evaluation. Their Carcraft simulator replays and perturbs every event from real driving, plus synthetic edge cases mined from a taxonomy of "interesting" scenarios. Each software release is gated on tens of millions of simulated miles plus targeted closed-course testing before any change is deployed to the rider fleet. The published Waymo Open Dataset (lidar, camera, motion forecasting, end-to-end driving subsets) became a primary academic benchmark.

Waymo's modular philosophy has clear strengths: components can be tested, interpreted and certified in isolation; sensor redundancy provides graceful degradation; the engineering teams can be partitioned. The trade-off is engineering scale, the codebase is enormous, behaviours emerge from interactions across modules, and improvements to one module can degrade others. The contrast with Tesla FSD is the cleanest controlled experiment in modern AV: same problem, opposite stacks, opposite operating envelopes (Waymo: geo-fenced, fully autonomous, expensive sensor suite; Tesla: ubiquitous, supervised, cheap sensor suite). Both approaches have been productive; which scales further remains an open empirical question.