The Road Is Reading
You Back

What the Machine Sees at Midnight

Inside the sensor fusion pipeline that keeps a robotaxi oriented during a Category 1 rainstorm

At 11:47 PM on a wet Tuesday in Phoenix, the vehicle's front-facing cameras were functionally blind. Rain on the lens, sodium vapor halation from the overpass, a truck's brake lights smearing into abstract expressionism across the wet asphalt. The lidar was doing its job — returning 1.3 million points per second, constructing a faithful occupancy grid of the world — but the neural network responsible for reading lane markings had been trained on 40 million frames of dry California highway, and this was not California, and this was not dry.

What happened next is the reason Waypoint exists: the vehicle didn't stop. It didn't call for remote assistance. It fused the degraded camera signal with radar returns and the high-definition map's stored lane geometry, assigned confidence weights that the camera team would later describe as "embarrassingly low," and proceeded through the intersection at a speed its motion planner had calculated as appropriate for a situation it could not fully see. The passenger — a policy researcher from ASU, in the back seat with a laptop and a half-finished grant proposal — noticed nothing.…

Thirty-Seven Minutes Without a Driver

A first-person account of the cognitive vertigo that arrives about eight minutes into a driverless ride

The car arrived exactly on time, which felt important to notice. I had been told to sit in any seat, which I interpreted as an invitation and also a small test of my own comfort with ambiguity. I chose the back left — a position I'd never occupied in a taxi without a stranger in front of me — and felt immediately that I had made a category error, that the seat I had chosen was designed for a passenger accompanying someone, not for a passenger alone.

The first eight minutes were fine. The car drove competently through the Mission, stopped correctly at lights, yielded to a cyclist with what I can only describe as appropriate caution. I watched the in-app display show me what the vehicle was seeing: pedestrian icons blooming and fading, lane lines traced in confident blue. Then we entered the freeway on-ramp, and the car accelerated to 67 miles per hour in a way that felt — I am trying to be precise here — not aggressive, not smooth exactly, but algorithmic. The acceleration had no opinion about itself.…

Waymo One, Mission District to SFO. The empty front seat.