Waymo's Rosenband: camera-only self-driving stalls before 99.99%

Waymo's keynote at Hot Chips was framed less around a product launch than around the reasoning behind its Sensor Fusion Processor. Daniel Rosenband used a 2017 example in which a red balloon floating in front of a traffic light was misread as a red signal, and noted that even Gemini cannot reliably tell the two apart. That sets the bar for autonomous driving: matching or exceeding human perception remains an unmet capability, and by 2017 the required capability was already estimated at 50 times what it had been in 2012.

The talk also put Waymo's operating record in context. The company says it has reached its sixth vehicle generation and, by 2020, had driven over 200 million miles (320 million km), completed more than 20 million paid rides, and covered 15 cities. Rosenband's caution is that this represents only the first 1% of the problem; a product is defined by the long tail of remaining cases, where each additional "9" must be earned rather than accumulated through more of the same driving.

That view explains Waymo's sensor choice. Camera-only systems are the lowest-cost route, but Rosenband argued their gains flatten before the final steps, while adding LiDAR, radar and ultrasonic sensors brings extra parts, wiring, weight and failure risk. Waymo has evidently judged the accuracy payoff worth those costs. The company has separately said it will launch a fully driverless ride-hailing service in Japan in 2027 with Nihon Kotsu and GO.