Waymo, the company’s driverless car services, is rapidly broadening its fleet far more quickly than ever before, and while it now has nearly 4000 autonomous vehicles functioning on the roads of the U. S. the rapid growth that has led to this status is also constantly revealing an intriguing new array of unexpected issues that are notoriously difficult to plan for. The corporation currently us present in 15 U. S.
metropolitan areas (others have been launching totally driverless services in places including Denver, Las Vegas, San Diego and Tampa), and an average week is getting half a million paid trips completed. This is amazing, but the outrunning edge cases are not. These edge cases are edge cases because they are the rare circumstances that many software algorithms still can’t handle. In the last couple of months Waymo’s cars have encountered flooded streets, gone into busy freeway construction zones at high speed, frozen due to power failures and been unable to move from intersections on popular public holidays.
A large July Fourth fireworks event in San Francisco had over 100,000 visitors leaving the waterfront stalls clogging many of the streets that supply the city while driverless cars added to the the road congestion by bringing many major arteries to a halt. In Miami Beach, two robotaxis had a few deadlock at an intersection for some twenty minutes, creating miles of backup, until the machines stopped fighting one another and decided to budge. Last December, a 2-minute power outage near San Francisco paralyzed nearly 1,600 Waymo’s. Federal authorities have been paying attention.
Since last December, Waymo has recalled its software several times, for campaigns related to submerged roads and freeway construction practices. It suspended its highway service for weeks following repeated instances where vehicles entered closed work zones, then gradually reintroduced freeway rides following software upgrades. On surface-streets, rides persisted in most cities, but the disruption revealed how swiftly one class of issue can cascade through a national fleet. Based on Waymo, the incidents are declining in frequency even as the fleet and the cities expand.
Execs view the recalls as preventative and keep informing that each odd moment contributes to an ongoing improvement in the systems. Leaders mention that they speed up city openings, and the system continues to learn from real-world driving. If we look purely in numbers the argument has some validity. Half a million trips a week result in vast quantities of data. Most trips conclude without a hitch. The failures But, are the ones that make the social media or local media. The dash of human life can break it all down. Passing views differ across states, weather, and construction styles and personal comportment. A system heavily trained in one configuration can’t reasonably be expected to have also encountered all of the randomly possible interactions of these big variables with another city’s unique characteristics.
Flood level was/is hard to get right. Temporary construction signage risks dissonance with lane-mapped structures. Packed crowds or abrupt blackouts result in anomalies that don’t happen in training. As the fleet grows, the absolute number of anomalies grows. Riders’ experiences are varied. In mature cities, they have come to see Waymo as a comfort, dependable for the quotidian ride. In new ones they are still testing it, and a freeze that comes right when you need to turn or a miscalculation that leads you miles down the road can chip away at your trust. Municipalities are now left to handle the fallout when robotaxis block fire exits or cause local traffic snarls.
