By ,
Zoox employees enter a self-driving robotaxi as it drives between two Zoox headquarters in Foster City in July 2024.
Jessica Christian/The Chronicle
The light turned green at a busy intersection in San Francisco’s South of Market neighborhood, where Bryant, Division and 11th streets conjoin at acute angles — and traffic becomes an omnidirectional snarl.
What happened next was captured, probably in stark detail, by multiple cameras on a Zoox robotaxi. The company is not required to publicly share its footage, and has declined to do so.
But in a report filed to the Department of Motor Vehicles, Zoox gave a clinical description of the crash on April 4: A driverless taxi collided with a Bay Wheels e-bike whose rider had swerved into the car’s path. The Zoox sustained fender damage while the cyclist rode away. Nobody called police.
Though apparently minor, the SoMa fender-bender came at an inflection point for Zoox, an autonomous vehicle company owned by Amazon. Widely recognized as the No. 2 player behind Waymo, Zoox is testing its buggy-shaped robotaxis and retrofitted Toyota Highlanders in the wilderness of downtown traffic, and preparing to launch passenger service. It’s also endured some growing pains.
Four days after the e-bike jolt, another unoccupied Zoox vehicle collided with a passenger car in Las Vegas. Again, no one emerged with injuries, but the company opted to pause operations in Vegas for several days and recall 270 vehicles for a software update.
Such episodes could become more commonplace as self-driving cars flood the roads, treating cities like San Francisco not only as markets, but as laboratories. While the makers of these vehicles contend that a machine programmed to follow traffic laws is safer than an erratic human driver, the crash research is less conclusive. The stakes for these companies are high: Waymo, the most established among them, is rapidly expanding commercial service as Zoox test-runs its carriages in several cities.
As they integrate into the transportation system, robotaxis are drawing excitement and delight. At the same time, they’re clashing with other road users. Last year a self-driving Waymo car struck and injured a bicyclist in San Francisco’s Potrero Hill neighborhood, provoking anger from district Supervisor Shamann Walton. He questioned whether any car operated by artificial intelligence could quickly react to traffic chaos. Representatives of Waymo said an oncoming truck had occluded the Waymo vehicle’s view of the cyclist.
A review of California DMV reports for March and April suggests that so far, most of these incidents have not caused injuries, and have resulted in minimal damage. Still, they raise a vexing question for the average San Franciscan: Should I be worried a robot car might hit me?
The notion has incited a spirited debate among experts.
William Riggs, a professor of engineering and management at the University of San Francisco who studies autonomous vehicles, is so confident in the technology that he doubts they pose much danger to other drivers or cyclists — even the ones who violate traffic laws.
“You probably have a greater chance of being struck by lightning” than by a robotaxi, Riggs said.
Other researchers are less certain. Scott Moura, a UC Berkeley professor of civil and environmental engineering, said past studies have shown a similar “crashes-per-million miles” rate between self-driving cars and humans, though data also suggests that AV companies are learning from their experiences, and swiftly course-correcting.
In California, AVs are held to a higher standard than human drivers, with a requirement to report any collision that results in injury, death or property damage to the Department of Motor Vehicles and, in some instances, to the National Highway Traffic Safety Administration. Yet since the companies write those reports, they are by definition one-sided and could omit details.
The most notorious example is the crash of Oct. 2, 2023, in which a robotaxi operated by General Motors’ Cruise struck a jaywalking pedestrian. Although initiated by a hit-and-run driver, the incident led to Cruise’s suspension, largely because the company withheld video from state investigators.
When pressed about the April 4 collision with the cyclist in SoMa, staff at Zoox point back to their report to the DMV. It portrays two vehicles navigating a complex, unpredictable road space with crosswalks and traffic signals on all sides.
The scene unfolded shortly after 9:30 a.m., the fade-out of a Friday morning rush hour. According to Zoox, an e-bike rider had started crossing before the change in signal. Pedaling south down 11th Street, the rider zipped ahead of the robotaxi, which had no driver at the time. Both vehicles had nearly crossed the intersection when the cyclist abruptly turned left.
Sensing the maneuver, the robotaxi “braked hard,” yet a collision “was unavoidable,” Zoox staff wrote in their incident summary. The cyclist struck the self-driving car’s right fender and body, causing minor damage. As the Zoox pulled over on Bryant Street, the cyclist rode to the sidewalk, briefly stopped, and then rolled away.
By slowing down and stopping, the Zoox demonstrated its ability to drive defensively, and likely preempted a more serious accident, according to a company spokesperson.
“Autonomous vehicles have the capability to see farther and understand the speed and trajectory of those sharing the road, which is especially important when interacting with vulnerable road users,” the spokesperson wrote in a statement.
It continued: “Safety is foundational at Zoox, and we are continuously learning from our testing across markets to improve the overall safety of roadways.”
Whether the Zoox performed worse or better than a human driver in this instance is unclear, said Matthew Raifman, a transportation safety researcher at UC Berkeley’s Safe Transportation and Research Education Center.
“There are scenarios where an AV might see (a cyclist) that a human wouldn’t see, because they have 360 degree views,” Raifman said. Even so, he noted, an AV doesn’t have a human’s intuition or ability to anticipate behavior. For all of its perspectives and sensors, a robot car might not understand how a cyclist moves through space.
“I think that’s where there’s a heuristic process that human drivers have. Maybe a computer can replicate that perception and understanding, but it’s also possible that a human driver would have eye contact with the cyclist, or detect something in that cyclist that was indicative they were going to turn instead of go straight.”
Or alternatively, perhaps the robotaxis could over time become more perceptive than humans, Raifman said. What if, for instance, a network of them are passing through that intersection every day, collecting data, and sharing it with all of the other robotaxis in a company’s fleet?
Steven Shladover, a research engineer at UC Berkeley’s Institute of Transportation Studies, had a more cut-and-dry reading of the incident.
“It sounds like this was probably the fault of the cyclist,” he said, surmising that a human motorist wouldn’t have been able to avoid the collision either.
Every expert who spoke with the Chronicle acknowledged the basic math of exposure: More autonomous vehicles traveling more miles means more opportunities to crash. Yet the companies are making a case that their fleets will eventually make the roads safer. After all, a robot won’t drive while intoxicated, or start texting friends at a stop light, or get distracted and fail to yield to a pedestrian.
Robotaxis “are getting safer over time as the companies gain more experience with driving in traffic,” Shladover said. “But that’s a long, slow process.”