And now for the main event…

Five years ago, after the National Transportation Safety Board found Uber partially responsible for a fatal crash involving a self-driving car and a pedestrian in Arizona, the company temporarily shelved its fantasy of outfitting its fleet of vehicles with self-driving robotaxis. Now, the company is preparing to try again. Last week, Uber announced a partnership with Lucid, an electric vehicle maker, and Nuro, a self-driving tech company, to deploy some 20,000 self-driving vehicles, built explicitly for Uber’s ride-hail platform, in “dozens of markets around the world” over the next six years. A regulatory filing noted that production on these vehicles will begin next year. Meanwhile, Uber announced that the robotaxis will launch in a “major U.S. city next year.”

As part of the partnership, Uber will invest $300 million in Lucid and an undisclosed amount—somewhere in the “multi-hundred-millions”—in Nuro. When I asked Dr. Jesse Kirkpatrick, the co-director of Mason Autonomy and Robotics Center at George Mason University, about the scale and pace of the rollout, he told me that “ambitious is an understatement. It seems to me likely more of a pipe dream, to be honest.”

Indeed, the trouble with self-driving cars deployed at mass scale is that the technology remains vulnerable to the same problem plaguing all A.I. applications: reliability. In this case, however, the issue isn’t merely algorithmic reliability—it’s the reliability of the hardware that feeds those algorithms, too.

The Nuro Driver, for instance, features a suite of sensors that include multiple types of cameras, in addition to lidar, radar, and audio sensors. The software involves a series of overlapping A.I. models, each designed to enable autonomous mapping, perception, behavior, and controls. The details of Nuro’s algorithms (their type, size, energy intensity, etcetera) haven’t been disclosed, though the company has said they’ve been trained on a combination of real-world data gathered over the past decade or so, along with closed-course testing and simulations. Like all other self-driving car companies, Nuro employs teams of human teleoperators who can remotely take over a car if—or when—it gets into trouble, though researchers have pointed out that those remote operators are only genuinely useful if the cars operate at slow speeds. (According to Kirkpatrick, these teleoperators are the “dirty secret” of the industry.)

Those layers of technology are designed for redundancy, which is vital when it comes to safe self-driving. But there are still plenty of potential failure points. On the hardware side, unusual lighting or severe weather can disrupt the sensory data coming from the cameras; rain or moisture can disrupt or entirely disable lidar sensors; and radar systems can fail from environmental interference. Then, there are the well-known and documented (hallucinatory) flaws to the algorithms themselves.

Many of these failure points were on display during Tesla’s robotaxi launch last month. This brought to mind a conversation I once had with Missy Cummings, director of the aforementioned Mason Autonomy and Robotics Center, who noted that, for self-driving vehicles, hallucinations are “a guaranteed outcome. They just simply can’t not hallucinate.” This doesn’t mean robotaxis are unusable—only that they must be thoughtfully engineered to manage the inherent risks tied to all those failure points.

The dream of robotaxis, however, isn’t that they will merely be as safe as human driving—it’s that they’ll be much safer. But it’s unclear whether that will be achievable, largely due to the potential scale at which they’ll be deployed and the number of vehicles already on the road. In 2023, the National Highway Traffic and Safety Administration reported a fatality rate of 1.26 per 100 million vehicle miles traveled; there were a total of 3.19 trillion miles traveled that year, and just over 40,000 people died in car accidents. That was all spread across approximately 284 million cars and trucks that, on average, drove 11,000 miles each that year.

Waymo, the autonomous vehicle ride-hailing service already on the streets in Los Angeles, San Francisco, Phoenix, and Austin, recently surpassed a total of 100 million autonomous miles driven, all without any major incidents. And while the company claims to be “safer” than human-driven vehicles, its total fleet consists of just 1,500 or so cars—a mere drop in the very large bucket of all the cars on the road today. In its 10-ish years of operation, Nuro has racked up slightly more than a million autonomous miles without “an at-fault incident.” But again, Americans drive more than 3 trillion miles annually. As Cummings said last year, we have nowhere near enough data to make statistically strong claims about the true safety of robotaxis.

According to Kirkpatrick, Uber’s ambitious rollout plans would level up the risk. At the scale that Uber seems to be targeting, he said, “the exception can become the rule.” In other words, if 20,000 autonomous cars drive just 50 miles a day, that equals a million daily miles; in turn, that rare, one-in-a-million edge case that wasn’t included in the training data could occur on a daily basis. “Autonomous vehicles have to generalize to rare events with near-perfect reliability. At present, no end-to-end A.I. model has demonstrated capability to do that,” he said.

And the bigger a fleet gets, the greater the “aggregation of risk”—meaning, for example, a software bug that gets pushed fleetwide could cause a lot of harm without the proper safeguards. To deal with that risk, operators should take a page out of the aviation industry’s playbook, and implement rigorous, standardized software-quality assessment, real-time monitoring, and rollback systems. “But all of that has to be built intentionally. At present, that’s not what’s going on,” Kirkpatrick said. “Instead, these systems are operating in what is effectively a test environment—a civilian test environment that can become very dangerous.”

At this stage, so much about Uber’s plan remains decidedly unclear. It’s unclear how the partners will split up safety-observability responsibilities, or what that will look like in practice; it’s unknown how many teleoperators there will be, or who will employ them; it’s unknown how much these robotaxis will cost to build and maintain, or which geographies Uber is targeting first, or the constraints that will be applied to the fledgling fleet.

Finally, it’s unclear whether the rollout will coincide with a reduction in human-driven Ubers, or what the robo-Ubers will cost, or whether this partnership will in some way impact Uber’s current partnership with Waymo. These are questions, according to an Uber spokesperson, that are “best answered as we get closer to the rollout in 2026.” (Nuro did not respond to a request for comment.)

Activist groups, community groups, first responders, and unions have all advocated against the rise of the robotaxis. And rolling them out without the rigorous testing necessary could come back to haunt Silicon Valley if accidents end up stalling out the industry. Dr. Zach Asher, director of the Energy Efficient and Autonomous Vehicles Lab at Western Michigan University, told me that, generally speaking, more competition in this space is a good thing—but that he’s “worried, because aggressive expansion and dangerous activities, such as driving on busy roads, don’t mix.” (Currently, no robotaxi operates on highways in the U.S., although it’s an area that Waymo wants to expand to.) “All it takes is a major public accident to set the whole industry back a few years again.”

That’s all for today. If you’ve taken rides in a robotaxi—or have had some sort of encounter with one in the wild—I’d love to hear all about it…

I’ll see you Thursday.
Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St, New York, NY 10006