Robotaxi Scaling Is Just Beginning

We are only at the end of the beginning; scaling is yet another 99% of the effort. But the industry won't get there if they don't change their approach to building trust.

The robotaxi industry is starting another upward hype swing, with Waymo scaling up, and other players wanting to be seen as still being in the race. As has been graphically illustrated by Tesla “robotaxi” videos from Austin, we are nowhere near the end of this journey. The next challenge is scaling, and it will be harder than all the previous challenges.

That last 99% is tough

I like to say that getting the first hundred robotaxis operational is the first 99% of the effort … and scaling up to a national fleet is the other 99% of the effort. So we’re at 99% done (for one company at least) … and 99% left to go.

Imagine the scaling complexities if robotaxis saturate the market in every city. https://pixabay.com/photos/taxi-traffic-jam-strike-traffic-4121829/

One way to look at scaling robotaxis (and robotrucks) is that for every factor of ten growth in fleet size, one should expect a fresh batch of challenges to graduate from quirks to problems. For a fleet of 10 cars, you can declare victory by being able to drive down the road without hitting stuff and, hopefully, not getting caught on camera going the wrong way down the street.

When you get to a fleet of 100 cars you start dealing with things that are not every-mile events, including: clogging streets, passenger discharge safety, hitting bicyclists, hitting towed vehicles, wet concrete, hitting mid-street utility poles, and so on. At this point, if you are a robotaxi company that is paying attention, you might reflect that more humility is needed while scaling up. Or you might charge ahead while trying to deflect attention away from the embarrassments that show up — until the high-profile mishap that puts you out of business. Deflection tends to stop working at the next 10x improvement in fleet size, when one-off situations mature into continual incidents.

At 1000 cars you can expect to see things that require a lot more training to handle than you’ll get from a few million miles of material used to create training data. A fleet at this size will experience frequent novel unstructured situations that people can figure out how to muddle through, but that tend to flummox robotaxi technology. By the time you get to a million cars there is no way to avoid addressing rare circumstances, because even a one-in-a-million situations will happen to the million cars in your fleet.

Most players are closer to 10s of cars, but Waymo is up in the 1000 car ballpark. It is completely unsurprising (to me anyway) that they are seeing unusual situations that are hard to handle. A recent one was a Waymo robotaxi with a passenger unable to handle a road closure due to a grass fire in the SF Bay Area (news story link). It’s unclear why a grass fire near a road in California should be something so inconceivable that it wasn’t handled by the design, but there you have it. A police officer had to jump into the car to move it. Maybe that works if there is one errant car. But police are going to get awfully tired of playing on-demand valet down the road to a fleet of a million confused robotaxis when that level of scale is achieved.

Please excuse the mess while we save lives

The autonomous vehicle industry, and Waymo in particular, likes to drive home the narrative that they are Saving Lives! — even though their statistical argument has some overly aggressive assumptions. They have gotten a bit better, perhaps in response to earlier criticisms of their studies I and others have made. And as a hint, any company who insists that crashes for which they are not mostly to blame do not count is playing inappropriate games with the net harm statistics. But I’ll let others take up the issues with their current methods.

Even if you believe everything Waymo says about their current safety analysis, a big problem for market success is that “Saving Lives!” or even just “fewer crashes” does not address the very real social burdens and negative externalities that are already being imposed by this technology. Blocking fire trucks, night-time parking lot honk-fests, engaging in overnight residential permit parking spot arbitrage, failing to stop for pedestrians in crosswalks, problems dealing with road flooding, driving past a stopped school bus, and so on all cause public concern and degrade public trust in the technology even beyond the continual stream of embarrassing pictures of robotaxi operational snafus.

The inevitable response from the industry amounts to “Don’t mind the mess — We’re Saving Lives!” or some variant. Alternately, “Well that problem is fixed, so you’re not allowed to mention it was ever a problem”, which I fully expect will be a response to the examples linked in this piece. (However, for the most part we really have no public-facing evidence of a fix, but rather just the company’s word that they’re always making things better.) Really, that just amounts to a variant of the Tesla go-to of “The next software release will Blow Your Mind!” More importantly, it amounts to a pattern of needlessly imposing costs on others and cleaning up after the mess has been made. Which is a problem, even if a company responds to public pressure (which is an additional cost for us to create) to clean up their messes. To be clear, this is common practice across robotaxi companies, with Waymo attempting to position themselves as a better actor than the others.

I realize I might be accused of stifling innovation! by pointing out areas in which robotaxi technology needs to improve. But making sure that innovation benefits society requires care and attention to issues not necessarily on the short and direct path to profitability. Innovation for its own sake is not really the point once you’re into scaling up, and past doing research or trying to find product/market fit.

I sincerely hope the technology someday saves lives. But that does not give carte blanche to play fast and loose with other users of our shared public roads on the journey. If you make a mistake, own it and fix it. Don’t pretend it’s someone else’s fault, or blame the messenger for pointing out that improvements are needed.

Working hard or hardly working?

The typical robotaxi industry deflection to a driving misbehavior is “we’re always working to improve.” But why are their cars scaling up when they can’t get the basics right, like driving the correct way on the road? Safety drivers or remote monitors could prevent these behaviors if implemented properly. Putting up with such problems is not necessary to advance the technology. Rather, it simply seems that various players in the autonomous vehicle space believe they can get away with doing this while they continue to work on the other 99% of maturing their technology.

The next time you see a story about a robotaxi causing disruption, ask the hard questions:

1. Did the company acknowledge it caused disruption, display regret, and articulate a specific, actionable, transparent, measurable plan to improve in the future? Or did they deflect blame and argue to “trust us bro, we’re busy saving lives”? Where is the concrete, verifiable follow-up to show they fixed the problem?

2. Was the disruption foreseeable? It is one thing for a company to stumble with a truly unforeseeable situation. It is another for them to aggressively scale while being willfully ignorant of scenarios that human drivers can handle and are obviously foreseeable. Grass fires in California are readily foreseeable. Wildfires have been mentioned in the UL 4600 autonomous vehicle safety standard since the first edition in 2020. Why did Waymo stumble when that happened to them? They’ve got remote assistants. If those remote assistants cannot handle foreseeable issues such as wildfire response and other unusual situations as well as human drivers can, they need a better plan.

3. Is the incident likely to happen again when the fleet size is 10x, 100x, or 1000x larger? (The answer is almost always: yes, and more frequently.)

4. Is the company compensating those who had a cost imposed on them? Is the police department compensated for filling in with an on-road assistance effort caused by an automation failure? What about drivers who spent a half-hour stuck in a traffic jam when a gaggle of robotaxis stopped in the middle of a street? What about for other incidents? To be sure, these are not fatalities and major injuries. But the costs imposed on others are not nothing.

5. Was the company held to account at least to the degree a human driver doing the same thing would have been? Police still can’t give real traffic tickets to robotaxis in California. A free pass on mistakes to a company with 10 cars does not scale up to thousands of cars without causing serious problems. At that point you’re not innovating; you’re deploying at scale.

Asking for responsible innovation

Any new technology requires societal adjustment. But we should not put up with unnecessary disruption because a company is too busy meeting its KPIs and showing progress to profitability to take the disruption they are causing more seriously. Honking horns at 4 AM next to sleeping residents is not required to someday save lives. Nor is blocking fire trucks, or driving the wrong way in a traffic lane, or any of the rest. We can ask for responsible road citizenship without being anti-progress.

The robotaxi industry’s refrain of the problem being society resisting inevitable assimilation into their new techno-utopia is growing old, as such arguments always do. Talked about less often are the significant messes being made and cleaned up behind the scenes. (There was a highly revealing podcast about issues with Waymo remote operations on the Web for a few days until it was scrubbed — one might speculate due to pressure from Waymo or their remote assistance service provider. If they want to really be transparent about operational safety perhaps they could repost that podcast and explain how they have addressed the concerns discussed by someone who spent time as an insider to their operations. But I’m not holding my breath.)

I’ve heard robotaxi folks complain about negative public sentiment, and that somehow it is unfair. Some of this is likely due to normal human caution around new technology. But I think a lot of this they have brought on themselves. They can do better, but it is up to them to change their strategy for building trust. I have specific ideas on that. Those ideas do not involve cover-ups of embarrassing facts.

Innovation is a means, not an end. What is important is responsible innovation. The robotaxi companies need to stop playing hide-the-ball on blame and accountability. Instead, they need to display real transparency when it comes to taking proportional blame and displaying evidence of improvement on metrics that matter to public stakeholders beyond net injury statistics. They also need to come to terms with addressing a huge set of situations that are not as graphic as a fatal crash, but every bit as important to integrating this technology into our society in a responsible fashion. Because we can expect a whole lot more of those incidents as the technology scales.

We can hope that robotaxi technology will improve quality of life and reduce traffic harm. But it won’t amount to anything if the technology can’t be trusted to operate at scale. And scale won’t happen unless the industry changes its tack on trust and transparency. Technology push runs out of steam eventually unless demand pull is cultivated based on mutual respect and trust. Having the industry blame the messengers (pro-consumer advocates and the press) for talking about what’s really going on won’t solve their problems.

Changing the current dynamic requires a new game plan from the robotaxi companies to build genuine trust. All of them that are successful will get to where Waymo is now, and likely struggle with the same problems.

This is an industry problem, not just a Waymo problem. The ball is in their court.

---

Phil Koopman has been working on self-driving car safety for almost 30 years. For more on this and related topics as they applied to any embodied AI system, see his new book: Embodied AI Safety.

---