The Death of KitKat
Why urban AVs struggle with the last two feet
Before we get into the main subject today, a quick promotional note: on Tuesday, 2 December at 1:00 PM ET (10:00 AM PT), I’ll be doing a live conversation with of
about The End of Driving, right here on Substack, courtesy of Substack Live. In our talk, Jeff and I will explore the policy choices that will determine whether automated vehicles improve, or worsen, our cities and our transportation networks. It should be a fun and interesting conversation!If you’re already a subscriber, you’ll be notified when we go live, but I encourage you, if interested, to sign up here. I hope you will join us!
The memorial appeared within hours. By October 28, the sidewalk outside Randa’s Market on 16th Street in San Francisco’s Mission District had transformed into a makeshift shrine of a sort that city-dwellers everywhere would recognize: candles, wilting bouquets of flowers, photographs, and handwritten notes of grief and gratitude.
It was a show of mourning for KitKat, a nine-year-old tabby. A bodega cat and fixture of the neighbourhood, he had spent nearly a decade greeting customers and napping in produce boxes. The night before, he had been killed, struck by a Waymo robotaxi as it began to enter traffic.
Artist’s rendition of KitKat, after this photo
The firm responded quickly. Waymo confirmed its vehicle’s involvement, expressed condolences for the community’s loss, pledged a donation to a local animal-welfare organization, and offered its account of what happened: the cat “darted under the vehicle as it began to pull away”.
I regret KitKat’s death. And I regret more that there will be more incidents of this sort. It’s a sad truth that automated vehicles (AVs) struggle in this specific domain. While AVs perform brilliantly at highway speeds, they falter in the final two feet: AV technology does best in environments of structured predictability, and city curbs are anything but.
An Unfortunate Pattern
KitKat is not the first animal killed by an autonomous vehicle in San Francisco, nor even the first killed by Waymo. In May 2023, a Waymo vehicle struck and killed a small dog; the company filed required documentation with the California DMV describing the incident as “unavoidable”. Six weeks later, a Cruise robotaxi hit a Labrador retriever, which survived with injuries.
Let’s be clear that three incidents, across two-and-a-half years, involving two different operators, is not statistically significant. As per the New York Times, hundreds of animals are killed by human-driven cars each year in San Francisco. Given the facts, I don’t think anyone is seriously arguing that AVs pose a greater threat to urban wildlife than conventional traffic does.
That said, let’s not ignore the fact that these cases do point to a threat to be taken seriously.
These cases share a distinctive signature. All three occurred at low speed. All three happened near the curb. All three took place during or immediately adjacent to pick-up and drop-off manoeuvres.
And notably, San Francisco’s robotaxis—which have now logged millions of miles in the city, and are held to stringent reporting requirements—have no documented high-speed collisions with animals. If AVs were generally worse at avoiding animals than human drivers, we would not expect that to be true. Instead, we would expect incidents distributed randomly across operating contexts: at high speeds on highways, at medium speeds on residential streets, and at slow speeds in parking lots or at intersections. Instead, these collisions cluster tightly in this specific scenario.
This concentration doesn’t suggest across-the-board failure. It suggests a specific technical limitation in a specific environment, namely perceiving what’s happening immediately adjacent to the vehicle as it begins to move from a standstill at the curb… that is, the place where almost all robotaxi trips will begin or end.
In other urban environments, robotaxis excel. At a range of 50 metres, today’s automated driving systems (ADS) simultaneously track dozens of moving objects, including vehicles, pedestrians, and cyclists; model and predict their trajectories; and react faster than any human driver could. But at a range of 50 centimetres—the space directly beside and beneath the vehicle—those same systems have persistent issues with detecting living things, an instance of what engineers call the “near-field perception problem”.
The near-field perception problem stems from fundamental physical and computational constraints.
Start with geometry. All sensor modalities—LIDAR, radar, cameras—have minimum-range thresholds below which detection reliability degrades. LIDAR systems typically can’t resolve objects closer than 30 to 50 centimetres, for example. Radar, excellent at distance, struggles with stationary or slow-moving objects at close range. And cameras mounted at windshield height or above can’t see under the car, or objects close to it.
What this means is that an AV typically can’t sense, in real time, the space directly beside and beneath itself. Instead, it infers the nature of that space from earlier observations. And these inferences can be tragically mistaken; which is what, I imagine, occurred in KitKat’s case.
The most prominent example of a near-field perception failure was the October 2023 incident that led to the death of Cruise. A Cruise robotaxi struck a pedestrian and then dragged her roughly six metres because it failed to detect that she was pinned underneath the vehicle. Investigators noted that parts of the pedestrian’s legs were partially visible to the vehicle’s sensors, yet the system failed to recognize her as a pedestrian and remain stationary. To the best of my knowledge, neither Waymo nor Zoox has publicly announced any dedicated under-vehicle sensing modules added in the wake of the Cruise crash. Operators rarely discuss this specific vulnerability, likely because acknowledging it touches precisely the area—low-speed, close-quarters interactions—that the public seems to find most viscerally troubling.
Speaking of things operators prefer not to discuss, the Cruise incident and the death of KitKat reflects a broader ADS principle, namely that it performs best in familiar, common scenarios; it flourishes in structured interactions, and struggles in unstructured ones.1 AVs excel at predictable interactions, like managing lane flows, obeying traffic signals, and respecting the right-of-way. But the curb is not predictable: it’s visually cluttered and dynamically complex. And yet AV trips will almost always begin or end at a curb.
The question sharpens: if this weakness is known and the environment is unavoidable, why haven’t straightforward mitigations become standard operating procedure across all AV operators?
Solutions That Are Simple, Intuitive, and Wrong
The potential mitigations seem clear enough. Institute ultra-low-speed roll-outs, say 5 kph, for the first three metres of travel. Conduct full 360-degree scans before initiating any movement. Deploy additional cameras lower on the vehicle. Add audible movement cues—a tone or voice announcement before the vehicle begins pulling away, giving nearby pedestrians and animals a moment to react.
It would seem that there is both technology and behaviour that mitigate these risks. So why isn’t it standard practice?
It’s because the obvious fixes have non-obvious problems.
Start with placing cameras lower on the vehicle. The robotaxis I’m familiar with—Waymos, Zooxes, and conventional Teslas—have cameras mounted at windshield height or higher. If they were lower, they would be able to see ground-level objects immediately adjacent to the vehicle… but they aren’t. Why not? Because, to ensure durability, weather resistance, maintainability, and protection from debris and curb strikes, they really can’t be placed lower. In theory, manufacturers could add low-mounted or under-body cameras to monitor the blind zone directly beside or beneath the vehicle. But in practice, these positions are too difficult to keep clean, operational, and reliable in urban conditions. They accumulate grime, snow, and road spray; they suffer frequent impacts; and their fields of view are often occluded by curbs, wheels, and darkness.
The challenge is therefore not ‘just add a camera’, but figuring out how to place a robust, safety-critical sensing modality into one of the harshest parts of the vehicle exterior. The fact that they’d be attempting to detect low-profile objects makes the problem worse. A cat crouching against a curb, a small dog darting between parked cars, curbside debris: these present resolution challenges, especially against visually-complex backgrounds of pavement, shadows, and street clutter.
So solving this problem with technology is tricky. Can we solve it with behaviour? We could, but that would invite other costs. Longer dwell times at curbs would create traffic friction. In dense urban environments, a robotaxi that takes 15 seconds to complete a pre-departure sensor sweep and then crawls away at 5 kilometres per hour would become an obstruction. It would block bike lanes, impede bus service, and frustrate drivers queued behind it. At scale—thousands of pickups daily—these delays would compound into system-level inefficiencies that reduce fleet productivity and increase operational costs. Perversely, more conservative curb behaviour would paradoxically increase exposure to rear-end collisions from impatient human drivers who don’t expect a vehicle to pause for five seconds before pulling out.
In time, these solutions may mature to effectiveness. It’s always possible we’ll invent a better sensor, one effective enough to spot near-field people and animals; robust enough to survive the rough environment of a car’s underbelly; and cheap enough to deploy at scale. It seems unlikely to me, but the engineers have their ways, and one can’t rule it out. And as the number of AVs on the roads increases, driver behaviour will become more and more AV-friendly, making slow roll-outs normal rather than burdensome. So, over time, response will improve. The question is, will it improve fast enough?
The answer may be no. Political tolerance for edge-case failures may be eroding faster than technical improvements are accumulating.
The Political Reckoning
On 5 November 2025, nine days after KitKat’s death, San Francisco Supervisor Jackie Fielder introduced legislation to grant cities and counties explicit authority to regulate or restrict robotaxi operations. The measure cited KitKat’s case directly.
In California, cities do not possess this authority today: AV operations are regulated exclusively at the state level, with the DMV governing testing and deployment and the Public Utilities Commission overseeing fare-charging services. State-level preemption was a deliberate choice by California to support innovation over and against early municipal efforts to restrict or tax robotaxis. The state continues to hold that view; when state legislators considered a 2024 measure that would have restored city control over robotaxi permitting and allowed local governments to impose new fees, the committee warned that the bill risked “unnecessary local control” and potential “policing for profit” by municipalities.
State-level preemption still seems correct to me. I recognize that this will chafe local politicians, especially those with ties to the Teamsters. But game recognize game: Fielder has correctly identified that incidents like KitKat’s death carry outsized political weight, and is savvy to use it to her advantage. When a human driver strikes an animal while pulling away from a curb, it’s unfortunate: a momentary lapse in attention, perhaps, or an unavoidable accident. When an ADS does the same thing, it reads differently, because the strongest argument for ADS is precisely that it is safer than a human driver. Failing to detect a nine-year-old cat sitting near a curb suggests fundamental inadequacy.
Animals are sympathetic victims, which doesn’t help. Their deaths generate genuine grief which can be amplified by social media. But it’s not only animals at risk due to poor near-field perception. Children, the elderly, wheelchair users: all of these people occupy the same space where KitKat was killed: low to the ground and/or potentially occluded, moving unpredictably or not moving at all. If an ADS struggles to reliably detect a cat, how confidently can they detect a toddler in a dark jacket or a person who has fallen?
The paradox is that AVs can only refine near-field sensing through real-world deployment: machine learning requires data, which comes from operation. But deployment may be threatened politically by the very edge-case failures that would, if tolerated, eventually be trained away.
The paradox matters because of the broader political landscape. While I was drafting this piece, the New York Times published its own coverage of KitKat’s death, which took a sharply negative tone toward Waymo… despite the fact that the article’s content shows why that tone is at sharp odds with the facts on the ground.
Waymo today operates a fleet of roughly 1,000 vehicles across the Bay Area, with service now reaching San Jose Airport and expected to expand to SFO shortly. Public support for robotaxis has grown dramatically: from 44% in September 2023 to 67% by July 2024. These are not the hallmarks of a technology losing the confidence of a city; they describe a service becoming more mainstream, normalized, and trusted.
Further, in the weeks before KitKat’s death, San Francisco’s animal-control service processed the bodies of twelve other cats struck by cars; which, because they were killed by human drivers, received no media coverage, nor shrines, nor political resolutions. The human toll was worse: 43 people were killed by human-driven vehicles in San Francisco last year, while Waymos killed none.
None of these facts minimizes the sadness we must feel at KitKat’s death. But they do illustrate how this persistent issue of AVs killing animals will command attention, and can be exploited by motivated reasoners, to advance an anti-automation narrative, even when the evidence shows how AVs contribute to safety rather than detracting from it.
I think this is a problem the AV sector must prioritize.
KitKat won’t be the last animal to die in an AV’s near-field zone during pull-away from a curb. But deaths like these could be reduced significantly, if this engineering problem is taken seriously. It’s a specific, well-defined class of failure that can be reduced through better near-field sensing, improved pull-away protocols, and clearer operational constraints. And it needs to be solved, because public tolerance for edge-case failures is not infinite. High-salience incidents erode trust faster than operational improvements rebuild it. Each failure like this one is a data point in a public narrative about AVs not being ready for urban deployment. And once that narrative solidifies, dislodging it will become extraordinarily difficult, even as the underlying technical reality continues to improve.
The technology will get there, eventually. But the sector can’t afford to test whether public patience and political license will last long enough to let it. I hope Waymo and Zoox understand this, and act accordingly.
Respect to for feedback on an earlier draft.
This is true whether the ADS relies on transformer-style AI that absorbs immense quantities of driving footage, or on traditional programming that relies on driving simulations or in-the-field data, or—as is increasingly the case—both.