Latest / The Robotics Podcast with Fexingo: Autonomous Systems, Industrial Robots, and Hardware / The Robotaxi Dilemma When Safety Becomes a Headwind
Transcript
- Lucas: So there's this number that's been stuck in my head since the California DMV published its 2024 autonomous vehicle collision report. Autonomous vehicles in the state had a collision rate of about 3.2 incidents per million miles driven. Human drivers in the same cities were at roughly 12.9 incidents per million miles. Luna: That's a 75 percent lower rate for the robots. Which sounds like a slam dunk for the technology. Lucas: It should be. But instead of a green light for expansion, the last eighteen months have seen regulatory pauses, public hearings full of angry testimony, and at least one major operator effectively shut down in its home market. Luna: Right — we're talking specifically about Cruise and Waymo, mostly in San Francisco and Phoenix. Lucas: Exactly. And the puzzle is: if the data says the machines are safer, why is the political and regulatory momentum moving in the opposite direction? That's the question I want to unpack today. Luna: So let's start with the incident that really changed the conversation. October 2023, San Francisco — a pedestrian was hit by a human-driven car, thrown into the path of a Cruise robotaxi, which then dragged her about twenty feet before stopping. Lucas: That event was a catastrophe for Cruise. The company lost its permit to operate driverless vehicles in California within weeks. But here's the thing that doesn't get enough attention: the human driver who initially struck the pedestrian fled the scene. The robotaxi was the secondary actor. Yet the public and regulators focused almost entirely on the machine. Luna: And that's the asymmetry at the heart of this whole debate. Human drivers kill roughly forty thousand people a year in the US, and we accept it as background noise. But one robotaxi drags someone twenty feet, and suddenly the technology is too dangerous. Lucas: Right. And I want to be fair — the dragging incident was genuinely disturbing. The vehicle detected the collision, stopped initially, then attempted to pull over, not realising the pedestrian was trapped underneath. That failure mode — a robot not understanding what it had done — is viscerally frightening in a way that a human error isn't. Luna: So the safety data is better, but the failure modes are alien. That creates a trust gap that statistics alone can't close. Lucas: Precisely. And that's where the regulatory paradox kicks in. The DMV report I mentioned actually covered data through the end of 2024. It showed that Waymo's fleet in San Francisco had zero at-fault collisions in over seven million miles. Zero. But the California Public Utilities Commission still moved to slow down expansion, citing 'uncertainty about public safety.' Luna: Let's pause on that — because I think listeners might wonder how zero at-fault collisions is possible. I mean, robotaxis get into fender benders all the time, right? There are YouTube compilations. Lucas: Sure, they get hit. The data distinguishes between crashes where the autonomous vehicle was at fault and crashes where it was struck by a human driver. In Waymo's case, virtually every collision in 2024 was a human driver hitting the robotaxi. The robot was obeying traffic laws, and the human wasn't. Luna: Which is actually a remarkable safety record. But the public perception is 'they keep crashing.' Lucas: And that's the communication failure. The companies have done a terrible job explaining what 'safe' means. They lead with technical benchmarks — disengagements per mile, miles between interventions — which are meaningless to most people. Meanwhile, critics point to any collision, regardless of fault, as evidence the technology isn't ready. Luna: So what would a better metric look like? Something a normal person could grasp. Lucas: I think the industry should standardise around a single number: fatalities per hundred million miles, compared to the human baseline. That's the metric we use for aviation, for rail, for every other mode of transport. If a robotaxi fleet can demonstrate it's safer than the average human driver on that measure, the regulatory question shifts from 'is it safe enough?' to 'why are we blocking it?' Luna: But that runs into the data problem. Robotaxis simply haven't driven enough miles to produce statistically meaningful fatality numbers. Waymo has maybe thirty million driverless miles total. That's not enough to establish a fatality rate with confidence. Lucas: It's a catch-22. You need more miles to prove safety, but regulators restrict miles because you haven't proven safety. And meanwhile, thirty-seven thousand human drivers died on US roads in 2024. Luna: This is actually the kind of conversation that makes me appreciate why we do this show — because without people digging into the specifics, you just get headlines and outrage. Lucas: Totally. And it's also why listener support matters so much. We don't have advertisers pulling the strings or pushing a narrative. A couple of dollars a month is genuinely what keeps these going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: It makes a real difference. And it keeps us free to follow the data wherever it leads. Lucas: So let's follow it to Phoenix, which is where Waymo has been operating the longest and most quietly. They launched there in 2020, and by early 2026, they're running what amounts to a citywide taxi service across about two hundred square miles. Luna: Phoenix is interesting because the regulatory environment is much friendlier than California's. Arizona basically said 'show us you're safe and we'll let you expand.' Waymo has had far fewer public incidents there. Lucas: But even in Phoenix, there's pushback. In 2025, the city council held a hearing after a series of non-injury incidents — robotaxis blocking intersections, circling a parking lot, getting confused by construction zones. None of them dangerous, but all of them annoying. Luna: And annoyance scales differently than danger. A thousand people mildly inconvenienced by a robotaxi that won't move out of a crosswalk is a thousand people who show up to a city council meeting angry. Lucas: Exactly. And the companies have been slow to address that. Waymo's response to the intersection-blocking issue was to tweak the routing algorithm so vehicles avoid certain complex intersections. But that's a patch, not a solution. The underlying problem — that the vehicle's planning system can get into an edge case it can't resolve — remains. Luna: So let's talk about Tesla, because they're the elephant in this room. Elon Musk has been promising fully autonomous driving since 2016. Tesla's approach is vision-only — no lidar, no radar — and they're deploying Full Self-Driving to hundreds of thousands of consumer vehicles. Lucas: And the safety data there is much harder to assess. Tesla doesn't report disengagements or collisions per mile in any standardised way. They've argued that their fleet is so large — millions of cars — that even if the per-mile incident rate is higher than Waymo's, the total number of lives saved could still be bigger, because they're deployed at scale. Luna: But that argument only works if you accept that Tesla's system is actually safer than a human driver. And the evidence there is contested. Several studies have shown that Tesla's Autopilot and FSD reduce crash rates, but the effect size is much smaller than what the dedicated robotaxi companies claim. Lucas: There's also a fundamental difference in architecture. Waymo and Cruise build vehicles from the ground up as robotic platforms, with multiple sensor modalities and redundant systems. Tesla is adding software to a standard car. That makes it cheaper and more scalable, but it also means the vehicle can't handle the same range of scenarios. Luna: And that's the trade-off the industry is wrestling with right now: do you go the Waymo route — slow, expensive, but demonstrably safe in the operating domain — or the Tesla route — fast, cheap, but with a higher tolerance for edge-case failures? Lucas: I think the answer in 2026 is becoming clearer: regulators are pushing toward the Waymo model. The National Highway Traffic Safety Administration has proposed a framework that would require autonomous vehicle operators to submit detailed safety reports, including disengagement data and collision reports, before they can scale beyond a pilot phase. Luna: Which effectively raises the bar for any company that can't produce that data. And Tesla can't, because they don't collect it in the same way. Lucas: Right. So the regulatory environment is hardening around a standard that favours the conservative approach. That's good for safety in the near term, but it might slow down the overall adoption of autonomous vehicles, which has its own safety cost — the continued deaths from human drivers that could have been prevented. Luna: So what's the path forward? How do you break the catch-22? Lucas: I think you need an independent body — not a company, not a regulator — that collects and publishes a single, standardised safety benchmark across all autonomous vehicle operators. Something like the aviation industry's NTSB, but for roads. Then the public can compare, and the debate moves from 'are robotaxis safe?' to 'how safe are they compared to humans?' Luna: And if that benchmark shows robotaxis are, say, twice as safe as human drivers, do you think public opinion would shift? Lucas: It would take time. But I think it would shift. The challenge is getting there without another high-profile incident that sets the industry back years. And that's ultimately a question of engineering patience — something the tech industry has never been very good at. Luna: Patience and safety culture. It's not a coincidence that the companies with the best safety records — Waymo, and also some of the smaller players like Motional — are the ones that have been working on this for over a decade. Lucas: And that's the note I want to end on. The robotaxi dilemma isn't really about the technology. The technology is good, and it's getting better. The dilemma is about trust, regulation, and the gap between statistical safety and perceived safety. Closing that gap is the real engineering challenge for the next five years. Luna: And it's one that will determine whether autonomous vehicles become a fixture of city life or remain a curiosity in a few sunbelt suburbs.