How Do Waymo Cars Work? Sensors, AI, and Safety Data

Transcript

Seeing the Road: Lidar, Radar, Cameras, and the Maps They're Built On

NovaWelcome to YourTalks.

NovaWe're talking about Waymo today — how a car with no one behind the wheel can navigate a city, and what it takes to make that work at the level where safety data starts looking alarming for human drivers.

AtlasAlarming for human drivers, maybe. Or alarming for people who think a tech company running a robo-taxi fleet is a fine substitute for actual transportation infrastructure. But yes — let's start with how the car sees the road, because the engineering here is not nothing.

NovaSo Waymo calls its approach Sense, Solve, Go. The "Sense" part is a stack of three sensor types working in parallel. Lidar — light detection and ranging — fires millions of laser pulses in every direction and measures how fast they bounce back, producing a full 3D point cloud of everything within roughly 300 meters. Objects in the dark, at a curve, even a pedestrian on a moonless night a<chapter index="0" title="How a Waymo Actually Sees the Road">

AtlasWelcome to YourTalks.

NovaWe're talking about Waymo today — how a car with no driver actually navigates a city street, what the safety numbers say, and why Toronto is still arguing about whether to let them in at all.

AtlasIt's a legitimately impressive engineering story, whatever you think about the regulatory politics around it. So let's start at the hardware — what the car actually uses to perceive the world.

NovaThe sensor stack is three layers working together. Lidar fires millions of laser pulses in every direction and times the bounce-back to build a live 3D point cloud of everything within about 300 meters. It works in the dark, which cameras struggle with. Radar reads speed and direction and cuts through rain, fog, and snow. Cameras handle the visual detail — traffic lights, road signs, a pedestrian's body language — out past 500 meters.

AtlasAnd no single sensor is trusted alone. The whole point is redundancy. If rain degrades the cameras, radar fills the gap. If a sensor fails, the others carry the load. That's a more disciplined design philosophy than most human drivers who are operating on one distracted brain and a coffee.

NovaHigh bar. But the sensors alone don't explain why the car knows where it is. Before Waymo ever carries a passenger in a new city, it sends manually driven cars through the streets building centimeter-precise HD maps — every lane line, curb, stop sign, and crosswalk logged in advance.

AtlasSo when the car is running a ride, it isn't trying to figure out the world from scratch. It's matching live sensor data against a map it already knows cold. GPS is a backup, not the primary source of truth.

Predicting What Everyone Else Is About to Do

NovaThe sensing and mapping get you a picture of the world. The harder problem is predicting what the unpredictable humans around you are about to do.

AtlasAnd this is where the AI earns its keep. The system classifies every moving object — car, cyclist, pedestrian, motorcycle — and then runs out dozens of probable future paths for each one simultaneously. It doesn't pick one prediction; it holds the whole probability distribution and plans around it.

NovaWaymo says its driver has logged over 270 million real-world miles and more than 20 billion miles in simulation. That training library is what gives the predictions weight. The car has seen a pedestrian step off a curb mid-block ten million times in simulation and knows the statistical distribution of what happens next.

AtlasWhich is a lot more systematically than the average human who learned to drive from a three-week course and a lenient tester. The system is running constant what-if scenarios — what if that cyclist swerves, what if the car ahead brakes hard — and pre-computing responses before any of it happens.

NovaWhere it gets interesting is the edge cases. A construction worker waving traffic through, a kid chasing a ball, a driver making eye contact and nodding you forward. Those social cues are hard for a machine to parse. The car can hesitate, and that hesitation has caused its own problems — other drivers rear-ending a Waymo that stopped unexpectedly.

AtlasWhich shows up in the data as a Waymo crash. The IIHS study found Waymos get rear-ended 40% less than human drivers — which is better — but the metric still exists. The car being too cautious has real costs, not just in awkwardness but in traffic flow and in forcing humans to react to its behavior.

When It Gets Stuck — and How It Recharges Itself

NovaSo what happens when the car hits something it cannot resolve — a flooded underpass, a cop hand-signaling at an intersection, something that breaks the map?

AtlasIt pulls over safely and calls for help. Waymo has roughly 70 remote assistance agents on duty at any given time, working across the US and the Philippines. Critically, in a February 2026 letter to Congress, Waymo confirmed those agents do not remotely drive the car. They advise — they feed the system guidance that the car can accept or reject.

NovaWhich is a meaningful distinction. There is one narrow exception — a trained agent can nudge a stuck vehicle forward at two miles per hour at fixed steering angles. Waymo says that function has never been used outside of training. So the car retains control; the human is a consultant, not a backup driver.

AtlasOn the charging side, it's more automated than people picture. Cars drive themselves to depot facilities at least once per day — the Nashville depot that opened this month can handle hundreds of vehicles around the clock. The car pulls up to a fast charger, displays a "ready to go" message when it's done, and drives itself to the next station. Human workers inspect and clean, but the car navigates the depot on its own.

NovaThe charge cycle is about 50 minutes every six to eight hours. Waymo runs 24-7, so the depots run three shifts nonstop. In Dallas, Waymo is building a premium facility with Avis featuring 40 DC fast-charge points. In Toronto, Element Fleet Management has been named as the depot operations partner. The physical infrastructure is scaling fast alongside the software.

AtlasIt's a logistics operation as much as a tech operation. The cars, the depots, the chargers, the maintenance bays, the remote support staff — that entire stack has to work before a single rider gets in. That's not something a startup can fake with a demo.

The Safety Numbers and the Toronto Standoff

NovaThe safety data is where the argument should settle — and it mostly does, at least on the statistics. Across 271 million driverless miles through June 2026, Waymo's own analysis shows 82% fewer injury-causing crashes than human drivers, 95% fewer serious-injury or fatal crashes. An independent IIHS study covering 50 million miles found 68% fewer crashes overall and 81% fewer injury crashes. These are not small margins.

AtlasAnd they are statistically significant across multiple cities and multiple methodologies. Swiss Re — one of the world's largest reinsurers, not a company that gets sentimental about technology — found 92% fewer bodily injury claims over 25 million miles. When the insurance industry validates the safety case, that's the market telling you something real.

NovaThe honest caveat is that Waymo operates in mostly sunlit, well-mapped, relatively mild-weather cities — San Francisco, Phoenix, LA, Austin, Dallas. Those numbers do not automatically transfer to a Canadian winter with black ice, heavy snow on the sensors, and roads that don't match the map because they were repaved last week.

AtlasWhich brings us to Toronto, where the regulatory situation is almost a parody of bureaucratic lag. Waymo got its provincial permit and started manual mapping drives this week. It wants paying riders by 2028. Ontario's pilot regulation that covers this testing expires in October 2027 — months before the launch date — and the province has announced no replacement.

NovaAnd Toronto's mayor Olivia Chow is on record with safety concerns while also acknowledging the city can't actually block Waymo — that's the province's call. Doug Ford has raised job concerns. Meanwhile the city is reviewing whether its vehicle-for-hire rules even apply to a car with no driver, who gets ticketed for moving violations, how insurance works. These are not trivial questions.

AtlasThey're not trivial, but they have answers if the government decides to write them. Every US city that now has Waymo went through this exact same cycle of regulatory ambiguity followed by incremental resolution. Ontario is not facing a novel problem; it's facing a political scheduling problem. The technology is ready to be tested. Whether the province writes rules in time is a choice, not a mystery.

NovaThe gap between what the data shows and what cities are willing to authorize is where most people are living right now. Waymo's safety statistics are real and they matter. But a 95% reduction in serious crashes measured in Phoenix does not automatically answer what happens when the same car hits a January ice storm in Toronto, or what happens to the taxi driver who loses their livelihood, or who pays when something goes wrong in a jurisdiction without clear liability rules. Those questions deserve the same rigor as the sensor stack — and so far, they're getting a lot less of it.

AtlasThe regulatory drag is frustrating precisely because the cost of delay is also measurable. If Waymo is right about the safety numbers, every month of delay is statistically associated with preventable injuries on those roads. Politicians treating caution as automatically virtuous are making a choice too, and they should own the consequences of that choice the same way they'd demand Waymo own a crash.

Sources

1. September 24, 2026 - From the roadwaymo.com
2. waymo.com · blog · shortsFrom the road — June 24, 2026 - Waymowaymo.com
3. Self-Driving Car Technology for a Reliable Ridewaymo.com
4. Waypoint - The official Waymo blogwaymo.com
5. Waymo Safety Impactwaymo.com
6. Waymo makes case for robotaxi safety citing its data, including in Atlantaajc.com
7. Self-Driving Car Technology for a Reliable Ridesupport.google.com
8. Waymo, now being tested in Philly, has a safer track record than people, new study saysinquirer.com
9. Waymo's driverless cars continue to crash less often than peopletheverge.com
10. Waymo's driverless cars crash less often than people - IIHSiihs.org
11. March 19, 2026 - From the roadwaymo.com
12. Waymo robotaxis safer than human drivers: IIHS reportsmartcitiesdive.com
13. Safety – Waymowaymo.com
14. The Waymo Driver Handbook: Teaching an autonomous vehicle ...waymo.com
15. Impact sur la sécuritéwaymo.com
16. Waymo picks Toronto as first Canadian city, targets riders ...electrek.co
17. Flexdrive by Lyft opens new Nashville AV depotlyft.com
18. Transcriptsroadtoautonomy.com
19. Hundreds of self-driving cars coming to Nashville with new Waymo, Lyft facilitywsmv.com
20. Flexdrive by Lyft’s Era of AV Fleet Management in Nashvillelyft.com
21. We’re Learning to Live with Waymos. It’s Going Weirdnewyorker.com
22. Waymo robotaxis are coming to Toronto in 2028 | CBC Newscbc.ca
23. Waymo Plans Toronto Robotaxis for 2028 as Ontario Rules Expire in ...thedeepdive.ca
24. Waymo in Toronto | Learn more about Waymo in Canadawaymo.com
25. Waymo is taking steps to get its robotaxis on Toronto ...cbc.ca
26. Waymo Plans First Canadian Robotaxi Expansion in Toronto for 2028hypebeast.com
27. www.canadianaffairs.news › 2026/10/09 › a-future-ofA future of driverless rides draws closer - Canadian Affairscanadianaffairs.news
28. Waymo Heads to Canada & Another Hint in the MOIA Puzzleavmarketstrategist.substack.com
29. Toronto Mayor has doubts as Waymo plans to test robo taxis in Ontariomobilesyrup.com
30. How's Waymo doing these days?newmarketpitch.com

More episodes

• Why Is RAM So Expensive Now? The AI Shortage ExplainedTechnology · 15 min
• Tesla Roadster Flying Car ExplainedTechnology · 7 min
• The Social Reckoning: True Story Vs. Sorkin's FilmTechnology · 10 min
• Why OpenAI Paused Training on Its Most Advanced ModelsTechnology · 5 min
• CXMT DRAM: China's Memory Chip Breakthrough, ExplainedTechnology · 7 min

Browse every episode