How robotaxis work: Tesla vs Waymo, explained

The short answer: a robotaxi runs the same loop as every other robot: sense, decide, act. The big split is in the "sense" part. Waymo puts cameras, lidar and radar on every car and drives inside areas it has mapped in detail. Tesla bets on cameras alone. Both now carry paying riders with nobody in the driver's seat in some U.S. cities, and both have remote humans who can help when a car gets stuck. Waymo is much further along: more cities, far more driverless miles, and years of crash data. Tesla is growing fast in Texas and Florida. The most useful habit is to ask three things about any claim: how many miles, where, and with whom on board or on call.

Part 1: Why I wanted to understand robotaxis

Robotaxis are probably the robots most people will meet first: a car that pulls up with nobody in the front seat. They're also the story that keeps showing up on Servospace. In one week the news feed ran pieces on Tesla opening Robotaxi rides in Fort Worth, Waymo adding Torrance and Manhattan Beach in L.A., Tesla's Texas fleet registrations nearly matching Waymo's, and Kodiak's driverless trucks heading for I-45. I was writing headlines about these companies without being able to explain, in plain words, what's different about how their cars actually drive.

So I read the primary sources: the companies' safety pages, the federal crash-reporting rules, the earnings releases. This guide is what I learned.

The usual disclaimer: I'm not an engineer. I'm learning robotics in public, and I'm skeptical of demos by default. In my first guide I described every robot as sense, decide, act. In my second I wrote about how real robots can still leave you with the wrong impression, especially when a human is quietly helping. Both of those ideas turn out to matter a lot for robotaxis. Every number below has a link. Where I couldn't find a source, I left it out.

Part 2: How robotaxis work

First, what "self-driving" actually means

The industry uses a scale from SAE International that runs from Level 0 to Level 5. NHTSA, the U.S. road safety agency, has a plain-English version. The short version:

Two things from NHTSA's page stuck with me. At Level 4, the system "can only operate the vehicle in limited service areas, not universally." That's what every robotaxi on the road is today: Level 4 inside a fenced-off map. And NHTSA says outright that no vehicle you can buy is fully self-driving. Every car for sale in the U.S. "requires the full attention of the driver at all times."

This is where a lot of the Tesla confusion comes from, so it's worth being clear. Tesla sells two different things. The Robotaxi service is a ride-hailing service where, in some cities, nobody is in the driver's seat. "Full Self-Driving (Supervised)," the software Tesla owners buy for their own cars, is driver assistance. In Europe, the Dutch regulator RDW said on Nov. 3 that it had signed off on renaming it "Tesla Assisted Driving," and it stresses that the driver is still responsible (Electrek; our story: Tesla's FSD in Europe gets a new name). Same company, same brand of AI, very different responsibility. When someone says "Tesla's self-driving," it's worth asking which one they mean.

Sense: lidar, radar and cameras vs cameras only

A robotaxi has to know where everything is around it, all the time: cars, cyclists, a kid between parked vans, a traffic cone. Three kinds of sensor do most of that work:

• Cameras see color and detail. They read traffic lights, signs and lane lines. Like your eyes, they struggle in glare and darkness.
• Lidar fires laser pulses and times the bounce to build a 3D picture of distances. It works in the dark.
• Radar uses radio waves to measure how far away something is and how fast it's moving, and it keeps working in rain and fog.

Waymo uses all three. Its sixth-generation system has 13 cameras, 4 lidar, 6 radar and an array of external microphones, with overlapping views all around the car up to 500 meters away, day and night (Waymo). The previous generation on the Jaguar I-Pace had 29 cameras and 5 lidar (9to5Google), so the newer one is leaner. In February 2026 Waymo began fully autonomous operations on the sixth-gen hardware, starting with employees and their guests, and its argument is about backup: the sensors are meant to cover each other's weak spots, "leaving nothing to the imagination of a single lens" (Waymo).

Tesla uses cameras. Tesla took radar out of new Model 3 and Model Y cars in 2021, then dropped the ultrasonic parking sensors in 2022, betting on what it calls Tesla Vision (Reuters). Its Robotaxi safety page describes external cameras giving a 360-degree view, with a neural network processing "over one million pixels" of visual data "every millisecond" (Tesla).

That's the core of the Tesla vs Waymo argument. Tesla's bet is that cameras plus good enough AI can do what human eyes do, and cameras are cheap. Waymo's bet is that a robot should see better than a person, with backups. I'm not qualified to settle that. Both are carrying driverless riders today, so neither approach is impossible. Which is safer at scale is a question for miles and crash data, and Waymo has a lot more of both.

Maps: knowing the road before you drive it

Waymo also drives with detailed custom maps of its service areas, which it matches against live sensor data to know exactly where the car is (WardsAuto). That's part of why a new Waymo city starts with cars driven by people. In Toronto right now, Waymo vehicles are out with trained drivers behind the wheel, mapping and testing, with fully driverless rides targeted for 2028 (Waymo; our story: Waymo cars are now driving Toronto streets).

Tesla hasn't published the same level of detail on how it uses maps, so I'm not going to guess. What's public is that Tesla also launches city by city, with service areas it chooses, and it's testing in places it doesn't yet carry riders. About two dozen Cybercabs are running around Detroit with test drivers and temporary steering wheels, for example (The Detroit News via BND; our story: Cybercabs testing in Detroit).

So "Tesla can drive anywhere and Waymo can't" is too simple. Both companies pick where their cars carry riders. That choice, the service area, is one of the most important facts about any robotaxi.

Decide: the AI in the middle

Once the car knows what's around it, software has to predict what everyone else will do next and pick a path: slow down for the cyclist, wait for the left turn, nudge around the double-parked truck. This is the "decide" step from my first guide, and it's where most of the modern AI lives.

A lot of that learning now happens in simulation. Baidu says its new Apollo World simulator can generate up to 100,000 km of virtual driving a day (CNBC-TV18; our story: Baidu's robotaxis practice in a virtual city). Pony.ai (PonyWorld) and WeRide (GENESIS) describe similar "world model" tools in their latest earnings releases. Companies don't say much about how well skills learned in simulation carry over to real streets, so I treat those numbers as effort, not proof.

Act, and the humans behind the curtain

The "act" part is the boring part: steering, braking, accelerating. Electric cars are easy for a computer to control. The interesting question is what happens when the car doesn't know what to do.

Every robotaxi company I looked at has remote humans who can help. This is the robotaxi version of the teleoperation question from my real-or-fake video guide, and the details matter.

Waymo calls it Remote Assistance. Waymo says its agents don't drive the cars. They answer specific questions the car sends, like "is this lane blocked?", and the car can take the advice or reject it. In February 2026 Waymo said about 70 Remote Assistance agents were on duty worldwide at any given time, for a fleet of about 3,000 vehicles (Waymo). It also has a U.S.-only Event Response Team for harder situations like dealing with emergency responders after a crash.

Tesla also has remote operators. Its own safety report says it counts every reportable collision while Robotaxi is driving with no safety monitor in the car, "including when a remote operator is assisting or has taken control" (Tesla). "Has taken control" is Tesla's phrase, not mine. Tesla hasn't published how often that happens, or how many remote operators it has per car.

Neither of these makes the cars fake. A remote helper who answers a question once in a while is very different from a person driving the car from an office. But it means "no one in the car" and "no human involved" aren't the same claim. When a company gives a ratio of remote staff to cars, like Waymo did, that's useful. When it doesn't, the honest answer is "unknown."

There's one more human layer to watch for: the safety monitor or safety driver inside the car. Tesla launched in Austin in June 2025 with a safety monitor in the passenger seat (Teslarati). It started some rides with no monitor at all in January 2026, beginning with "a few unsupervised vehicles mixed in" with supervised ones (CNBC). In the San Francisco Bay Area it's different again. As of Oct. 22, the independent tracker robotaxistats listed Tesla's Bay Area service as "supervised," with a safety driver, and noted that California's DMV permit list shows Tesla with a testing permit that requires a safety driver, not a driverless one (robotaxistats). So the same app can give you a very different ride depending on the city.

Where you can ride with no driver today

Here's where things stood when I checked, in early November 2026. Service areas change almost weekly, so the company's own app is always the final word.

A few things jump out at me from that table.

Waymo is still the clear leader in the U.S. Waymo's safety hub counts 271.3 million miles with nobody in the driver's seat through June 2026 (Waymo). Tesla's own report counts 3.1 million autonomous miles through October 18, 2026 (Tesla). That's roughly 87 to 1. The dates don't line up, but the mismatch flatters Tesla, not Waymo: Waymo's count stops in June, four months before Tesla's, and Tesla's only covers rides with no one in the car since January 22.

But Tesla's fleet is growing fast. As of Nov. 1, Texas DMV records showed 1,244 Tesla vehicles registered as automated vehicles, just 15 fewer than Waymo's 1,259 in the state (Robotaxi Tracker; our story: Tesla's Texas robotaxi fleet is now just 15 cars behind Waymo's). A registration is not a ride, though. It doesn't tell you how many of those cars are carrying paying riders, or how many drive with nobody inside.

It's not just taxis. Kodiak's technology already drives 35 trucks with no one in the cab in the Permian Basin oil fields of West Texas (as of mid-2026), for customer Atlas Energy Solutions. Kodiak says it's on pace to start driverless runs on Interstate 45 between Dallas and Houston by the end of 2026 (Kodiak; our story: Kodiak's safety case is 99% complete). That launch still depends on Kodiak finishing its own safety case.

The safety data that's actually published

The federal baseline: NHTSA's Standing General Order. Since 2021, NHTSA has required named companies to report certain crashes involving automated driving systems and Level 2 driver assistance (NHTSA). For robotaxis (the ADS category), a crash has to be reported if the system was in use at any time within 30 seconds of it and it caused certain property damage or injury. NHTSA publishes the reports as downloadable files, with its latest data running through September 15, 2026.

NHTSA's own warnings about this data are worth reading in full, because people misuse it constantly:

• The counts aren't adjusted for miles. NHTSA says companies don't have to report how many cars they run or how far they drive. A company with more crash reports may simply drive many more miles.
• Better sensors mean more reports. A robotaxi fleet with constant data uplinks knows about every bump. A privately owned car with driver assistance may never tell its maker about a crash.
• Some details can be hidden. Companies can ask NHTSA to treat the crash narrative as confidential business information.
• Reports can be incomplete or unverified. They're due quickly, before all the facts are in.

So when you see a headline like "Company X had 40 crashes," the first question is "out of how many miles?"

Waymo's safety hub. Waymo takes its own SGO reports, adds its mileage, and compares itself with human drivers on the same kinds of roads in the same cities (Waymo). Through June 2026, over 271.3 million rider-only miles, Waymo says that compared with an average human driver over the same distance, it had:

• 95% fewer crashes with a serious injury or worse
• 82% fewer crashes with any reported injury
• 82% fewer crashes where an airbag deployed in any vehicle
• 93% fewer crashes injuring pedestrians, 86% fewer injuring cyclists, 82% fewer injuring motorcyclists

Those are big numbers, so here's what makes me take them more seriously than a typical company claim. Waymo counts crashes regardless of who was at fault. It publishes the SGO report IDs, so outside researchers can match every crash to NHTSA's public data. It explains how it built the human benchmark, including a 32% adjustment for injury crashes humans never report to police. The methods are in peer-reviewed papers, like Kusano et al. in Traffic Injury Prevention.

And here's what keeps me careful. The research is written and run by Waymo's own team. The comparison is against all human drivers in those cities, including drunk and distracted ones, not against a careful sober driver. Waymo itself notes that its cities don't get much snow, so the data doesn't cover it. And Waymo says it doesn't yet have enough miles to say anything statistically solid about fatal crashes alone. All of that is on Waymo's own page, which is a good sign in itself.

Tesla's Robotaxi safety report. Tesla now publishes a safety page too (Tesla). For January 22 to October 18, 2026, covering only rides with no safety monitor in the car, Tesla reports zero people injured, zero Robotaxi airbag deployments and zero pedestrians or cyclists hit. It says human drivers covering the same Texas miles would have had about 9.1 injury collisions and 4.1 airbag collisions.

That's a real dataset, built on the same SGO reports, and I'm glad it exists. But reading the methodology, there are differences from Waymo's that matter:

• The sample is much smaller. About 3.1 million miles, compared with Waymo's 271 million. At that size, a couple of crashes either way can swing the picture.
• "Injured" is counted differently. Tesla counts injuries only in collisions "in which Robotaxi made first contact." Tesla describes one crash where another vehicle pushed an SUV into a stopped Robotaxi and someone was taken to hospital, and it didn't count that as an injury. Waymo counts crashes regardless of fault. Neither choice is crazy, but they aren't the same yardstick, so you can't line the two pages up side by side.
• The airbag count is the Robotaxi's own airbags. Waymo reports both its own airbags and airbags in any vehicle.
• It's a company scoring itself. Same caveat as Waymo. Tesla does link to NHTSA's public files, which is how anyone could check it.

My takeaway: both companies publish more than they used to, and both build on the same federal reports. Waymo's dataset is far bigger, older, and has been through peer review. Tesla's is new and promising, and too small to settle much yet.

China. Baidu, WeRide and Pony.ai all describe their safety records in earnings materials, but I couldn't find a public, crash-by-crash database for China like NHTSA's, so I can't check their numbers the way I can check U.S. ones. That's not an accusation, just the limit of what's public.

How to read a robotaxi claim

Here's the checklist I use now when a robotaxi headline lands in the Servospace feed:

And for videos specifically, the checklist from my real-or-fake guide still applies. A slick robotaxi clip can be a real car on its best take. Tesla's Halloween clip of Cybercabs carrying skeleton riders, for example, didn't say where it was filmed or whether the cars were driving themselves. When a post doesn't say, I don't fill in the blank.

FAQ

What's the difference between Tesla and Waymo robotaxis? Waymo uses cameras, lidar and radar plus detailed maps of each service area. Tesla uses cameras only. Waymo has far more driverless miles (271.3 million through June 2026 vs Tesla's 3.1 million from January 22 to October 18, 2026, by each company's count) and operates in more cities. Both carry riders with no one in the driver's seat in some cities, and both have remote staff who can help.

Is Tesla Full Self-Driving the same as Tesla Robotaxi? No. FSD (Supervised) on cars Tesla sells is driver assistance, and the driver is responsible. The Robotaxi service is a ride-hailing service that, in some cities, runs with nobody in the driver's seat.

Are robotaxis safer than human drivers? Waymo's published, peer-reviewed data shows large reductions in injury crashes compared with human drivers on the same city streets, across hundreds of millions of miles. Tesla's early data shows no injuries in its own count, over a much smaller distance. Neither covers snow-heavy cities, and neither is independent of the company yet, though both can be checked against NHTSA's public reports.

Do remote humans drive robotaxis? Waymo says its Remote Assistance agents give the car advice but don't drive it. Tesla's safety report says a remote operator can assist or take control. Neither company says the cars are remote-driven as a rule, but humans are part of both systems.

What I'm still figuring out

I don't yet know how to compare a camera-only car with a lidar car on equal terms, because no one publishes data in a way that lets you do that cleanly. The honest version of "which is safer" is "Waymo has proven a lot more, and Tesla is starting to show its work."

I'm also curious what changes when these cars hit snow. Waymo's Denver service is the first place I'll watch, and both companies are now testing around Detroit.

If I got something wrong here, or a number has moved since I checked, tell me (how to reach me is on the About page). I'll fix it and note the change.

Last updated November 4, 2026.

— Nate