I’ve been writing about and working on Robotaxis for close to two decades. I worked on Waymo’s early team, and helped craft their robotaxi strategy, and my writings and talks have guided many, including the leaders of many of the top robotaxi companies. In addition, I own a Tesla with FSD 12, and 15 years ago also sat down for an hour of private conversation with Elon Musk to try to convince him to do robotaxis. Not saying at all I’m the reason Tesla is doing what it’s doing, but I do know all sides fairly well.

Many people think there’s really no comparison between the companies and that one is the clear leader–but they differ on which one that is. I’ve made a number of articles and videos on the issues around this issue. Now I’m here to clear that up. It’s all be made more visible by Waymo’s growth–they just launched in Atlanta–and Tesla’s pilot launch of a robotaxi service in south Austin on Jun 22, with supervising human safety drivers inexplicably in the passenger seat

This is a longer piece, but the matter isn’t simple. Let’s start with Tesla’s advantages.

Tesla Advantages

Tesla is a car company, and not just any car company, it’s the most innovative and dynamic in the non-Chinese world, and by a large margin. They essentially invented the US EV market and made everybody use the phrase “Software Defined Vehicle.” When it comes to making the physical robotaxis, Tesla will be very good at that.

Tesla also has Elon Musk, who is both a visionary and a risk taker. Many would say he takes too much risk, but if reward needs high risk, he’s your man. He is simultaneously sometimes a negative, a man also of deep flaws who sometimes scares away employees and customers but investors think he’s a big net positive.

Tesla’s taking a big, and longshot gamble that they can make self-driving work with just cameras and machine learning, while almost every other active player, including all those who have working robotaxis, incorporates other sensors like LIDAR and radar. Most of them also don’t use a pure end-to-end machine learning approach like Tesla is switching to, though they all make extensive use of machine learning and vision. Tesla’s hardware approach is a fair bit cheaper today, and will be a little cheaper in the long run, because cameras are low-cost. Their approach to software also needs less programming and more data, and Tesla has a giant fleet of cars driving around with their cameras, collecting data. They have so much data they can’t use more than a small part of it, but it’s definitely an asset.

Tesla hopes that, should they get the software working, there will instantly be millions of Teslas already out on the roads, ready to drive themselves. Other teams still need to build their fleets. Existing Teslas could become Robotaxis, but Tesla also plans to manufacture a custom Robotaxi fleet. Tesla at one point even said they would take off-lease Teslas and convert them to Robotaxis, which lets them acquire vehicles at a nice low cost, but they seem to have paused that plan.

Tesla has more compute in their cars than any old-school carmaker, and they have built huge compute server farms to train their software. They’ve shown particular prowess and building and running such farms, faster than other AI companies. Most of the cars in their fleet have their 3rd generation hardware, which they now admit won’t be enough for self-driving, but Tesla has promised to upgrade the cars of people who bought their FSD package. Newer cars have their 4th generation AI chip, and a 5th is coming.

That fleet has many hundreds of thousands of customers eagerly willing to pay for the privilege of testing Tesla’s prototype software, and to provide training data for it. All other companies except MobilEye had to pay employees to do that testing. Tesla gets more testing data than they can handle.

All this fits in the “machine learning maximalist plan.” Under that philosophy, all problems can be solved by adding more data and more compute power. This approach is, to some degree what has brought us recent revolutions like chatGPT and its cousins--no wonder people love it.

Tesla already has some of the infrastructure you need to run a robotaxi service. In particular, they have the largest charging network, and they have service centers in many towns. Unfortunately, their new CyberCab uses wireless charging, for which they have no charging network installed, but they can add it to some of their existing charge locations at a lower cost than building new depots. Tesla has an app, with prototype ride-hail functionality.

For a few months, Elon Musk held a prominent position in the U.S. government and sway over the President. It seemed like that might protect them from regulation, but that bromance has faded and this advantage has gone away.

Tesla Disadvantages

Overwhelmingly, Tesla’s biggest problem is that their current self-driving platform can’t self-drive. In the industry, to say a car can “self-drive” means being able to operate with a “bet your life” level of safety so cars can drive the roads with a sleeping person in them. In fact, Tesla’s not even close by the standards of the industry. This surprises many Tesla FSD users, who report they are highly impressed with the system based on taking it for several drives. They get this false impression because you can never learn that a system is good at self-driving by driving with it yourself. Today, estimates suggest Tesla’s FSD 13 can take a few trips – perhaps 10 to 30 in a row – before it needs some sort of safety intervention by the driver. That’s a recent improvement; last year it had trouble doing a single trip. As good as that improvement sounds, the bar is vastly higher. A car must be able to make not just a dozen trips without causing a problem, it must make around fifty thousand – a whole lifetime. Perhaps 100,000 to meet the bar Musk has set of “much better than a human.”

Check out my article and video on this topic, where I also explain how, if a car is improving greatly in your view, that means it’s immature, as mature vehicles have almost invisible, incremental improvements.

As bad as humans are, most have an average of around one crash in their lifetimes, and robocars, as Musk says, need to be a fair bit better than humans.

Where Tesla is today, from a safety standpoint, Waymo was over 8 years ago. That doesn’t mean they will take 8 years to catch up, but it means they have far to go. The good news is that the second mover gets to move faster than the pioneer. Tesla can take advantage of tools and techniques that didn’t exist when the earlier players were working on their vehicles.

And, as noted previously, Tesla is taking a longshot bet of doing it a different way, a cheaper way, possibly a faster way. That bet might pay off. Indeed, many experts feel that some day, the vision-based approach will work. The concern is that nobody can name the year when it will work, and some think it won’t ever work, or at least not with the current Tesla hardware. Later, I’ll look at some factors which might affect that. Elon Musk has notoriously said it would work “next year” for close to 8 years–he clearly has no crystal ball as to when it will happen, and nor does anybody else. It might be next year. It might be ten years. It might be effectively never. If it does work sooner, then Tesla can take advantage of that giant fleet. Otherwise it will need to look at other strategies. They promised a service that runs unsupervised, with nobody in the car but could not deliver, at least for now.

Elon Musk says that LIDAR is a “crutch” that just distracts you from solving the big AI problems of self-driving, and that once you solve those, driving with just cameras is solved as well. Many hope this is true because humans drive with just cameras, namely our eyes. But that’s not true–humans drive with our eyes and the human brain, and of the two, the brain’s the most important part. In addition, machines almost never use the same techniques to do something as humans–airplanes use propellers not flapping wings. No AI we have today remotely matches the power of the human brain at solving general problems. The hope is that driving needs only a much simpler set of tasks, which AI might live up to. To give it a leg up, though, most teams use sensors like LIDAR and radar which provide superhuman ability. It’s vital to know the distance and speed of everything on the road with you, and LIDAR and radar measure that inherently and reliably. Humans and vision tools must try to guess it, and they’re getting pretty good, but not nearly as reliable as the superhuman sensors. If LIDAR’s a crutch, computer vision still has a broken leg.

LIDAR’s big advantage is that near-100% reliability, a must when you plan to bet your life. You’ve seen videos of camera-based systems unable to see something as big as a truck on the road. That simply won’t happen with a LIDAR that’s still working, and “won’t happen” is pretty important.

Waymo Advantages

Waymo’s big advantage is that they have a working self-driving system, one that works in a variety of cities, and not only can you bet your life on it, 250,000 riders are doing so every week. That’s so far ahead of Tesla that you might wonder how people can even compare them.

The difference between tens of thousands of trips in a row and dozens is overwhelming. It doesn’t look like much on a logarithmic graph, but you can’t even see Tesla’s score on a linear graph. To be fair, the path of progress is often exponential, but it still has taken about a year for each order of magnitude, though Tesla claims they did two orders in 2024–progress does tend to be faster in the early stages.

One point sometimes made is that Waymo is only operating in a handful of cities, in restricted taxi service areas which some incorrectly call “geofences.” While Tesla isn’t self-driving on any public road it does do driver assist on most roads and conditions. Waymo is trying to build a robotaxi, and a taxi inherently has a limited service area, where it not only drives, but has a lot of associated infrastructure. Even Uber, which has humans do the driving, opened its taxi service in just one city 15 years ago and grew, city by city until today it has thousands. But it’s vastly easier for Uber to deploy than any robotaxi because the people do all the work.

Tesla fans may not realize that Waymo could also do a driver assist on every road, should they wish to, and in fact almost certainly much better than Tesla’s. Waymo’s so-called “geofence” is an advantage, which lets them solve one problem at a time, including the key one of safety, rather than the disadvantage it is painted as. Waymo’s cars use a map, which gives them better understanding of the road, with the opportunity to look at it from every direction and every distance multiple times. Tesla tries to build its map on the fly, from what it can see from a distance as it approaches. Wamyo’s AIs build the map using as much compute as they like, Tesla’s must do it in seconds with just what’s in one car. Then Tesla forgets what it learned, while Waymo remembers. Alphabet is the world’s #1 mapping company, and in fact the Streetview/Maps team was the team that built Waymo.

Waymo cars still have to drive–and with bet-your-life safety–when the road changes and the map is wrong and no longer useful. They do this every day. In addition, Waymo’s CEO has said their vehicle can drive even with just one sensor (such as the cameras) just as Tesla does, though it doesn’t reach the full safety level they would like, but neither does Tesla.

Mapping started expensive but it keeps getting cheaper. In fact, any car that can drive without a map (as Waymo can but Tesla can’t) is a car that can make a map-–you just have to remember it and your mapping is free. MobilEye, another competitor, builds its maps from data that comes in from 50 million regular cars equipped with MobilEye chips, ten times as many as Tesla, and mapped all of Europe in about a month without any cost of mapping cars.

While you never depend totally on your map, it helps a lot, and so far Tesla has decided to forgo that advantage.

Waymo gets the advantage of being part of Google, the world’s #1 brand, and leader in mapping, machine learning, navigation and mobile phone platforms, as well as search and advertising. The support of Larry Page, who founded Waymo, has given them the capital resources to get through big spending that killed other projects. Most people who summon a car from Uber, Lyft or even Tesla, do it on Google’s platform. (Their main impediment is possible anti-trust fear.) Waymo makes its own LIDARs and radars.

Waymo Disadvantages

Waymo’s hardware is definitely more expensive, but the overwhelming historical trend is that computer/electronics hardware costs get very small at scale. Their architecure is more complex and takes more coding effort. As noted, their use of limited service areas is more of an advantage than a disadvantage in the robotaxi world, it’s an issue if you want to sell consumer cars. Perhaps Waymo’s biggest disadvantage is its size and age. As noted below, that can make a company slower to resist change and make it hidebound. Waymo has to hire other companies to make cars for them, but that’s not difficult to do, but not as flexible as being a car company.

Real World Experience

It’s surprising how important real world experience of operating a robotaxi with no safety driver is. When Waymo and Cruise got their vehicles to the safety level they wanted, they deployed, in Waymo’s case even before Covid. The headlines quickly filled with problems on the streets, especially with fire trucks and other emergency vehicles. It was shocking how much there was left to learn even when the safety level was good. Tesla is over 5 years behind in taking out the safety driver and learning the reality, though they can learn some things from their competitor’s mistakes.

There’s also lots to be learned by taking real, paying customers. No battle plan survives first contact with the customer. Tesla has done some trial service for employees, with safety drivers behind the wheel, though soon it will take actual paying customers that aren’t invited press.

Waymo makes a driver platform, while Tesla makes cars. You can buy a Tesla car and drive it yourself, or with supervised FSD, almost anywhere. You can’t buy a Waymo though in the future Toyota may make one you can buy. They plan for now to sell trips, not cars. If you could buy a Tesla that self-drives, that would be a very different product. But it’s an incredibly hard product to make.

A robotaxi is a viable business if it can drive in a useful service area, even just one area. It comes back home every night for you to maintain it and update it. Private robocars leave the factory and ideally never come home, except digitally. But every customer wants to drive a different area. You can’t make a private car that only works in one city or a few cities, so you have to do vastly more work before you can sell car #1. (Well, Tesla did figure a way around that, pre-selling self-drive tech long before it was ready, but they can’t deliver until it drives over a wide area.)

One of the components Waymo is fairly advanced on is their remote operations team. While Tesla has a small team in action, which resolves problems the systems aren’t able to handle on their own, as far as is publicly known, they just started, while Waymo has over a decade of experience in this vital area.

Can Tesla Catch Up?

Quite often the second player can move much faster than the first. They learn from the pioneers’ mistakes. They get to use the latest tools and technologies. Deep Learning didn’t exist when Google Car began, its inventor then went to work for Google. The top pioneer in unsupervised Reinforcement Learning is DeepMind, a unit of Alphabet. Transformers (the key new technology used by Tesla, Waymo, ChatGPT and everybody else) did not exist until Google invented them in 2017. You may see a pattern there–While Tesla works hard to build using these technologies, Waymo’s parent Alphabet is often where they were born. Tesla licenses processor designs from good Silicon Valley labs, but Waymo is the only company which gets Google’s TPU AI accelerator, designed in-house.

That’s not an insurmountable advantage though. Often being first can make you hidebound. You invest too much in what you first built, and can be slower to innovate. Waymo isn’t using end-to-end machine learning like Tesla, but they say it’s because they tried it, and found it just didn’t have the power to deliver the performance needed. Nuro and MobilEye have said similar things. But they might be wrong.

One challenge is that machine learning approaches make mistakes–including the famous hallucinations of the LLMs–and nobody’s ever made one reach the near perfection that betting your life requires. Tesla’s quest is to be the first to pull that off.

Among automakers, Tesla is the king of innovation. But their competition here isn’t with stodgy OEMs. It’s with companies like Google and Amazon and Baidu–tech leaders and startups that are also great innovators. Motional, who is part of Hyundai, is a startup they bought to escape their old-world thinking.

In the end, it comes down to that gamble. Few will say it’s impossible for Tesla’s less complex, big-data approach to work. But it competes with approaches that are already working. The race is over on the question of which system will work first, the only question is if the pure-ML approach can develop faster, and take advantage of Tesla’s existing fleet and manufacturing capability to catch up. It could happen. For now it seems more likely it will not happen any time soon.

Some Tesla representatives have made claims that the safety performance of FSD has been greatly increasing of late–as much as 1000-fold in the last year, and that it will continue to do so. Unfortunately, there’s nothing to back that up in terms of user experience–remember that personal anecdotes reveal next to nothing–and the few who are trying to measure it see no such improvement.

We could know more if Tesla were willing to publish safety data. They publish safety data so misleading about Autopilot that they’ve been accused of deliberate deception, and statistics on FSD have not been released in any concrete form. Many wonder why, since they have good data, they refuse to reveal it, and that the lack of transparency suggests the data are poor. But only through bulk statistics can we judge the progress on these systems–even thousands of individual rides, no matter how much they impress, tell us little. Waymo releases the data. Nobody else, including Tesla, does the same.

In addition, even if Tesla were seeing such great improvements, the most likely explanation would be that they had so much room to improve. Most methods, including adding more data to machine learning systems, tend to “plateau,” and so far only combinations of methods have reached the necessary high levels. Again, that’s not a law of nature, but it is the typical pattern.

Unfortunately for Tesla, there are factors which may mean their existing fleet isn’t much of an asset. In addition to the question of the HW3 computers, Tesla’s cameras on older cars don’t show a full 360 view, or right in front of the car. They are quite old and low resolution cameras compared to the new ones. Tesla chose to pick the hardware first, then write the software, while every other team decided to use the latest hardware and not try to deploy it at scale until they got it working. Again, Tesla has taken the longshot bet that their old hardware pick will be right. If it is, they are in great shape, if not they must start from scratch.

Tesla Robotaxi Pilot

On June 22, Tesla launched a pilot robotaxi service south of downtown Austin, Texas. They had repeatedly promised it would be a real robotaxi service, with nobody in the car, and unsupervised by humans. They worked hard, but weren’t able to meet these promises, so they deployed a small service with around 10 Model Y cars with a human safety operator, moved to the passenger seat. There, the safety operator supervises the vehicle and can press 3 different buttons to stop it if things look bad. Like a driving school instructor, they can also grab the wheel to steer. They don’t seem to be able to command the vehicle to go when it gets stalled as somebody in the driver’s seat could. There is no operational or safety reason to put this person on the right side, it appears to have been done purely for optics, so they can present a car with nobody behind the wheel. They aren’t the first to use this trick, and it seems to work somewhat, but at its core this is no different from the pilots many other companies have done with an employee in the car. They could have done that a year ago.

Unfortunately, removing that human is “the big one,” the hardest step in a robocar’s life–the moonshot, so this is, for now, a big miss on their promise. We have no information to say how long they will need to be supervised. The early cars are making quite a few mistakes, but as of June 25, they have not had any crashes. Even on Day one, however, they did things like wandering into an oncoming lane, diving over debris, going over a curb and more. For now, rides are limited to influencers with a record of praising Tesla, but the videos some have released are unfiltered.

You may have seen a video Tesla released on a car delivering itself, empty, to a customer. I have more details on why a vendor-selected trip (which everybody has done) doesn’t demonstrate as much as people might think.

The pilot may have trouble in some types of rain–Tesla FSD also shuts down then–and it keeps to simpler streets, and easy pick-up spots. Otherwise they have done a fair bit of work on adding taxi features like pick-up and drop-off and entertainment for riders based on their profile in the Tesla app. All rides are $4.20, which is more of a marijuana joke than an actual price. It shows they are getting serious about the “non-driving” elements of a robotaxi service, of which there are surprisingly many.

The jury still remains out on Tesla’s fate, but for now, Waymo is the strong leader, and Baidu is the strong #2.

I would like all players to succeed, so we get robocars sooner, and with competition to promote innovation and low prices. I think it’s a shame, especially as a Tesla FSD owner, that they’re taking this longer-shot, slower approach. A few small changes in thinking about maps and extra sensors might have Tesla in the game much sooner. They also can afford to acquire working technology from one of the other companies that has it, though it won’t match Waymo.

I’ll close by again pointing out that while Tesla probably has a long way to go to get working self-driving, the companies that did build real self-driving have taken at least 5 years from that point to being ready to scale and go commercial. Maybe later comers can do that faster, but it won’t be overnight or in just a year. This leads to the conclusion that Tesla still isn’t really even in the race yet, or at best they’re in last place among the major players. If they can make self-driving work, they’ll be in the game, and with some advantages, but still a long way to go.

The above is the detailed content of The Deep Story On The Waymo Vs Tesla Robotaxi Battle, With Video. For more information, please follow other related articles on the PHP Chinese website!