Waymo announced today that it would shortly begin service in Austin, TX. Cruise also announced it will serve Nashville, following the announcement of plans (with safety drivers, for now) for Dallas and Houston. Waymo is waiting on a permit from the California PUC to do service in Los Angeles.

Both companies are demonstrating their ability to scale and serve new cities. It is somewhat surprising to see Waymo pick Austin, where Cruise already operates, since one would imagine a desire, at least when in full production, to operate without competition in a city rather than go head to head. At this time, neither service is really in commercial operation, and they are both pricing in a manner similar to services like Uber. Both services have fairly high costs, and lose money on operations, though GM recently reported that Cruise has been dropping those costs at a rapid rate, and hopes to see them reach around $1/mile fairly soon. (A base cost of $1/mile can probably be sustained at the Uber level retail price of $2/mile.) As such they are not expanding for financial reasons.

These recent expansions are interesting in relation to a frequent argument made by supporters of Tesla FSD, or rather what they hope FSD can become. They argue that Tesla is a leader in the self-driving space because FSD is operating in a supervised mode on almost all streets in the USA, and they say the robotaxi providers are inferior because they currently are “geofenced” to a few cities, even though their level of service in those cities is many thousands of times better than Tesla’s performance if you try to judge it as a self-driving system. At present, Tesla FSD can not self-drive on any street. It only will sometimes complete a drive without intervention. That’s a driver-assist system, and it performs well as one. In the self-driving world, “drive” means “drive so reliably I could go to sleep” and in particular it means “need a safety intervention less than once in a full human lifetime of driving, “which is in the range of 20,000 trips. Being able to complete 2-3 trips is very distant from being able to do 20,000 — and most self-driving teams want to surpass the human number and are probably looking to see 50,000 trips between safety interventions.

To those in the self-driving space, the current Tesla performance level is not even on the same planet. It’s like saying, “Your Tesla car can only drive on roads, which are a very limited geofence. My horse can do any road or trail and is thus clearly superior.”

Elon Musk expresses unending optimism that his team will increase the safety performance of Tesla FSD, and every year — including this year — has predicted it will reach self-driving performance levels that year. Let’s imagine for the purposes of argument that this can in fact happen in some small number of years.

The robotaxi companies have constrained their operations to specific service areas for a variety of reasons:

1. One must gather map data on the new area. Thanks the the improvement of crowdsourced map approaches, this has become low cost — MobilEye brags to mapping most of Europe in a month.
2. One must test the system in the new area, or take the risk of it driving on a road with an unexpected feature that has never been seen or tested before. Problems found must be fixed and confirmed to be very rare
3. One must make arrangements with governments to operate in the area
4. One must have resources and staff to support the new customers in the new area, unless one is only allowing old customers to travel in the new area.

A driver-assist system is saved most of these burdens because it can always fall back to the monitoring driver to deal with problems, even urgent ones. An uncrewed car can only fall back to remote assistance operators for non-urgent problems where it has the ability to pause and wait for that assistance.

This makes such a large difference in the problem that it’s likely that companies like Waymo and Cruise could, should they feel motivated to, produce a system would could drive in random new territory with just navigation maps and a supervising driver. Indeed, with their greater experience and superior sensor suites, it seems likely this system would perform far better than Tesla FSD, though that can’t be said with certainty as they have not done it and are unlikely to wish to. The Tesla system starts from a lane geometry map, and looks at sensor data to attempt to build a more complete and accurate map on the fly. It then drives based on that map. Many of the errors made by Tesla FSD systems are caused by errors in the on-the-fly map, but it also makes errors when trying to solve the same problems as any other car — perception, prediction of where things are going, and planning.

A challenge for an attempted “drive all roads” system is problem #2 above. Without extreme confidence, nobody wants to ride in, or send out a vehicle, to drive a road or situation which has never been tested before.

One way to compare the systems more readily would be to see how well the Tesla one performs if given a fully correct map, and how well the other systems perform if required to build a map on the fly. Teslas are constrained to using only computer vision, while most other teams use vision (usually with significantly superior cameras) as well as LIDARs and radar. It seems likely that with this superior sensor suite they would do well at mapping on-the-fly. When they come to roads that have changed from their map, they do mapping on-the-fly, but only need consider the regions that have changed — because they keep a fair bit of detail in their maps, they know precisely what parts of the road are the same as their map, and which are different and need reinterpretation. Tesla FSD has a somewhat harder job — it’s lane map may say there are 3 lanes, and its live mapper may see only 2, and it has to decide if the road has been restriped, or the live mapper is making one of its frequent initial mistakes. (It is normal for Tesla’s live mapper, visible on the display screen, to make a poor first guess at the geometry of a distant piece of road as it approaches, and then refine it to a better result as it gets closer.)

Waymo claims their techniques are highly scalable to other locations. Waymo group product manager Aman Nalavade says, “By leveraging cutting-edge ML, we’re building a system that can generalize from one place to another, leading to quicker scalability. In Los Angeles and Austin, the Waymo Driver demonstrated solid performance right from the outset, serving as a testament to the effectiveness of our approach.”

Tesla’s approach does take advantage of a valuable asset — the hundreds of thousands of Tesla drivers who test out the FSD software. They can greatly speed up the process of seeing how well the system does on segments of road and whether it makes mistakes requiring intervention. Data from such events can generate trouble tickets to be fixed, and provide data to load into training new machine learning models. When it comes time for Tesla to certify performance on roads, they will be able to do that more quickly and thoroughly than companies that need to send out cars, or pay drivers, to do it.

At the same time, Tesla owners are also generating problems at a rate vastly faster than any team of human coders could keep up with. The hope is that a self-supervised machine learning system might be able to make use of this data. Tesla is building a very large supercomputer which it hopes to use for this purpose — some other companies in the space, like Waymo (an Alphabet unit) and Zoox (an Amazon unit) as well as NVidia have access to such computing already, others have to rent it.

This large source of data should be helping Tesla more than it appears to, but they always promise more for each new release.

Tesla may find there’s quite a lot more to do once they get to the point of being able to perform tens of thousands of drives in a row. The last year’s headlines have been full of stories of issues with Waymo and Cruise and the cities they drive in, notably San Francisco. These started with Pick-Up/Drop-Off (PuDo) issues as the cars often do not pull over to the curb for this, and still don’t in many cases. There have been a number of incidents where cars stalled, blocking streets, waiting for a remote rescue. A lot of attention has come on problems at emergency scenes or with emergency vehicles, where vehicles were accused of blocking vehicles and disrupting the scenes.

We haven’t heard about any of these issues with Teslas, not simply because Tesla is not yet attempting to operate a ride service. If a Tesla ever behaves improperly to block traffic or at an emergency scene, the driver takes the wheel and resolves the problem — people may be unaware it was even in FSD mode. When Waymo and Cruise operated with safety drivers, they may have encountered these problems but there would never have been complaints because the safety drivers immediately dealt with them. If Tesla plans to operate a robotaxi service or let vehicles operate with nobody in them, all of the learnings from the last few years for Waymo and Cruise lie ahead of them.

Tesla hopes that they will be able to use machine learning to resolve these problems. That remains to be seen.