Today Waymo has commercial service only in suburban Phoenix. AutoX is doing it near Shenzhen. Tesla TSLA -1.2% declares their “full” self-driving will operate just about anywhere, in the even they get it to work anywhere. Other teams are doing testing in a limited number of cities and environments. But everybody knows to win the big prize you do need to cover the needs of a lot of customers — not driving everywhere, but doing a decent number of larger cities.
What will Waymo and other companies have to do when they want to operate in a new city, new climate or even new country? Will it block their progress or be a straightforward incremental improvement? My colleague Michael Sena wonders in his monthly Dispatcher whether a taxonomy of ODDs must be produced to help sove the problem.
When NHTSA and the Society of Automotive Engineers created some counterproductive “levels” of self-driving defined by the role of the human being, they also added a more useful concept, termed the “Operational Design Domain” or ODD — the set of roads, times and conditions in which the car can operate. Right now most vehicles operate in a fairly limited ODD and are in pilot projects. Driver-assist systems, like Tesla Autopilot and it’s beta “full’ self-driving operate in much wider areas because they can fall back on the supervising human driver. Trucking companies are largely operating only on freeways and a few roads from the freeways to trucking depots. One startup player, “Voyage” was operating only on private land in a retirement community — and just sold itself to Cruise in what may have been an acqui-hire since no terms were announced. Delivery robot companies like Starship operate only on sidewalks and crosswalks, also a simpler ODD that can be solved sooner.
The ODD is not just a set of streets. it can also be limited by time (for example, only at night) or speed (only below 50 km/h in a traffic jam) or it can include weather — most vehicles don’t drive in snow, though Yandex YNDX +1.2% just announced a large track record there in Moscow.
People often incorrectly call the ODD a “geofence.” The domain is not strictly geographic — it’s typically a network of streets, and many not include all streets within a region, or all lanes on every street. The fence is more about how difficult the street is to handle than where it is, and the only thing geographic is that it’s useful that the streets be connected so you can drive between them.
Is it going to be too hard? My colleague Michael Sena wonders in his monthly Dispatcher whether a taxonomy of ODDs must be produced to help sove the problem.
- Enumerate and handle the local rules, conventions and conditions
- Learn patterns from crowdsourced data
- Make friends with the local officials
- Drive and map the streets and encode all official and unofficial rules into the maps
- Buy the fleet, arrange logistics and staff to run it (the biggest cost.)
- Test everything extensively in simulator and then on the roads
To drive in a new place, a robocar is certainly going to need to know the different local rules of the road. For most teams, this starts with having the detailed maps the car uses to drive. Almost everybody except Tesla expect to have a fairly detailed map showing all the roads, lanes, signs and traffic rules. Even Tesla needs both a regular navigation level map, and some more detailed mapping of complex areas like difficult intersections.
For those teams — most teams — who depend on a map, they must build it. This costs money, both to build and maintain it, but generally it’s well within the budget for operating in a new city. Mapping costs per mile, but cars will drive that mile again and again to make it scale if your fleet is large enough. This is a challenge, though, for anybody wanting to just “drive everywhere” as you can’t map everywhere all at once very easily, and it’s harder to recover the cost.
Yes, different regions have different rules. The “Pittsburgh left.” The crazy drivers of Boston. The crazy pedestrians and cabbies of Manhattan. The snow-driving habits of Montreal. But the list of these is quite finite, and well known to anybody who has come to the city from outside. It’s not intractable.
One solution to that is relying on crowdsourced data — information that comes from human driven cars already driving the roads and recording what they do and see. Tesla has access to that, as will any other car company with a large fleet of customer cars that they can update software in. Another company which does this is MobilEye, the Intel INTC -0.6% unit which makes camera/chip systems for driver assist already installed in many millions of cars.
It also makes sense to make nice with the local authorities. In the more distant future, the arrival of robotaxi fleets will be a common thing, and cities will know how to handle it, with state and national laws answering most of the questions, but for the first 100 or so cities it will be new territory for the city and you can’t just hit them by surprise.
If you’re going to have a robotaxi fleet, you must also buy that fleet. That’s a huge expense. A fleet needs to be very large if you hope to make a big, profitable service in the city — tens to hundreds of thousands of cars. That’s billions of dollars, and makes the cost of the mapping, city relations and other factors pale in comparison. This is one reason teams are not that scared of this cost.
Another cost will be testing and certifying in the new location. There may be different government regulations, but even without those, just making sure your new maps and new code to understand locations are good enough to bet your business on is a significant cost. You are betting your business — if you get something seriously wrong and hurt somebody, things could go very badly.
Nobody has deployed a production service except Waymo and AutoX, and they only did so after a great deal of testing. It is yet to be learned just how much extra testing will be needed to conquer a new city, or a radically different new city. It’s not likely to be nearly as much as doing it the first time, though even that might be affordable once streamlined.
So what’s involved in understanding what’s different about how to drive in a new ODD? We know that human beings can generally plunk down in any other city not too different from where they live and pick up a rental car and go successfully. We can handle a fair bit — Americans can handle the left-side driving, small roads, small cars and non-English signs and rules of Japan, but might have trouble with the chaos of India or Indonesia added on to that. Of course, humans have far more intelligence than any robot, so the question is, does being this adaptable require something unique to humans, or can robots do it, particularly with a month’s help from skilled humans?
The variety of driving styles in the world is not infinite. The variations in rules are enumerable. The task may be reasonably suitable to robots given a leg up by humans. After all, for a long time, people thought the task of driving even in a suburb of Phoenix was something clearly only humans were up to.
One demonstration was done recently by MobilEye. MobilEye has been training their self-drive system mostly in Israel, which is a fairly chaotic place to drive (though again, within the abilities of the typical western driver.) To do a demonstration in Munich they sent two cars and two non-engineers to Munich and had them drive around for 2 weeks gathering data. Engineers back in Jerusalem analyzed the data. More importantly, they combined it from data learned from huge numbers of ordinary cars driving around Munich for years with MobilEye cameras on board. MobilEye chips are in many BMWs which are everywhere in Munich. All that data had already produced crowdsourced maps of the roads and objects around them. They knew to go where a thousand BMWs had gone before and not to go where none of them went.
MobilEye reports they then had their car working. Not perfectly — I doubt it’s ready to deploy for taxi service since it isn’t ready for that in Jerusalem either. But they now claim that they can bring up a new city in a couple of weeks when the time comes.
Waymo has spent more work moving to their new territories. They do a lot of mapping in their custom cars, wanting to drive each street at least a couple of times. Everybody is working hard at making the initial map generation as automatic as it can be, so humans only have to tweak a few things here and there. Waymo hasn’t released data on how long it takes them to map or what it costs.
Several other companies are also mapping, or helping customers map quickly. (The author has a consulting relationship with DeepMap, one of the leaders in that space.) Nobody seems scared of the cost compared to the other large costs of conquering new territory.
Taxi vs. Private Car
Nobody is scared among the robotaxi providers. If you want to sell private cars, that’s another story. A private car with a self-driving feature that only works in a few towns isn’t really an option for a big car OEM. They don’t want to sell models that only work in one town, while a robotaxi operator can be quite happy doing that, at least to start. They have a bigger problem because they have to drive most of the cities in a whole country before they can put a car on the market.
Tesla is of course the most unusual car company, but they still have that need to produce a car that will work everywhere. Their driver-assist offerings easily work in all locations by relying on the human. But their ambition is to produce a real true actual “full self driving” some day. The beta version, which is a driver-assist tool, is nonetheless impressive for its ability to at least do something in most of the United States. It’s also very far from ready. In order to drive without a map or other special local knowledge, it has to build its map “on the fly” as it drives. They display the map it is building to beta testers, and it’s usually pretty good — but far too often for safety’s sake it’s wrong, and interventions are needed to stop the car from making a mistake. Tesla makes bold promises that it will get better, good enough to bet your life on, but there is as yet, no sign of that.
Videos of the Tesla beta often show it handling a decent drive with no interventions, but many show it not doing that. An hour’s drive is nice, but for real self-driving, you have to drive for almost a full human lifetime of driving without any serious mistakes.
The robotaxi companies can be more picky,. You can have a perfectly profitable robotaxi service handling just the most lucrative regions of a city, particularly if you can offer human-driven rides to the less popular zones (as Uber UBER -2.5%/Aurora and Lyft LYFT -1.4% can do.) You can have a great service in just the places with no snow, and do the snowy cities later — the only risk is that somebody else might claim them first and attempt first mover advantage. You can do just one country and handle the other countries when you are ready, again with only that risk of losing first mover advantage.
You can even offer a car-replacement service without a complete network, as long as all the popular destinations for your customers (including of course their work and home) are on your network. If somebody lives off the network, they just aren’t a customer, at least for car-replacement. There isn’t a rush, and whatever it costs to add new territory should be small compared to the value of the territory — or nobody bothers to claim it.