Tesla's "Shadow" driving and how they test Autopilot

Tesla's simulator is not that important to them, they say

Continuing my series on Tesla based on their Autonomy day, here's an examination of their approach to the hard problem of testing driving software. Tesla takes advantage of the fleet of 400,000 cars deployed with their choice of sensor suite and all paid for and driven by customers. This allows them to test new software revisions in real world driving situations, though mainly in an instantaneous way.

Read my Forbes site article Tesla's shadow testing offers a useful advantage on the biggest problem in robocars


When Brad Templeton casts a shadow, it is scientifically called an eclipse. Hence his experience as an eclipse expert.

I was surprised you said Waymo might soon have a large enough fleet for this. I was optimistic 2+ years ago when they showed their "ready to deploy" low cost LIDAR suite. Sure enough, they started pulling safety drivers in late 2017 then ordered 60k Pacifica vans and announced commercial service. Momentum was building then . . . pffft. Other than some minor logistics news (new depot in Phoenix, small shop in Detroit to install sensors) their pace is glacial.

Is there anything specific that makes you think they're close to putting a lot more vans on the road?

Waymo "may soon have a very large fleet on the road," but unless they have plans I don't know about that fleet will likely be driving in a very limited number of locations.

The population of all of Arizona is about one tenth of one percent of the population of the Earth.

Relying on lidar is probably the single biggest mistake robocar companies are making.

I have seen no statements from Waymo that they intend to deploy those 80,000 vehicles in Arizona alone. In fact, they have said they plan to deploy in Silicon Valley next, and presumably have other plans in mind. They have said little.

It was just an example. The population of Silicon Valley is even less than the population of Arizona.

It's possible Waymo has plans to deploy a worldwide service soon and aren't saying anything about it. That doesn't seem to be their strategy, though. Building that many cars (each with lidar), and hiring that many safety drivers (or regular drivers in jurisdictions where driving with safety drivers isn't allowed) would cost a fortune that they don't have. Even if they teamed up with someone like Uber who already has a worldwide taxi service, installing sensors (including lidar) on the millions of cars would cost tens of billions of dollars.

Betting the farm on lidar (and skipping ADAS) were big mistakes. Relying on maps built specifically for cars was a mistake too, although that mistake was derived from the reliance on lidar. Even if they are the first to succeed technically, they won't be able to deploy in very many places for a long long time.

I don't think Waymo plans to have safety drivers once they deploy the 80,000 cars. That's beyond even the Waymo budget. Waymo manufactures their own LIDARs and at $3K per car it would cost $240M for the LIDARs on the 80,000 cars, which is definitely within the Waymo budget.

Should Waymo not be ready to deploy, but have the cars, several options are open to them, including simply running an Uber style service at Uber prices with the vehicles which would pay for their cost, while having them map and shadow test with the more sophisticated sensors, until they are ready. They could, however also offer a safety driver based Uber style service where the car is driving, like Waymo One, and pay for all those vehicles with it. Trained safety drivers are more expensive than Uber drivers, but not so much that Waymo can't afford that.

I have a Tesla. Autopilot makes mistakes. Every Tesla owner has experienced them. We don't have hard data on the frequency or severity of the mistakes, but they are at a much greater rate than Waymo's system makes mistakes. Probably 2-3 orders of magnitude.

If Waymo's system makes so many fewer mistakes, why don't you use that instead of driving a Tesla?

Because Waymo's system isn't available at all in the places where you need to go, of course. Comparing Autopilot (with Hardware 2.0/2.5?) with a Waymo car in autonomous mode is of very limited usefulness.

Even comparing a Waymo car in autonomous mode to a Tesla full-self-driving with 3.0 hardware wouldn't be a good comparison. It'd be like comparing chess software that only works if the opponent makes opening moves within its very limited opening database to chess software that always works, but occasionally asks for input from a human.

Neither one has solved the problem (as I defined it earlier, in the post where I forgot to log in first). But my bet is on the second method being able to solve it sooner.

(As far as whether or not Waymo plans to have safety drivers, I don't know. Maybe they'll at least have remote safety drivers. I assume the plan is to charge close to Uber prices at first. Demand might even be high enough that they could charge a little bit more than Uber. Even then they'll likely be losing money for every mile, but as long as they aren't losing too much per mile that'll probably be okay. The point is to do this for R&D purposes, right?)

Waymo will have remote operators, but not remote safety drivers as far as I know. Tesla probably will as well.

If somebody offers me a taxi ride (robotic or otherwise) for 60 cents/mile I will take it over the Tesla, because I can get stuff done on the ride, and not have to pay for parking. If they offer me that service around town, and will drive me to a waiting regular car (or Tesla) when I want to go outside the service area, I would give up car ownership for that, as would many.

I'm not sure that anyone is going to offer you a taxi ride for 60 cents a mile ever. Not for short rides, anyway. Do you think anyone is anywhere close to that?

I find it interesting that you think Waymo is going to deploy tens of thousands of completely driverless vehicles carrying paying passengers any time soon. Are you basing this on something public?

If remote drivers would be safe enough, this could be a way for Tesla to provide an inexpensive short-term rental service that comes to pick you up. Use remote safety drivers while deadheading, and then just run in ADAS mode while someone (who would have to be a licensed driver) is in the car. I'm not sure that remote drivers would be safe enough, though. And who knows if it'd be legal (I wonder if current state laws require the driver of a vehicle to be inside the vehicle).

We just know what Waymo has said, that they have 80,000 cars on order, that they are making a factory in Michigan to finish the manufacturing of them. So they definitely plan to do it. Depends on what you mean by soon.

60 cents for a short ride does require that you distribute the cost of vehicle moves into all miles, rather than per ride, which is a subsidy of short rides. However, there may be reasons to subsidize short rides. However, with single person taxis, 60 cents for a short ride is not out of the question. 60 cents per mile for medium rides is very much a reasonable and profitable number.

Could Waymo test V2V "between company vehicles" in stealth mode in current Phoenix valley fleet and not have to make test public?

Airwaves are regulated, cannot imagine any way to sidestep

The spectrum for V2V is allocated and they could use it, but they have no particular desire to use it, nor should they. There is nothing they could use v2v for.

We know Waymo said they ordered a bunch of cars. What exactly they're going to do with them, if in fact they get them, is speculation. Also, this was back in June of 2018, and Waymo seems to have suffered some setbacks since then. I'll leave "soon" open-ended, but having 80,000 cars driving a significant number of miles completely driverless within 3 years would be surprising.

Single-person taxis are science-fiction at this point. As far as I know no one is even starting to build them yet.

This is as expected. Everybody but Zoox and a few shuttle companies knows the right strategy is to solve one problem at a time, so you base your vehicles on existing popular designs (like the Pacifica or the Bolt or the Tesla.) Google always said the Firefly, a 2 person taxi, was just an experiment.

The narrow 1-2 person taxi comes only once there is competition on price. When there is no competition, you charge $1/mile or more. Once there is competition and prices drop, you start realizing that 80% of the rides you sell are solo, and you could serve them in something cheaper, at a more competitive price.

Brad Templeton you stated: "If you're making one, you must first convince yourself, along with your lawyers and board of directors, and you must convince the public. When you have accidents, you will need to convince the courts. In time, you will need to convince regulators too."

Why did you leave out insurance companies? Did you forget their necessary buy-in or did you have another reason?

Large players like those making these cards do not need or want insurance companies, other than for their expertise at claims handling. They will self-insure. The people making and testing the cars know about their risk, the insurance actuaries know nothing about it.

Percentage comparison of ICML papers by Industry vs. Academia by Kaparthy needs a follow up

Add new comment