A16Z partner Ben Evans: Accidents occur frequently, and fully autonomous vehicles can actually be implemented in “levels”

This way of discussing it seems pretty straightforward and clear, but not if you think about how deployment might actually be possible and the fact that some places might be easier to drive than others.

As we've seen from early tests of prototype self-driving cars, any level of autonomy works differently in different cities—Phoenix is easier than San Francisco, which is easier than Naples or Moscow. This diversity is reflected not only in different cities, but also in different urban landscapes: highways are simpler than city centers, which may in turn be easier or simpler than suburbs.

Naturally, cars will achieve certain levels of automation capabilities sooner in some ("easier") places than in others. There are huge safety and economic benefits, so we're going to deploy this car - we're not going to do nothing until a perfect L5 car can drive itself from Kathmandu to South Boston. So, if we call a car L4, what are we talking about? What we probably mean is "suitable for autonomous driving in most parts of the country." But it’s more likely that L4 will fit in one block, L3 in another, and only L2 in a third – a car might encounter all three on a journey. Put your route on the map and it will tell you whether today is a day where you can use L5 mode.

So, for example, several companies are developing long-distance driverless trucks. Since these trucks are mainly driven on highways with greater environmental constraints than city streets, their mileage accounts for 80-90% of the total mileage of long-distance trucks. So It’s worth doing just that part of the highway, even if you need human drivers at both ends of the route, like a navigator guiding a ship into a harbor. If you look at it this way, the truck is in the L4 or L5 driving mode on the highway, and in the city road it is the L2 or L3 driving mode.

Suppose, instead, that a medium-sized city announces that self-driving will only be allowed in its downtown area during the day, and that only driverless buses, cars, and taxis will be allowed in. If those autonomous taxis never leave these reservations (and never hit the highways), they might look more like golf carts than cars. If they can handle conditions in central Cambridge without assistance (as long as there is no human driver), but not on suburban roads, but never leave central Cambridge, do these cars count as L5? What exactly does “all use cases” mean—all use cases for all cars, or all use cases for this car?

Then, imagine a situation where a garbage truck can drive itself on the street, but requires a human driver to return it from the street to the recycling station. Is this considered "fully automatic"? Or is it considered "advanced cruise control"? Does it matter?

These may seem to be purely definitional questions, but it can easily lead to the question of "when will autonomy be achieved?" But "when" depends on "where" and "what" autonomy)” and it is a matter of gradual and incremental deployment of many different models in different places.

In addition, this deployment process will also affect the final situation. The picture at the beginning of this article reminds me painfully of some discussions I had with people in 1999 and 2000 about the future of the "mobile Internet." Imagine a group of people from operators sitting down in 1998 and trying to define what "partial multimedia terminals" and "advanced multimedia terminals" would look like, what technical specifications to follow, and then tell the manufacturers, or actually "suppliers" "Businessmen" will produce products that comply with this specification 10 years in advance. That’s not how things work out—rather, the process of technological creation shapes the final outcome. We tried many paths and found the one that worked (by the end the carriers were no longer in control and the "suppliers" had all changed). We don't start with a predetermined conclusion and then proceed. Similarly, I sometimes think that the discussion of L4 or L5 also presupposes the shape of the result - but in fact the final result is shaped by the process.

Obviously, the problem with this multi-faceted, incremental, process-based model is that there are actually some pretty binary issues with automation - "Can I stop looking at the road and look at something else? Can I be behind the car?" Sleeping in your seat?" A progressive model can only take you so far - you really need to tell people whether the car is "autonomous driving." A car that needs to switch from "autonomous" driving mode to manual mode as shown in the figure presents some significant challenges: Humans are not good at taking over driving at high speeds and paying no attention to their surroundings. On the contrary, some people think that L1 is actually safer than L2 or L3 - at L1, you clearly know that you are always driving the vehicle, but sometimes the car will apply the brakes to save your (or someone else's) life when you are not paying attention. .

So that's why there's so much work to be done on how the car communicates with the user - how does it say, "This is a Level 5 journey, you can take a break," or "The next hour is with me. Come and drive, I will notify you 5 minutes before I need you to take over”? Will that self-driving golf cart refuse to cross an invisible line into a neighborhood that’s not self-driving certified? Can you throw the uncaring robot taxi driver out of the car?

The robot driver that made Schwarzenegger furious

Editor's note: There is a scene in the science fiction movie "Total Recall" starring Arnold Schwarzenegger, where Arnie has a conversation with the "driver" in a self-driving taxi.

Driver: "I'm Johnny Cab, where do you want to go?"

Arnie: "Drive! Drive!"

Driver: “Can you tell me the address again?”

Arnie: "Anywhere, just drive! Drive quickly!"

Driver: "Please tell me the specific street and house number."

Arnie: "Shit. SHIT!"

Driver: "I'm not familiar with this address. Can you repeat it?"

In the end, Arine was so angry that she threw the robot driver out of the car and drove herself to avoid pursuit.

As with most issues surrounding autonomous driving these days, there are far more questions than answers - again, it's like trying to predict the shape of smartphones in 2018 in 2000. But my point here is to say that even the terms discussed can be misleading. It is almost certain that the first “autonomous car” with an “L4 certification” label will not be launched in 2023 or 2027. There may never be the "first" L4 or L5 car, or there may be many different "firsts".

Original link: https://www.ben-evans.com/benedictevans/2018/3/26/steps-to-autonomy

Produced by the compilation team. Editor: Hao Pengcheng.