Will the winner get everything in the world of robomobiles?

Original author: Benedict Evans
  • Transfer
image

Today, several dozen companies are trying to create technology for robotic vehicles - these are OEMs, their traditional suppliers, existing large technology companies and startups. Obviously, not everyone is destined to achieve success, but a sufficiently large number of them have every chance - therefore, it will be interesting to think about what the consequences of the “winner gets everything” effect will be and what leverage can be in this area. Will there be a network effect due to which one or two of the largest companies will squeeze out the rest, as happened in the world of smartphones or OSes for PCs? Or is there a place on the market for five to ten companies that will compete for a very long time? And what floors in this pyramid will victory bring power over other layers?

Such questions are quite important, for they indicate a balance of power in the auto industry of the future. A world in which carmakers can buy products such as turnkey autonomy from any of five to six companies (or make one yourself), just like they buy ABS technology today, is very different from the world in which Waymo and Uber probably remained the only real candidates capable of independently building a business model, as Google did with Android. Microsoft and Intel found pain points in the PC world, and Google - in smartphones; What kind of points can robots have?

It is immediately clear that the goods will be equipment and sensors for autonomy. There are a lot of engineering efforts and scientific research in them, just like, say, in LCD screens, but there is no reason to use one instead of the other just because everyone else does it. The economies of scale are quite powerful, but there is no network effect. Therefore, let's say LIDAR goes from a $ 50,000 “rotating bucket from KFC” to a small widget without moving parts worth several hundred dollars, and there will be winners in this segment - but there will be no network effect, because the LIDAR winner will give you there are no more levers of influence on other floors of the pyramid (unless you can capture the monopoly) than the best Sony photomatrixes (which she sells to Apple) give her in the world of smartphones.

On the other hand, there will hardly be direct parallels with ecosystems of third-party software developers, such as they currently exist for PCs or smartphones. Windows squeezed out Mac, and then iOS, and Android squeezed out Windows Phone thanks to a vicious circle tied to developers, but you won’t buy a car based on how many applications you can run on it. Uber, Lyft and Didi will work on all of them, and everyone will have Netflix built into the screens, but other applications will work on your phone (or watch, or glasses).

You need to search not in the cars themselves, but higher in the pyramid - in stand-alone software that allows the car to move along the road, without crashing into anything, in city-wide optimization and in building routes, during which we can automate all the cars in the system, and not only each separately - and in the wake of all this, Robotaxi parks will exist. Network effects in the field of automobiles on demand are obvious, but with the advent of autonomy they will become much more complicated (autonomy will cut the cost of any trip on demand by three quarters, or even more). Robotaxi fleets will dynamically redistribute their cars, as well as coordinate with each other, and, probably, together with all the cars at once, their routes in real time to achieve the greatest efficiency in order to avoid situations, in which all cars simultaneously choose one route. This, in turn, can be combined not only with fluctuating pricing, but also with path-based pricing - you may need to pay more to get to the place faster during peak hours, or choose an arrival time depending on cost.

From a technical point of view, these three floors (movement, paving the way and optimization, leaving on demand) are quite independent from each other - presumably, it will be possible to install Lyft in the GM car from GM and allow the Waymo module installed in the system to control the car. Obviously, someone is hoping for levers of pressure to appear at different levels, or, possibly, combining them into one package - Tesla plan to ban people from using their robomobiles with other travel services on demand, except for their own. It doesn't work the other way –Uber will not insist on using only its autonomous system. But while Microsoft used joint leverage in promoting Office and Windows, both of these positions won first place in their own markets as part of their own network effects: if a small OEM insists on using its small robotax service, it will look as if in 1995 Apple insisted on buying AppleWorks instead of Microsoft Office. I suspect that the result will be a more neutral approach. Especially if we have long-distance coordination of car traffic, or even direct communication between cars at intersections, then some common floor of the pyramid will be required (although I am predisposed to decentralized systems).

All this, of course, is a complete theorization, and looks like attempts to predict today's traffic jams, living in 1900. The only area in which we can talk about key network effects is autonomy itself. It all depends on equipment, sensors, software, but primarily on data. And in the case of autonomy, two types of data matter - maps and driving data. Let's start with the cards.

Our brain constantly processes data from the senses and builds a three-dimensional model of the world around us, in real time and at the subconscious level - when we run through the forest, we usually do not stumble on the roots and do not bang our heads on the branches. In the world of autonomy, this is called SLAM.- Simultaneous Localization and Mapping (Method of simultaneous localization and map building). We are engaged in marking up our environment and localizing ourselves within it. This is obviously a basic requirement for autonomy - the robomobile must understand where it is on the road, what can surround it (rows, turns, sidewalks, traffic lights, etc.), and it must understand where other vehicles are located and how fast they move.

So far, implementing this technology on a real-time road in real time is a pretty difficult task. People drive using vision (and sound), but building a fairly accurate three-dimensional model of your environment on the basis of only one image (especially a two-dimensional) remains an unresolved problem: machine learning makes this possible, but so far no one has achieved the accuracy necessary for this driving. Therefore, we use workarounds. Therefore, almost all autonomous projects combine image processing with LIDAR 360 degrees: each of the sensors has its own limitations, but their combination (“sensor fusion”) helps you build the whole picture. Building the world around you with just images alone will surely become possible at some point in the future, but the use of a large number of sensors allows you to speed up this process, even if you have to wait until the cost and form factor of these sensors are reduced to practical. LIDAR is a workaround for building the world around you. Once this is obtained, then machine learning is often used to understand what this world is like - is this form a machine, a cyclist. But in this case, the network effect is not observed - you can collect quite a lot of images of cyclists, without even having a fleet of cars. what this world is - whether this form is a machine, a cyclist. But in this case, the network effect is not observed - you can collect quite a lot of images of cyclists, without even having a fleet of cars. what this world is - whether this form is a machine, a cyclist. But in this case, the network effect is not observed - you can collect quite a lot of images of cyclists, without even having a fleet of cars.

If LIDAR is one of the workarounds to SLAM, then the other, more interesting one is pre-built maps, that is, high-resolution 3D models. You go around the road in advance, calmly process all the data, build street models, and then put them in any car that goes along it. Then the Robomobile no longer needs to process all this data and extract turns or traffic lights from them against the background of the rest of the noise in the real world at a speed of 65 mph - instead, it knows where the traffic light is and can localize itself inside the model by key milestones roads at any time. Therefore, your car uses cameras and LIDAR to determine its position on the road, search for traffic lights, etc., comparing what he can see with a previously created map, and he does not need to do this on his own and from scratch.

Cards have a network effect. When any robomobile moves on a previously marked road, it compares the road with the map and updates the map. Each robomobile can simultaneously be a panoramic car. If you sold 500,000 robomobiles, and someone else - 10,000, your cards will be updated more often and more accurately, so your cars will have less chance of meeting something completely new and unexpected. The more cars you sell, the better they behave - a network effect by definition.

The risk of this is that in the long run, all machines will be able to carry out SLAM without LIDAR, and drive without pre-installed cards - because people can do this. Whether this will happen at all, and when exactly, is not yet clear, but now it seems that this will not happen immediately after the first robomobiles are on sale, and by then everything will change.

If cards are the first network effect in the data, then the second is obtained from what the machine does after it understands its surroundings. Driving on an empty road or on a road filled with other robomobiles is one task, but determining what other people are going to do on the road and how to deal with it is a completely different task.

One of the breakthroughs that support autonomy is that machine learning should be a good fit: instead of trying to write down complex rules that explain what you think about how people can behave, MO will use the data, and how the bigger the better. The more data you can collect about the behavior and reactions of real drivers in the real world (both about the behavior of other drivers and the behavior of the drivers of your own sightseeing cars), the better your software will understand what is happening around, and the better it will plan next steps. As with the cards, before launching your test cars collect this data, but after launching, every car you sell collects this data and sends it home. So just like with cards, the more cars you sell,

Driving data can be used in another place - in the simulation. This should answer the questions “if X happens, how will our stand-alone software behave?” One way to answer it is to make a robomobile and send it on a trip around the city to see how it reacts to the random actions of other drivers. The problem is that such an experiment is not controlled - you cannot return to the same situation with new software and see what is changing and whether the problems are fixed. Accordingly, now a lot of efforts are aimed at creating simulations - put your software for the robot car in Grand Theft Auto (almost literally) and check it out as you like. This will not necessarily help you catch all of the event options (“will LIDAR detect the presence of this truck on the road?”), And some simulation scenarios will be looped, but it will tell you how your system will behave in certain situations, and you can collect these situations from data on trips in the real world. So there’s an indirect network effect: the more data you have about traveling around the real world, the more accurate your simulations will be, and therefore the better your software will be. The simulations also show obvious advantages depending on the scale - how many people will work on this, how much computer resources you can devote to this, and what theoretical experience you have in working on large-scale computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. and you can collect these situations from data on trips in the real world. So there’s an indirect network effect: the more data you have about traveling around the real world, the more accurate your simulations will be, and therefore the better your software will be. The simulations also show obvious advantages depending on the scale - how many people will work on this, how much computer resources you can devote to this, and what theoretical experience you have in working on large-scale computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. and you can collect these situations from data on trips in the real world. So there’s an indirect network effect: the more data you have about traveling around the real world, the more accurate your simulations will be, and therefore the better your software will be. The simulations also show obvious advantages depending on the scale - how many people will work on this, how much computer resources you can devote to this, and what theoretical experience you have in working on large-scale computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. the more data you have about traveling around the real world, the more accurate your simulations will be, and therefore the better your software will be. The simulations also show obvious advantages depending on the scale - how many people will work on this, how much computer resources you can devote to this, and what theoretical experience you have in working on large-scale computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. the more data you have about traveling around the real world, the more accurate your simulations will be, and therefore the better your software will be. The simulations also show obvious advantages depending on the scale - how many people will work on this, how much computer resources you can devote to this, and what theoretical experience you have in working on large-scale computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. how much computer resources you can devote to this, and what theoretical experience you have in working on large computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations. how much computer resources you can devote to this, and what theoretical experience you have in working on large computing projects. The fact that Waymo is part of Google definitely gives it an advantage: its robomobiles wind 25,000 real miles every week, and in 2016, they traveled an average of 19 million miles a week during simulations.

We can say that Tesla has an advantage both in terms of maps and driving data: since the end of 2016, those cars whose owners bought the “autopilot” add-on have eight cameras providing 360-degree visibility supported by the radar, directed forward (and there is also a set of ultrasonic sensors working at a fairly short distance and mainly used in parking). All this can collect both cartographic data and driver behavior, and send them to Tesla - and, apparently, the company has recently begun to collect such data. The catch is that since the radar is only forward, Tesla will have to use only image processing to build models of most of the world, but, as I noted, we still do not know how to do this with sufficient accuracy. That means Tesla, in fact, it collects data that no one is able to read today (or at least read well enough to work out a complete solution). Of course, this problem will need to be solved both for data collection and for driving the car itself, so Tesla makes an uncharacteristic bet on the rapid development of computer vision. Tesla saves time by not expecting the appearance of cheap / practical LIDAR (today Tesla would not be able to place them on all of its cars), but without LIDAR, computer vision software will have to solve more complex problems, so it can also take a long time. And if all the other parts of the software necessary for autonomy - the parts that decide what the machine should do in a certain situation - will not appear soon, during this time LIDAR may become cheaper and become more practical, and the Tesla workaround will lose its meaning. We will see.

So, the network effects - the winner gets everything - are in the data: in the driving data and in the maps. This raises two questions: who will receive this data, and how much do we need?

Data ownership is an interesting issue of power and value. Obviously, Tesla plans to independently create all the important parts of the technology and implement it in its own cars, so that it also owns the data. But some OEMs claimed that if the machine belongs to them, and their client, then the data also belongs to them, and not to some other technological partners. This is a fairly reasonable position that should be considered in connection with the manufacturers of sensors: I’m not sure that it will be possible to sell GPUs, cameras or lidars by themselves, and not to share available data with anyone. But the company manufacturing the robomobiles needs to have the data - without them, nothing will work. If data changes do not affect the technology in a loop, the technology will not improve. This means that OEMs add network value to the supplier, and they themselves have nothing from this value, except for some improvement in autonomy - but this improved autonomy itself becomes a commodity among all products of any OEM using it. This is similar to the position of a PC or Android: they create a network effect by agreeing to use software in their products, which is why they sell their products, but their product has become a commodity and the network value goes to the tech company. This is a vicious circle in which value goes largely to the seller, not the OEM. Therefore, most OEMs now want to make robomobiles on their own - they do not want to finish just like they create a network effect by agreeing to use software in their products, which is why they sell their products, but their product has become a commodity, and the network value goes to technology companies. This is a vicious circle in which value goes largely to the seller, not the OEM. Therefore, most OEMs now want to make robomobiles on their own - they do not want to finish just like they create a network effect by agreeing to use software in their products, which is why they sell their products, but their product has become a commodity, and the network value goes to technology companies. This is a vicious circle in which value goes largely to the seller, not the OEM. Therefore, most OEMs now want to make robomobiles on their own - they do not want to finish just likeCompaq .

And this leads me to the final question: how much data can you need? Can the system improve endlessly with the constant addition of data, or will we observe an S-shaped curve - will there be a certain point after which adding new data will have very little effect on improvements?

That is, how strong is the network effect?

Pretty obvious question for cards. To make the cards good enough, what density of cars and what frequency of trips do you need, and what is the smallest market share that results? How many participants can the market accommodate? Could there be a dozen companies, or just two? Can a bunch of second-rate OEMs get together and merge all their data into one database? Will trucks be able to sell their data the way they sell different cartographic information today? The situation is different from consumer ecosystems of software - RIM and Nokia could not merge Blackberry and S60 user bases together, but the cards can be combined. Is it a barrier to entry or a condition for entering the market?

This question applies to the data needed for travel, and in general to all projects using MO: at what point do the improvements become insignificant when adding new data, at what point does this curve begin to straighten, and how many people do you need to get so much data? Suppose, for general-purpose search engines, the improvement effect seems endless - the answer can almost always be more relevant. But for autonomy, it seems necessary to have a ceiling - if a car can drive around Naples for a year and not get confused, what else can be improved there? At some point, you will essentially complete the improvements. The network effect means that the product improves with an increase in the number of users - but how many users will it take for the product to stop improving significantly? How many cars need to be sold, so that your autonomy is closer to the best on the market? How many companies can achieve this? And, meanwhile, the MO itself is changing rapidly - it cannot be ruled out that the amount of necessary data needed to achieve autonomy will decrease dramatically.

All this hides the assumption of the very existence of concepts such as better or worse autonomy. What does “worst” autonomy mean? Slightly more likely to die in an accident, or a little more likely that the car will get confused, slow down at the side of the road and connect to the support center so that the operator takes control? Will manual control levers packed in polyethylene jump out of the console and will the car give encouraging comments?
“Computer, what kind of maneuver can we take?”
“I'm afraid no guys, the ship control system is paralyzed by silencers,” the computer cheerfully explained. - I'm starting the countdown. Forty-five seconds to hit. Just call me Eddie if you feel better.
Zaphod tried to decisively escape in several directions at once.
- So! He cried out. - Uh ... You need to take control.
“Do you know how to control a ship?” Ford asked politely.
- Not. And you?
- Also no.
- Trillian?
- I do not know how.
“Wonderful,” said Zaphod with apparent relief. - We will manage together.
“And I do not know how,” Arthur decided to remind himself.
“No wonder,” Zaphod caustically remarked. - A computer! Turn on the manual control system!
“Please,” said the computer.
The panels leaned back in the wall, and a huge, incredibly intricate remote control appeared from the bowels of the ship, its handles, buttons, levers and switches were tightened with cellophane. They have clearly never been used.
Zaphod's eyes ran wildly.
“Good, Ford,” he firmly ordered, “full back and ten degrees to the right!” Or something else ...
- Good luck, guys! - Cheerfully inserted the computer. “Before thirty seconds.”
Ford jumped to the remote control and grabbed the first leverage. With a terrible roar, the ship reared up and twitched convulsively as the shunting engines pulled it in different directions. Ford let go of half the leverage: the ship made a sharp turn and rushed towards the rockets.

The answer, apparently, is that the fifth level of autonomy will appear in the form of evolution of the fourth level - each machine will have manual controls, but they will be used less and less, and the fifth level will appear step by step, and the controls will decrease, then hide, and then completely disappear - atrophy. Perhaps first the 5th level will appear in Germany, then in Naples, then in Moscow [ dirty insinuations - approx. perev. ]. This will mean that data is collected on a network scale and used long before full autonomy.


[ “Perestroika in Moscow,” commented the author of the tweet. It is unlikely that this egregious case can be called the norm. For comparison, it is enough to google “usa road wars” - approx. perev. ] The

answers to these questions are not yet known to us. Few experts in this field expect the emergence of fifth-level autonomy over five years, most of them leaning toward ten years. However, they point to a wide range of results that can lead to very different options for influencing the auto industry.

One extreme is that the network effect will be weak, with the result that 5-10 companies with a more or less autonomous platform will appear on the market. In this case, the automaker will buy autonomy as a component, at a price similar to ABS, airbags or satellite navigation. The industry will change anyway - autonomy will lead to a drop in the cost of trips on demand by at least three quarters of the current price, as a result of which many people will think about the need to own their own car. At the same time, the transition to electric cars will reduce the number of moving parts in the car by 5-10 times, which will radically change the engineering dynamics, the base of suppliers and barriers to entering the market. But the situation will not reach the Android level.

The other extreme is that only Waymo will be able to create a robot car, in which case the industry will look completely different.

Also popular now: