Why a classic automatic car is impossible and has no commercial prospects

Why a classic automatic car is impossible and has no commercial prospects

Those who at least follow the automotive innovations, well know that a full-fledged automatic car should go on the roads for a long time. But its appearance every 5 years is constantly postponed for another 5 years. Obviously, something is going wrong, and even the most obstinate optimists are beginning to be tormented by vague doubts, and those who remember the "winter of artificial intelligence," so generally become disheartened. Artificial intelligence has rested against the barrier of understanding . It is obvious that if Google (Weimo) with its unlimited possibilities for 10 years didn’t do anything, in fact, apart from expensive toys, then it’s time to admit that with the current artificial intelligence technologies a fully-fledged commercial automatic car (AA) of the fifth level impossible to do. Indirectly, it was recognized even in Veimo herself, and in the last tests in December 2018 (Waymo launched a paid unmanned taxi service in the USA) returned the driver to the safety behind the steering wheel (In each car, just in case, the drivers would be put behind the wheel to react -major situation or failure of the unmanned system). And the articles that in 2019 a full-fledged AA has become further from us, and not closer, no longer gain a bunch of minuses (the 5th level gets further, not closer) .

In this article we will examine the insurmountable obstacles that have arisen during the numerous and lengthy attempts to create a classic AA. That is, we will talk about the real state of affairs today, and not about the advertising promises of a brighter future in N years.

What is an Automatic Car (AA) in general?

In the view of the man in the street, this is a car driven by a computer, that is, a kind of Artificial Intelligence (AI). From here there is an obvious solution: we make a normal AI, connect it to one pivot stereo camera, put it on any car, send it to regular driving training courses and get a qualified driver. However, we see nothing close to this decision. That is, the existing AI can not compete with the driver person. Therefore, the developers of AA are forced to go on deception, offering a kind of substitute (a bunch of expensive sensors, neural networks that do not understand what they are doing and so on). Please note that all this is a temporary solution, until a full-fledged (strong) AI. With the advent of which, all these expensive radars and lidars will be thrown out as useless (a person does without them). That is, the very classical approach to creating AA smacks of some kind of deception. Let's figure out why.
')
Let's start with what we define, and what do we mean by full-grade AA of the fifth level? By this is meant a certain automated complex (ACS) capable of driving a car in real road conditions with the characteristics of an average human driver. A problem immediately arises - what kind of animal is this “average driver”? Is this the average temperature in the hospital? But I propose to leave the solution of this issue to the PR people, and greatly simplifying the task, compare AA to the level of a novice driver who only received a driver's license and, accordingly, admission to public roads.

That is, we will compare the novice driver and the most advanced autopilot, not even a specific one, but the maximum potential at the current level of technology development. To do this, it is necessary to build a comparative table of all the characteristics of the vehicle control by a novice driver-person and the most advanced autopilot.



And what do we see? That, according to one criterion (responsibility), AA in principle cannot reach the level even of a novice driver. That is, strictly speaking, classic AA can never reach the level of a novice driver even theoretically.

Or maybe her, this responsibility? Let's figure it out. We see that even the potential autopilot characteristics are “good,” but not “excellent.” Because everywhere there are insurmountable restrictions. Short:

The accuracy of the control depends on the pavement information. And in general, the completeness of this information is not.

The road assessment is generally based on incomplete information and, accordingly, cannot be accurate.

Decision making, with a generally inevitable lack of information, is probabilistic in nature and cannot be guaranteed to be correct even theoretically.

As a result, it is obvious that even the ideal autopilot cannot guarantee the absence of accidents even theoretically.

Then who will be responsible for the inevitable fatalities in an accident?

Therefore, when the next PR people will say that a full-fledged AA is already fully prepared, but do not believe the problems with archaic laws. If manufacturers would take responsibility at the level of a human driver, the laws would be changed instantly . But the reality is that there is no real AA yet, and there are already dead bodies from its “fakes”, and no one wants to answer for them. Mask said that he had nothing to do with it, but this sucker himself was to blame for not understanding that Tesla's “autopilot” is not an autopilot, but “TM autopilot”, that is, just an advertisement lies the name of an ordinary assistant driver. And it was necessary to believe not Ilona Mask and advertising, but what is written in the end in small print ( all responsibility lies with the driver-person ).

Thus, the problem of responsibility of classical AA cannot be solved even theoretically. So, a full-fledged classic AA is impossible even theoretically .

This, by the way, does not mean that it cannot be made or even sold, it means that first , you have to somehow solve the problem of responsibility, or push the way of how to otmatsya from this responsibility.

And here we understand that it is very difficult to make it in a natural way (without direct directives of the state).

Moreover, the problem is even worse. Responsibility is a consequence of establishing guilt. But this is a big problem now. On the one hand, the presence of a black box on AA simplifies this task, and on the other hand complicates it, since all information from the AA manufacturer is information of the interested party. There is no need to go far for examples: naturally, the transition of the car to the “dieselgate” mode was not reflected in the logs. Or maybe even worse, knowingly false information can be written to the logs. Here is the report page from the gas sticking investigation:



Slide 14 of the report - the black box can lie about what the driver was pressing.

And then, AA manufacturers are companies of specific countries that travel countries will support, last but not least paying attention to the truth. For example, in very similar cases of Toyota and GM, the US authorities make significantly different decisions.



Let me briefly remind you that there was a technical flaw there, they knew about it, they hid it for a long time, and those who talked about it were fired.

As this problem will be solved, PR managers do not tell us.

But suppose we still found AA guilty. What to do next? It is necessary to find out the reason, and here we understand that the current pseudo-AI, that is, neural networks, in principle, cannot explain why an erroneous decision was made, which led to an accident.

A small digression about neural networks. Their problems are well known to specialists. “Reviving AI” is no more than expensive iron and advertising, thrown on the implementation of the old idea . And many are already beginning to speak openly about the second winter of AI. If you explain on the fingers, then you need to know about neural networks one thing: that it’s just multidimensional statistics. Not less, but not more. From here their possibilities and restrictions are obvious:

1. Like any statistics, they are very dependent on the input data, that is, even small changes in the input data can seriously change the final result.
2. Since this is an empiric, not an analyst, and multidimensional, the developers, in principle, do not understand how the neural network works and which elements affect what. This can be illustrated on any complex empirical formula, which is a polynomial with multiple coefficients. It is impossible to understand for what and how the specific coefficient is responsible, otherwise it would be possible to write a simple analytical formula.
3. From the first two points, it follows that any change in input data requires a full test of the entire system. Otherwise, it turns out that correcting one error, we can add other errors that weren't there before. That is, the results of past tests lose their meaning.
4. Adequate collection of primary data is a task that is organizationally and economically very heavy.

As a result, the classic error correction approach does not work. And as a result, each new software version is not a fact that it will be better than the previous one. How then to certify AA?

But we all consider the situation on the assumption that, from a technical point of view, a perfect AA already exists. But this is far from the case. Here are a few reasons, each of which, if it does not put an end to the classical approach, then at a minimum, postpones the decision and increases the cost of the final AA . And he is still very expensive.

Impossible to see everything from the car

The problem in the classical way has no solution at all, since there are simply no such sensors for any money. Therefore, in the general case, the AA must make decisions in the context of a lack of information, that is, think about what is around the corner. The person is responsible for these his own ideas, and the computer? Please note that depending on real responsibility, the same person with the same experience and qualifications makes different decisions on a real road and on a car simulator in different ways . And if the driver has to be responsible for a large number of passengers, then he is required to receive an additional category of driver's license. Therefore, responsibility is a fundamental management criterion on which other management characteristics directly depend . A classic AA is completely devoid of this most important criterion.

It is impossible to make decisions without a full-fledged even non-AI, but a full-fledged artificial consciousness

And in dangerous situations it is often required. For example, following the written rules (we brake directly) and unwritten (we leave to our side) would lead to an accident. And a detour on the left with the exit to the oncoming traffic allowed to avoid an accident. It is obvious that a problem without a full-fledged artificial consciousness does not have a solution, since it requires making a decision about a deliberate violation of traffic rules in the conditions of a lack of information. Who will be responsible for the consequences, in the conditions of a real lack of information and probabilistic forecasts?

There is no correct trajectory.

It is not hard to guess that the issue of resolving conflict situations between AA and a man-driver cannot be unambiguously resolved (a person will always say that he would agree with the other person about interaction). The decision of this issue in court is very problematic, since the judge is unlikely to be objective. Since the judge may be in the driver’s seat, the driver’s computer will never be in the driver’s seat.

It will turn out all skipping lohovozka that does not suit the owner

Real AA can be programmed only for strict adherence to traffic rules , which unfortunately do not differ in severity and unambiguity. Therefore, AA will hypertrophically observe the rule “give a fool a way”. And naturally, fools-drivers will use this, cutting AA with impunity. And he will ottormazhivatsya skipping all forward. How quickly will this annoy the AA owner?

Cannot guarantee security (which will not crash into you)

The task of the trolley, which does not have a safe solution, guarantees the dead through the fault of AA. Naturally, professional drivers who are left without work will use any accidents for political pressure on the authorities in order to ban AA. And these are real voters.

Unacceptable low reliability and noise immunity

Absolutely all AA sensors are easily jammed or cheated. Thus, it is very easy and cheap to disable all AA on any part of the road. We give children "laser katafoty" for a bicycle and ask to ride along the road. We scatter from the quadcopter on the trees wipers wifi, GPS, white noise for radar. We draw the "original" road marking. Slightly modifying road signs, confusing conventional video cameras Tesla can fool a 40-dollar set of Arduino and an ultrasonic transmitter that generates sound waves - the car will not be able to find free parking space or miss real close obstacles . Cheap and efficient. And we do not even consider hackers! These are just Neo-Luddites.

Professional drivers-people are strongly opposed and will interfere

So, a simple and cheap jamming (such a simple EW) can be guaranteed to stop any AA. Who in such conditions will not be able to work even theoretically, but the driver-man can . Try to come up with at least one reason for which people drivers who are out of work will not use this method. But transport, which can stop at any time for an indefinite time, is completely unnecessary. And taxi drivers, people who have raised prices at this moment, will quickly reduce the number of people who want to save on AA-taxi to zero.

Hackers

Even now, hundreds of companies from dozens of countries make AA. How to check, certify and maintain all this? And how do you certify software that is constantly updated over the air? How to ensure that another offended employee did not merge access codes? In the future, all auto pilots will be repaired in the provincial car services. How to control which firmware will be installed there? Who will be responsible for these firmware? The driver will say that I was put in the service and I did not know anything about the activated God mode, I thought that this was the case with everyone. That is, a huge number of security problems, and whatever one may say, classic AA is just a hacker's dream! A successful target for hackers_ how to cheat the AI ​​and make him make dangerous decisions

Subtotal: for classical AA, there are many technical and legal problems that currently have no acceptable solution . Theoretically, of course, everything can be solved or “smoothed out”, but this will require a lot of time and money.

And here we fall to the ground. That is, from “wow! AA "to" but at whose expense the banquet "? In reality, there are many significant additional costs that are not included in the AA price tag:

1. AA himself. It's just expensive here, but at least honestly.
2. Exact high resolution maps. And how to maintain their relevance? There are two ways and both are bad. First, collect the company itself to the manufacturer AA. But how many cars are needed for this? The second way, let the cars collect themselves. But how to determine the accuracy of this data (and not hackers or competitors are trying to teach your neural network to the bad). That is, this data will have to be seriously checked, and all checks require money. But the question of the actual relevance of these cards generally has no solution at all for any money .
3. These exact maps must be uploaded over wireless networks, otherwise they will not be able to be relevant in principle. So it turns out we still need broadband Internet everywhere, and even on federal highways we do not have cellular communications everywhere. At whose expense? Please note that even rare updates Tesla allows you to do only on free WiFi. This is not to say that this is not said at all. But at best, in some places in small print and always without a price.
   
   
4. 4) Insurance against accidents. For some reason, it is believed that insurance can resolve the issue of liability. But this is not the case; insurance only distributes responsibility to a large number of innocent persons. Naturally, Tesla owners will not want to be responsible for the mistakes of the unmanned KAMAZ and vice versa. It is difficult to say how things will go, but at the initial stage insurance will be prohibitive (millions of dollars for one car for tests in California). And then the classic vicious circle (little AA-expensive insurance, expensive insurance-little AA). At the same time, the statement itself that classic AA is safer than a person is not based on any facts, since they cannot be due to the absence of real AA .

And where did the myth about the economic efficiency of AA come from?

At first glance, obviously, we throw out the driver, save on his salary. But the reality is not so simple. After all, we are not just throwing out the driver, we are throwing out the person. And for the only person's truck in the car. But this is a dream of robbers! We suppress the connection, imitate the accident, stop the unmanned truck on an empty road. We overload the cargo, remove the lidar, radar and so on. We set fire and leave. And all this can be classified as not even theft, but hooliganism. Man is not there, it means not robbery.
Lidars we shot play for experiments. And the cargo was taken away by random people. And to counter this scenario is very difficult. Since malefactors will jam communications regularly and in many places, it is physically impossible to send riot police for each case.
Putting means of defense on the car is also not an option, since you will be responsible for the dead children, which the intruders will specifically substitute.

So a person should be in any case. Yes, it may not be a driver, but a forwarder. But the presence of a person inevitably requires a regime of work and rest. That is, the main advantage of AA - around the clock work is lost .

And if because of the hackers, several trucks firmly stood up and completely blocked the road for a long time, for example, an ambulance? That is, in the foreseeable future, every car should have a driver. That is, it’s not even the fact that it will be possible to get along with one driver, having loaded the work of two drivers on him. Yes, of course, it is possible for him to lower his salary, since he is needed only in the wake of the autopilot. But this is very similar to the principle that the cow needs to be milked more, but less fed. And the driver’s salary is simply nowhere to be reduced; there is a huge shortage of drivers in the US for the existing salary.

And there is another problem: if the car most of the time in standard conditions will go on autopilot, then the driver will quickly lose the skills . And how can he steer with such a low qualification in difficult conditions when the autopilot fails? This problem is well known with aviation autopilot. Yes, it is successfully solved by a system of simulators, examinations, and so on. But these costs will also have to be added to the AA estimate!

And there are still problems with potential unscheduled repairs (wheel replacement) and so on.

Intermediate conclusion: the very concept of the economic profitability of classical AA is not as obvious as it is considered to be . Please note that after three years of widespread successful testing of truck traffic by caravans, three years have passed, and the technology has not become commercial, although there are no technical or legal problems with the technology of caravan traffic.

A classic automatic car will increase traffic jams and complicate their regulation:

But let us assume that all of the above problems were somehow miraculously solved, and we have a full-fledged AA of the fifth level of autonomy without a steering wheel and pedals. Wow Wow But who needs it, who will pay for it? Naturally, the process will not occur instantaneously, and someone will surely boil back and forth at these pitfalls, but the logic of introducing classic AA looms like this:

1. The technology is most effective in an automatic taxi.
2. Lack of salary to the driver will dramatically reduce the price of the trip.
3. This will sharply increase the demand from public transport passengers. It is proved by Uber that they are transplanted to cheap taxis not from private cars, but from public transport.
4. Meeting the demand, business and private owners will dramatically increase the number of automatic taxis, according to the Uber model.
5. The movement will rise, since the effect of reducing the distance between cars is significantly less than the effect of a sharp and multiple increase in the number of simultaneously moving cars (now 80% of cars are without a driver, and could go in an automatic taxi mode).
6. Obviously, the classic business of taxis when using AA will not be able to compete with private AA, because the private owner does not need to depreciate the value of the car (he still bought it for trips to the cottage on weekends). Thus, since the number of moving automatic taxis can change quickly and unpredictably, it will be much more difficult to regulate traffic jams (it is very difficult to control private owners, it is proved by Uber).
7. The state will have to respond in the only way available to the state and raise the price of AA ownership (hi Vision Zero). As a result, tada! The price of a taxi ride will return to the current level, which corresponds to the current number of taxi cars, which corresponds to the adequate capabilities of the road network and the real solvency of public transport passengers.
8. As a result, the price of transport services will increase for everyone (laid-off drivers, businesses, taxi passengers, public transport, and especially ordinary car owners). It will increase, since the prices after the jumps rarely drop to the initial level.

So the result is that classic AA is not beneficial to anyone:

1. AA tax business is not profitable, because instead of the driver’s salary, he will have to pay about the same amount as a “tax on AA”. It is the AA tax, because the number of taxis with human drivers is strictly limited to the number of taxi drivers, and this is a sufficient natural restriction. And the number of automatic taxis is limited only by the credit rate, that is, de facto nothing. But the cost of AA itself is significantly more than the same as usual. As a result, from the point of view of business, sheer losses.
2. AA's private trader is not profitable, since he will also have to pay "AA tax". Because the state will be “lazy” to determine who uses AA only for himself and who for taksovaniya.
3. Taxi passengers, this situation is also not profitable. Since the price, at least, will not decrease, but traffic jams and their unpredictability will increase.
4. Passengers of public transport, the increased traffic jams too to anything.

And this is all without taking into account the “nightmare of UAVs”, when in order not to pay for parking, UAVs will simply circle around the center of the city. That is, paid parking disappears. Paid entry to the center, too, since there will simply be a transfer at the border. It remains only the total discrimination of AA , the benefit of the Vision Zero seriously trained.

Pay attention that experts know, in principle, about most of these problems. Forecasts: unmanned vehicles will create chaos on the roads . Philip Road, LSE Cities_ is the most nightmarish scenario - drones driving around the city not to pay for parking . dead end, starting with the seagulls . But no one tries not to talk about what this will inevitably lead to. This is a “professional effect”, because only professionals can tell about these problems, but if they say loudly about the futility of classic AA, then investors will no longer give money to these prospective classic AA technology . And his own shirt is closer to the body.

Conclusion: a full-fledged classic AA has a huge number of security problems, technical, economic, traffic, legal and so on. Most of these problems at the current level of technology development have no solution at all, moreover, even the very existence of these problems is carefully ignored. Thus, the emergence of real commercial AA of the fifth level is postponed .

The question whether a full-fledged classic AA is possible is to answer two questions:

1. Can classic AA guarantee the complete absence of those killed in an accident?
2. If not, then who will be responsible for the victims?

Without a legally meaningful answer to these questions, classic AA will remain an expensive toy . That is, you can do classic AA, but you ca n't do it.

PS This is the third article from the cycle description "ITS SPRUT", the first article here .

This concludes the analysis of the current situation with the classic approach to road safety and the creation of classic AA, and in the next article I will begin to describe in detail how to solve all these problems.