When will autonomous driving cool down, and how long will it take?

--Pay attention to the reply to "information" and receive Tesla patent technology analysis report--

There are some problems with current autonomous driving: the data requirements are too high, the cost of future sensors is difficult to accept, the massive map data is endless, and there are also safety problems that require high requirements for the robustness of the system. These problems all seem to be difficult to solve with current AI technology. Therefore, someone recently asked on Zhihu: "When will automatic driving cool down, and it is estimated that how long will it take?" "It sparked a heated discussion. In this regard, a computer engineer @zhzz discussed in detail the dilemma he believes currently faces in automatic driving with sufficient arguments, and put forward his own views and prospects for the future of this field, which has also been recognized by many people. This article is the engineer's answer!

Can ask this kind of question to show that it is an insider, the current difficulty of automatic driving is mainly in perception and decision planning, pure visual route is basically game over, although in the future, the computing power will be larger and larger, cheaper and cheaper can run more complex neural network models in real time, but these models, or visual means itself is very limited, and, the training cost is also high, affected by the environment, resulting in the basic impossible to reach the commercial level of robustness; and the complexity of the real road, With the near-endless variation of the real world, it is difficult to completely fit it with a limited number of mathematical models (neural network models). A typical scene, traffic lights or traffic markers are relatively simple and limited graphics, but in Hong Kong, Tokyo's narrow streets full of flowers and green billboards are currently difficult to accurately and robustly identify. Of course, it can be helped by some technical means, such as locking the search area and using V2X 5G car networking, but this is no longer a purely visual problem;

In addition, the decision-making planning piece, I personally feel that this belongs to the deep water area, known may only google begin to touch this depth. When driving, people will make a lot of decisions and predictions in real time, a lot of empirical judgment, do some subconscious logical thinking, the current deep learning is good, the traditional scheme is good, there is no way to do human thinking, reasoning, analogy, associative thinking ability, for example, a plastic bag blown up by the wind on the road, or a large number of falling leaves, radar, or visually seems to be an obstacle, may have to stop or do emergency avoidance, but people will recognize this thing, directly drive past; or, At present, it is often encountered that a small puddle, or shadow, vision may be mistaken for an obstacle, of course, this time may be radar tells you that you can pass here, this time, how to write your environmental fusion modeling logic? Is it more trusting of radar, or vision? Is it walking or stopping? Of course, this also involves the problem of perception, you say that I perceive it, I will drive past, this is basically unrealistic, your decision-making planning module is difficult to write an if-else judgment for each special scene, such a situation is almost endless in reality; in addition, see whether the small animals on the side of the road slow down, hear the siren of the police car or ambulance whether to stop and avoid, keep a distance from the muck car, and even the passengers on the car are different, whether it is self-driving to the nearest hospital or the police station, etc. Is it all written as if-else judgments?

With a little understanding of technology, you can immediately realize that the problems mentioned above are hardly visible at present, and there are engineering methods that can be solved well and thoroughly.

Of course, you said that in the future, we can all rely on big data, we can collect a very, very much driving experience data, through the cloud big data to judge, which is equivalent to we have a world's excellent driver experience pool, self-driving cars do not have to understand the reasons for these behaviors, just according to the scene to make the most reasonable reference;

In fact, this is also a major development direction of the current automatic driving, is the vehicle road collaboration, may do automatic driving people are now the world's most eager for 5G, cloud computing early roll out of a group of people, the current pure car intelligence encountered a variety of bottlenecks, rely on the wisdom of the road to break through, the limitations of the car side of the intelligent road to make up for the wisdom of the city, for example, the current existence of surveillance cameras everywhere, and can enjoy the future may appear more variety of layout in the road, the city's sensors, Can help the vehicle in advance and a wider range, more accurate perception of the surrounding environment, through a more real-time network to send information to the car end; on the road vehicles can self-network with each other, tell each other their location and driving status, mutual coordination, the front car for the back car to share experience, such as the possible congestion in front, or avoid the maintenance of the road section, etc.; the data center will coordinate the traffic flow, so that problems such as the current intersection, roundabout and other problems will become very simple;

Of course, this is a very beautiful vision; then you may also realize that to achieve this degree may be much longer than the current most optimistic automatic driving landing time, the investment required is also a day, after all, this means that we almost have to renovate the entire urban transportation infrastructure, change the design and construction ideas of the entire city; in the middle, there may be unimaginable technical and engineering problems, all of which require time and huge investment to solve;

So, you say, why is it so complicated? Can't we let the car be like a little mouse on the street, drilling around, hiding when it sees people, and driving with needles in the seams? I did think about this problem very seriously, because at present, relying on the combination of a variety of radars, using the currently known technical means, it is basically possible to detect almost all the obstacles around the vehicle that may constitute a danger, then, I only use the simplest logic to judge, whether it is leaves, or pedestrians, the vehicles are all brainless to avoid, the use of powerful computing power of the computer, to achieve the world's martial arts for fast and unbreakable, just want to run for life on the street like a small mouse can? Of course, there is also a lot of dynamic modeling involved, not necessarily to avoid all obstacles, for example, to avoid the side of the car, but due to the limitation of mobility may hit the front car or guardrail;

In fact, this design should already exist, used to avoid potential external impacts, and not to actively hit others. But driverless on the road still needs to have a certain interaction with other vehicles, and this interaction must follow the traffic rules, for example, with the car to maintain distance and a certain speed, turn to make a straight, such as can not change lanes at will, traffic lights, roundabouts, intersections through, there are corresponding rules to follow; driverless system developers must go to achieve these rules; and specific to the implementation of rules, the rules to comply with the application, it is back to the above proposed behavioral decision-making problems. Corresponding to the current situation, in fact, the L4 unmanned driving scheme equipped with multiple lidar, millimeter-wave radar and Baidu is basically unheard of for a crash, but it is often very silly, especially when passing through the intersection, often experiencers are more slow to respond, or too cautious; (Tesla's design is very aggressive, without lidar, only a millimeter wave and some cameras, and it seems that the driving strategy design of the decision planning module is also more radical , the feeling is the logic of the little mouse mentioned above, so always bump);

Therefore, to sum up, the perception ability is limited, and it does not have the ability to think and judge in the true sense;

So back to the main question, when is the unmanned driving cool, if you carefully read the dilemma faced by the unmanned driving analyzed above, then it is possible to get such a conclusion:

Driverless driverless, which can really be applied on a large scale and can make people completely free of hands and feet, is likely not to be a good business attempt at this stage.

Or that the general unmanned driving technology (note that the universal unmanned driving that excludes the limited scenario) is not suitable for engineering and commercialization at this stage;

According to past experience, things that can be rolled out on a large scale and commercialized on a large scale will inevitably be able to make it easier to use the existing technology and industrial base, first engineering, and then, through large-scale industrial production, reduce costs, and then widely apply and create profits.

The premise of being able to industrialize production is that the technology itself has completed three preparatory stages: 1. Theoretical breakthrough is a thing that scientists have foreseen theoretically very early and proved its feasibility. 2. Technological breakthrough, this stage is basically equivalent to breaking through the barriers in technical realization in the research institution with a very elite professional team, making demos and samples that meet or approach theoretical expectations; 3. Engineering, mainly to solve product design, program optimization, complete functions, performance improvement, yield rate, robustness, usability improvement, large-scale replication of technical preparation, cost reduction and other engineering problems.

For example, mobile phones, wireless communication-related theories and attempts began about 100 years ago, and then, ternary batteries about the 80s to make the current prototype, low-power chips also basically appeared in the 90s, other radio frequency, networking and other theories and technical reserves have a history of several decades. It took quite a long time from theory to conceptual idea to the finished product.

In fact, the industrialization, engineering of high-tech products, large-scale integrated circuits, OLED, quantum dots, and their corresponding theoretical and technical reserves from the laboratory to the practical usually take 20 years or even longer; after all, the engineers in the enterprise generally work on the shoulders of scientific researchers, first of all, scientists, big researchers give us the direction of the name, pave the way, we do a thing. I think that when engineers in any other field are skillfully using mature methods for product development, only driverless engineers are staring closely at the top of the industry all day long, the so-called latest progress published in well-known journals is working, and I even saw that a large factory recruits driverless related engineers to clearly require that they be familiar with various states of the art research, it is best to issue the top journal and vote for the top meeting, which is a bit ridiculous. It can be seen that a group of students (doctoral students) are doing self-amusement development. It can also be seen how immature this field is at present. Normally, the engineering field is more inclined to mature and stable solutions that have been practiced and tested by products, and these are precisely not present in unmanned driving.

The most critical, people's theories can clearly prove from the very beginning that what kind of performance these products want to achieve is technically achievable;

And general unmanned driving, in fact, in the case that the first and second stages have not yet been completely completed, directly entered the third stage driven by capital;

There is no theoretical proof that the problem i mentioned earlier can be completely solved, which involves answering the current limits of artificial intelligence, machine learning technology (including but not limited to deep neural networks) can reach; at least I do not know of any research that can answer the boundaries of ability. Or more specifically, it can be demonstrated that the level of intelligence required for universal driverless driving is within the capability boundaries of currently known technologies.

Obviously, the unmanned demos made by the scientific research team participating in the DARPA competition at that time, including the demos that have been invested in so many large and small companies after so many years, have not been able to technically verify the feasibility of this. (This refers specifically to the driving ability and behavior ability of humans in a variety of real-world complex scenarios that are required for universal unmanned driving.) Running a park at a low speed, running a warehouse, this limited scene is currently possible for many solutions);

Therefore, in fact, I personally think that it is best for general unmanned driving to stay in universities or research institutions honestly, to explore basic work, when the basic theory, technology, and ability are accumulated to a certain extent, engineering and commercialization are natural things.

People who invest in unmanned driving are actually gambling at present, betting on the theory that is currently lacking, and technological breakthroughs can suddenly emerge in the near future, of course, there is this possibility, but this breakthrough may also be decades late, which is unknown

In my personal opinion, it will not even take 10 years, at most 5 years, if it is still impossible to break through the current bottleneck, the investment will withdraw from this field on a large scale, that is, the day when the landlord said that the driverless cool. In fact, the current trend is already obvious, and new investments have rarely come in. However, this technology itself will not disappear, or will continue to exist in a variety of limited scenarios of the application (the market size is very small, can not afford large-scale investment), at the same time, the second-leading solution, intelligent assisted driving derived from unmanned driving technology will be greatly applied and promoted, used to improve the driver's driving experience and driving safety.

And when one day finally comes, our cities, our roads are becoming smarter, perhaps the expected driverless will come.

Unified to explain some of the controversies and doubts in the comments:

First of all, again to explain the main point of view above, the perception ability is limited, the implementation of driving behavior decisions on the current mainstream of the industry or artificially write various rules (can be based on the state machine, based on various parameters of judgment, or some relatively simple and rough adaptive logic, the above is simply with if-else This is a relatively less rigorous statement, everyone knows what it means), in fact, it is based on some badcases found by the test, Cornercase to use some trick to avoid, to express is the current implementation of these rules or programmers tell the car (computer) how to do, the car itself is not intelligent, does not have the ability to reason, analogy, association, in essence, the implementation of these rules and the realization of a Taobao order, takeaway order business logic is not much different.

You for a specific scenario, you implement what rules the car how to drive (ideal, assuming there are no other bugs), you do not implement or the environment changes slightly, the car does not know what to do, behaves very silly, or there are some dangerous behaviors. This is what the industry usually calls generalization ability, the current situation is that this kind of generalization ability is very weak, or even no generalization ability, generalization is to pile up rules, or, for perception, is to retrain a new model.

Some people mentioned that the videos released by waymo, cruise, etc. look very beautiful, I do not doubt at all that these videos are all true (of course, there are also manufacturers of fake, for example, I clearly know that the video at a certain conference of a domestic factory is spent a lot of money to find people who make movies a small section of a small section and then edited out, there have been startups looking for venture capital flickering, behind the scenes is actually let people take the notebook remote control of the vehicle, this domestic and foreign have, is not a secret).

Industry insiders may be easier to understand, waymo, cruise and other videos look very good because of the scenes in the videos they released, the vast majority of videos, which are relatively ideal test environments, sunny days, open roads, not many cars, pedestrians and vehicles are also very compliant with traffic rules, traffic flow, people flow are relatively stable, need to pass through the intersection, fork the field of vision is also relatively open.

And in fact, most of the current unmanned engineers are trying to cover some of the tests encountered in the badcase, or to do generalization; that is, it is often said that unmanned driving to do 60 points 70 points 80 points is not particularly difficult, github on the open source program of each business module simply change it, take it to spell can spell a fifty or even sixty points can let you simple scenarios, on the road to run up the program, to do 85 points, 90 points difficult to the heavens, If you want to commercialize, you need to achieve 99 points or more.

The above is a case that has been encountered, a passage on both sides of the wall, a fork in the middle of the wall, there are vehicles pedestrians in and out, the vehicle before and after the fork can not see the fork inside, no matter how much radar, how good the sensor, is the blind spot of vision, and then there is a car to come out, fortunately decelerate in advance, the speed is not fast, less than 10km, if the safety driver does not take over may be directly hit. You said that I implement the strategy, see the intersection on the deceleration, it is true, deceleration, deceleration is not enough for me to stop, see clearly and then go, well, this is to the problem mentioned above, there are reports (can be found on the Internet) experience waymo, baidu has a response to the vehicle crossing the intersection, or when there are more cars, driverless performance is very dull, when the car is less, you are sluggish, you can imagine the rush hour, you follow a butt car, there is a sexual emergency, obviously you can go, you lie there and do not move, In fact, there are many road testing experiences have been encountered, when the first time it is not good, the driverless inexplicably comes to a sharp brake, or stops and does not move, you do not intervene and do not go, you can go back to check the log, play the bag, see what the situation is. You say that the above situation people also do not handle well, I can clearly tell you that people are much smarter, the vast majority of cases human drivers are very safe and smooth (pay attention to the word smooth) through, humans can listen to the sound (waymo has now been on the road sound inspection, it seems that the report is to detect sirens and the like, this kind of vague and complex judgment does not know whether there is any), or look at the passage has no headlights to play out, or look at the front of the car passing, you can reason out that there is no car at the fork or the car in the fork in advance to give way, Then follow up with the drive past, according to experience estimates will not suddenly have a US group, hungry small electric donkey out, in short, the old driver can be based on a variety of clues, as well as experience to dynamically make the best decision. It is difficult for you to simulate this human ability through artificial intelligence or some other means.

There is also interaction with other vehicles, pedestrians, there may be a mutual temptation in the middle, such as overtaking, narrow road opposite the wrong car, and finally form a consistent strategy, unmanned driving is very silly, you want to overtake I will let, you want to wrong car I will stop, because this is the simplest to achieve. But in some places that haven't been tested yet, this may also cause some unknowable problems. For example, if the opposite direction of the wrong car, if both are unmanned, both take the initiative to stop each other, the implementation of this simple strategy, may be two cars lying on their stomachs and waiting for each other to pass first, it will be blocked. Of course, there must be a way to design a more complex strategy to avoid this situation. And rush hour, a bunch of people, a bunch of cars crossing the intersection, I just want to watch these videos, these waymo, how come cruise doesn't put it? Of course, at its peak, the crossroads may not allow them to test it.

Not to mention heavy rain, snow, fog, road area water, long tunnels, neon lights all kinds of flashes in the dark streets. Anyway, I have not heard of any engineer looking for excitement (digging a pit for themselves to find overtime) to measure these, but if you do business, cover these are the most basic, because these environments can be opened, I remember that many years ago when I was in school, I asked a Japanese old man to tell us about unmanned driving, the old man gave an example I am still impressed, that is, they are in Hokkaido, Japan, a snow is white in winter, even the trees are white, and the visual knows that a piece of white is the lack of texture, and the current visual means have a high probability of eating up. People are not easy to drive, but they can drive, they can judge the lanes by undulating, green belts and the like, they can judge the lanes in front of them, how to do vision, burn incense and ask the grandfather to appear, let themselves work hard to train the traffic sign model, just right to identify the ruts as lane lines? Of course, you say that I rely on RTK, rely on high-precision maps, rely on lasers to maintain positioning, OK, not to mention that RTK will fail at some times, laser this problem we have also encountered, do not consider the cost, do not consider the number of what is very ideal, we originally took the laser map in the summer, to the autumn is not good, why? The leaves are falling in autumn, the point cloud does not match, you say that you update quickly, ok, the snow situation above, you did not snow the day before yesterday, the map of bare branches, the next snow may be just a few hours, the branches are full of snow, I estimate that there is a considerable probability that you point the cloud is still not worthy. Not worthy of good, I am afraid of matching crooked, snow sliding your wheel speed points may also be wrong, with crooked can put your kalman filter or good, sliding window optimization is also good to pull crooked, with chi-square detection, on the federal filter, with a variety of redundant verification can do most of the pull crooked situation. After all, there is also an IMU that is easy to make, and it is also good to make if there is no occlusion GPS/RTK. So is it possible that it is a situation where you are crookedly located in the ditch? It is very likely that the probability problem, this situation car is very stupid, must be boring head to the ditch, people will not, people can make a very comprehensive judgment according to the environment, this comprehensive judgment ability, it is currently difficult to use the program to achieve.

Finally, the problem of generalization, regardless of perception, positioning, and regulation and control, is the most headache at present. Generalization ability is the deep water area of artificial intelligence, and human intelligence is strong in being able to reason, summarize, analogy, and associate. Perception, regulation of the daily various parameters, what is the purpose of the parameter? It is to adapt to various scenarios, all kinds of badcases, but lack of generalization ability. A common phenomenon is that the laborious bar adjusted a set of parameters, wrote a rule to cover a badcase in the past, the scene changed a little, and it was not good to make it, even, move a parameter of this badcase mixed up, another scene that had been passed was hung up again, like a gopher, holding down this hole, and another hole appeared again.

There is also a big guy who does control origins, saying that in fact, everything is not far from its origin, vehicle control is only a few variables, those states, this sentence is very correct. However, the reality of the scene is ever-changing, for a specific scene you need to adjust these states, and then form an optimal behavior, then how to let the vehicle itself make this adjustment without human intervention is the most difficult. Now the mainstream is still developers to identify these scenarios, and then, programming to teach the car how to do, the car does not have this autonomous intelligence. Or just have a very, very limited ability to generalize. There are mentioned what fuzzy control, intelligent control, adaptive and the like, I do not understand, but also just heard that the comment area has a control professional big guy, can explain how practical these high-end things are in the actual project? How wide are the applications? Anyway, what I've seen myself is still generally PID-based, plus a variety of case-specific rules or a small amount of adaptive logic. In fact, this is what I mean by gaif-else above.

Again, the above discussion of the car end implementation, the article also mentioned earlier, the comment area also has instructions, with the cloud big data can provide some solutions to these problems, and Tesla, mobileye has been doing this exploration, and even the industry has doubted tesla's driverless capabilities in the past two or three years of rapid progress is because they have a huge number of first-hand human drivers actual road conditions driving data, convenient for them to train models or optimization algorithms in the background. However, the specific method of how to do it is not disclosed by the secrets of each factory.

To reiterate my view on the word "cool", it is not that driverless cars disappear and do not develop, but that commercialization is blocked, large-scale capital withdrawals will appear in a foreseeable short period of time, and the industry's driverless investment and research and development boom will cool down.

I only say that I learned the basic situation, do a good waymo, mobileye, how they specifically to achieve, are confidential, I can not understand, from the public information even if they are still trying to break through the bottleneck, the longest time Google almost did 10 years, still did not break through, Google has money, another 10 years can also be raised, but also for the sake of financial reports to look good to split this business out of the financing set up waymo, other rely on venture capital to eat, What about startups that make big news today and pull big projects tomorrow? Looking at my answer, I'm not saying that the technology will disappear, but that capital that pursues short-term returns can't be consumed forever.

There are comments asking for additional explanations on 5G, as well as cloud intelligence, here are some additional points:

5G standard is very large, the standard about the industrial Internet of Things has a local concept, such as 50 meters within 100 meters of the vehicle local networking, the latency in this LAN is very small, as if you have your own WIFI, if the iPad even your computer needs to go through the server side of a certain application manufacturer's data center and then come back is very slow, but the LAN only through the WIFI router directly access each other is very fast. This can solve the need for rapid response to the workshop interaction problem, the data center there is a large delay, not to mention the data center processing delay, that is, from the base station through a number of routers, through the various network segments to the data center, and then back to the car end, the delay of this section to my rough understanding of 5G and now will not be much different. The cloud solves a wide range of macro problems with low timeliness requirements, as well as collects massive data for some form of post-processing analysis and optimization. For example, the communication of cars in the local area network communicates with each other's location, speed, their own local driving path, the state of the entire traffic flow dispatched by the cloud, the observation information of road sensor nodes in a relatively large area, and the better driving strategy trained through massive data analysis.

There is a problem that must be made clear: whether it is 5G or cloud, it is not to replace the car-side intelligence, but to simplify the design of the car-side system, reduce the demand for various indicators on the vehicle-side intelligent system, and help break through the bottleneck problem that the current car-side intelligence cannot break through.

If you know through the network whether the state of other vehicles is equivalent to more than a very reliable observation data source, and in fact other vehicles can also observe and process the surrounding environment, the obstacles detected by other vehicles around you, the road conditions sent to you in real time, is not equivalent to a single car's perception ability to increase exponentially, a lot is still within your blind spot, such as the front car obscuring your line of sight, or, often encountered, some missed detection, false detection, But other vehicles can be better observed and more accurately checked at the observation angle where it is located, counting other sensors on the road, such as surveillance cameras, roadside speedometers, etc., for the perception of whether to find that sudden life has become much better. If the vehicle information is exchanged, overtaking, crossing the intersection, the wrong car, sending a request to the surrounding vehicles in advance, and then through unified strategic coordination, can it greatly simplify the design logic of planning control and improve safety and efficiency, is it safer than the current guess to predict the behavior of other vehicles? The end of the car you to identify yourself, and then design logic to avoid damaged road that needs to be repaired, or you have a sudden traffic accident somewhere before, however, the vehicle in front of you just blocks your line of sight, it suddenly brakes sharply, you have to follow the reaction, first do not discuss how to achieve the end of the car, how good can be achieved, this problem side of the car is not difficult, and unreliable, if your previous vehicle to share this information for you, you only need the simplest evasion logic is not, The sudden behavior of the vehicle in front of you, such as the sudden appearance of a dog on the road, it has to brake urgently or slam the steering wheel to avoid, through the data link to share with you, you can even react before it has an observable change in the state of motion, after all, the data transmission speed is much faster than the vehicle braking speed. Overtaking, crossing intersections, and wrong cars in advance to inform the surrounding vehicles for confirmation, is it much simpler and more reliable than you designing an incomparably complex behavior logic or artificial intelligence model? Coupled with the blessing of cloud capabilities, can the capabilities of the entire system be greatly improved?

Of course, as I said in the paragraph above, the transformation of the entire infrastructure is huge and protracted, but it is undeniable that some things will be done, and after it is done, it can achieve certain results, and I personally feel that looking forward to the upgrading of infrastructure, and then, to promote the improvement of the overall transportation system capacity, far more reliable than in the technical path that has not been able to squeeze out any oil and water.

And the communication end of the transformation cost after the dilution of really not much money, 5G base station is always to be set up at least the domestic has been included in the planning, this does not need unmanned people to worry about, spend money. Do not give unmanned driving is also to be framed, this is the communication network transformation of the money, the amount of day, specific to the car end, the cost is very small, equivalent to adding a 5G Network card only, this communication module is 5G mobile phones have something, the future will definitely be integrated into the car computer, after mass production will be very cheap. Compared with any sensor now, it is simply negligible. Of course, considering the interaction problem, many of the current design ideas must be changed, and then, because of these improvements, the reduction in research and development costs brought about by the reduction of the intelligent complexity of the car end is also very significant.

Mainly to refute the comments that say I am technically pessimistic and frustrating!

Polytechnic Dick Silk, from the small language is not good, maybe my writing, or the way of expression let the reader have this idea.

However, it should not be difficult for the reader to conclude that my main views and arguments are based on currently known facts and objective analyses. My personal attitude is very objective, even optimistic (for limited scenario driverless, intelligent assisted driving, and all kinds of new changes that infrastructure upgrades may bring).

For example, you can't say that I don't think I'm going to have a sudden genetic mutation and grow a pair of wings that can fly tomorrow, which is pessimistic frustration. Even with this mutation, and I haven't died suddenly because of drastic genetic changes, at the rate of cell division, the new substances I take up every day, and the transformation ratio, I'm unlikely to grow well tomorrow. This is based on basic facts and basic laws.

In the same way, you cannot say that Sun Yat-sen said before his death: The revolution has not yet succeeded, and comrades still need to work hard; Lao Mao wrote "On Protracted War" as pessimism and frustration. We know that this is a lifelong revolution, a comprehensive, rational, and profound understanding of the situation, and a thorough understanding of the conclusions that can only be drawn. The idea of expecting a quick revolution is very dangerous, and most of them have been given away in vain.

In fact, according to previous analyses, there is basically a consensus among the professionals among a reader that breaking through the current bottleneck requires the emergence of new technologies. There is not much oil and water to squeeze out of the existing path. Specifically, the best breakthrough is deeper intelligence, more human-like intelligence.

Usually carry out scientific exploration, engineering research and development, first observe the phenomenon, summarize the law, and then use the law to solve the problem. A typical example of what is now an artificial neural network is inspired by the study of the structure and basic working principles of the nervous system in the field of neuroscience.

Then you now need to engage in high-end intelligence, reasoning, association, analogy, etc., more specifically you have to program the implementation of the computer based on the current Von Neumann architecture, simulating these capabilities (not to mention that it may not be possible). You must at least know a rough idea of what is going on, and basically understand its formation mechanism and how it works, right?

However, it is a fact that cognitive science, neuroscience, biology, medical physiology, or even psychology that is less involved in basic scientific research. At present, the research on these problems is basically still stuck in some very superficial phenomena, and there is little understanding of the deep-seated mechanisms.

In fact, on such issues, any basic discovery will be regarded as an extremely important breakthrough by the entire scientific community, and it will certainly be put on the news network. For example, the research on the division of labor between the left and right brains has been awarded the Nobel Prize. Not to mention that you can make or simulate such high-end intelligence as living organisms, even if you take a small step forward in human understanding of such problems, the Explosives Award or the equivalent scientific recognition must be given to you.

Therefore, I think that even if there is any breakthrough, it must be the first breakthrough of the basic disciplines such as cognition, physiology, and neurology. I never understood why there was always a whole bunch of EECS coders who could have such an arrogant sense of superiority that they could be above all other disciplines. Or are you too ignorant or narrow-minded?

Back to the problem of artificial intelligence, even if you can't come up with more awesome technology, a more hanging way to simulate high-end intelligence. If you can answer the question I mentioned above, where is the current deep learning, reinforcement learning or ability boundary? Or the scope of the problem continues to narrow, at present, do you say that the L4 you want to achieve is within the capability boundaries of the currently known technology, or further, what breakthroughs does L4 need to do to be complete, without you having to give specific solutions, just pointing out the direction?

This problem is of great value both theoretically and engineering. If you can answer it, don't say more, give an academician, or no worse, the titles of various top outstanding talents will give you one. You can weigh the above questions which are you hoping to solve in the foreseeable future, if you really know clearly and have this confident genius, why are you still nestled in a small startup 996 moving bricks?

The above question about the boundary of ability can even be raised at my level, the blood is small and young, the big fools of the old and the young like to worship the gods in the industry, the grandfather, the grandmother will not think? Will you not see the value of these kinds of questions? It's just that the level of people is enough to see the depth and complexity of this kind of problem, and it is most likely that it is deliberately avoided by the chicken thief and does not touch this nail.

So, a lot of enthusiastic young optimism is nothing to understand blind optimism?

As for the blind optimism of many old, middle-aged and young people, it may be more complicated. Maybe some of them really don't understand, they haven't thought it through, and some of them are just fooling around and making profits, drawing big cakes for ignorant young people and beating chicken blood. For example, I was once a leader, a big expert, although I am not from a relevant profession, nor have I done work in related fields. At least the performance is extremely optimistic about the unmanned landing, I am very uncertain whether he is so optimistic, whether it is really ignorant, or purely to find the head of the project, funding, personnel to fool around, and give us the following small soldiers chicken blood. Because even from the first contact, I strongly felt that this person was an incomparable chicken thief, and had considerable technical understanding. After all, this society can mix well, and climbing high has its superiority to some extent.

Those friends who are still optimistic about the prospects of artificial intelligence and the short- and medium-term landing prospects of L4, I personally have a very simple point here. I stared at the annual Nobel Prize, the Turing Prize, the academician co-optation. If you can really produce something revolutionary, these are indispensable to you. I look forward to seeing you on the news feed and the headlines.

Oh yes, I would also like to insert a paragraph about the "experts" in this industry:

Don't be surprised, just think about it a little, 10 years ago, there were few people in the world engaged in things, Google first began to engage in industrial landing, full of calculations is only 10 years, how can so many practitioners suddenly emerge in just a few short years, even big experts? There are so many experts, is it so easy to mix? In fact, many of them have been transferred from other fields, some are related fields, and some are not even related fields. Oh there is also an interesting phenomenon, the more mature the field, you see can be called the industry recognized experts less to say that the field is deeper than 20 years up, and then not 10 years is also there, people are real experts, under normal circumstances, there is no engineering problem in their own field. You see the driverless, especially those who have only emerged in recent years from the startup experts, experts who have changed careers halfway, all kinds of young doctors who do not know what meetings have been watered by several articles are experts, even pay attention to this problem, give answers and comments, it seems that there are mouths to land L4, industry leaders who study L5, industry experts. The whole industry does not even have a reliable commercial solution, and the result is all experts, what kind of expert is this?

There are also many people who ask legal questions, how to solve ethical problems. How to deal with disputes in the event of an accident, etc., and whether there is any progress in this regard.

First of all, I don't know much about this aspect, and I personally think that the current unmanned driving should be far from progressing to this point. After all, nothing that can be commercially used now has been made, and it is still in the research and development stage, and it seems that it is far from the final stage of research and development. Law and ethics are all things that will be done later, and it is necessary to consider the industrialization and mass production of products and commercial promotion.

But I personally think that these problems are actually easy to say, as long as things are made and profitable, there will naturally be capital scrambling to promote these things (especially those who invest huge R&D funds in the early stages, and they can be harvested immediately, who is not excited?). ）。

Of course, there will definitely be a game of new and old capital, it will not be smooth sailing, just like now the domestic high-speed rail, infrastructure, Internet, 5G leading, can go to the world to cut leeks, domestic capital and political forces will promote this matter. Yankee, the Europeans see that you are going to make his money, cut its leeks, and will use all kinds of ethical, legal, political issues to do things with you. For example, the recently hyped up relatively hot ban on Huawei, forcing ByteDance to sell Tiktoc to American companies and so on. For example, in the future, driverless cars will be first made by Internet companies such as Google, then traditional car manufacturers will definitely attack driverless cars with the problems you said, until they make compromises on interests, such as letting them joint venture shares, engaging in technology transfer, or other ways to make everyone get a share of meat.

Even if you are an advanced productive force and represent the future direction of development, you cannot beat to death all those who are lagging behind you, otherwise as a fragile new affair, the possibility of being killed is greater. The normal process of social development is tortuously advancing in the game struggle between the old and new forces.

In short, the principle is that no one can take all the benefits, otherwise the whole world will come against you, which is also the eternal law of human society. Ethics or laws are merely tools for reconciling conflicts of interest. In my personal view, the biggest conflict of interest comes from the game of capital, specific to the end user, but not the main contradiction. The capital forces of the victorious side will certainly fix this problem by fixing a rule in their favor, and through a powerful propaganda machine to gain the widespread agreement of the whole society with this rule, or certain ethical ideas, as for the few with iron heads, it is never enough to create a substantial obstacle.

Many people say that legal problems, in fact, careful analysis, there is no so-called legal problems.

As long as it can be made, legal issues, claims and so on are well resolved.

In the first two years of the market, in order to dispel the public's doubts, the depot, or the driverless operator, will inevitably subsidize it substantially, package it in a big package, and take the accident claim. In fact, this does not cost too much money, because, if it can be commercialized, then at that time, the safety of unmanned driving must be far better than that of normal human drivers, and even if you touch porcelain, it is not easy to touch.

The analysis is as follows, the driverless will be a variety of cameras, sensors, equivalent to 360-degree driving recorder, including, the driving process will be the sensor data, and program processing process logs will be recorded. Then there is an accident, who is responsible for taking the data or log and analyzing it.

Most likely, pedestrians or human drivers are fully responsible, because if there is no error or failure in the program or sensor, driverless people will never violate traffic rules, because the program is written according to traffic rules.

If the analysis finds that it is a problem caused by equipment failure or program bugs, then the responsibility of the depot, or the driverless operator, this responsibility determination is very clear and easy, and the accident is equivalent to helping the depot or operator to detect a bug, and they are not wronged by losing money. Because, if it is not repaired, it may appear in large quantities, then they lose more.

This is equivalent to you buying a mobile phone, TV, the manufacturer gives you the warranty regulations, the warranty period is not abnormal use of the problem, must be responsible for the manufacturer to maintain. Manufacturers in order to lose less money must strive to put the quality, here is the safety to do high, yield rate to improve.

There is also a situation, that is, the general car factory sells the car to the user, there will be some use specifications, including the car itself will be designed to detect the fault early warning. For example, drive tens of thousands of kilometers you have to go to the maintenance, the sensor is not good, the self-inspection system to give an alarm, this car you do not use to repair, this time, if the user forced to use the problem, according to the exemption agreement, the car factory is not responsible, the user's full responsibility is also very clear. And this is also easy to analyze from the data recorded on the car, and the responsibility determination will be easier.

Finally, if the driverless accident rate is lower than that of human drivers, then insurance companies will be very willing to make this money, and it is conceivable that this premium will be lower than the current rate. It can be imagined that in order to push their products, the car factory even buys a driverless car and sends you two years and three years of insurance this promotion, but the wool is out of the sheep, this cost can be added to the price and transferred to the user.

Automakers can reduce their risk through regular maintenance agreements, insurance, and maybe even make money in this way, and they must have the motivation to do it.

The above analysis is for L4 and above level driverless, in fact, it is also applicable to L3 and below. Because the applicable conditions for L3 and below are more stringent, the corresponding user use agreement will also be more stringent. For example, full attention, the hand can not leave the steering wheel and other ghost things, then your user violated, sorry, even if it is the problem of the vehicle itself, my manufacturer does not recognize the account. I couldn't win the lawsuit, after all, my user agreement at that time was clear. This is also the reason why L3 and below automatic driving is not practical, unscrupulous manufacturers take these things to blow themselves how high-tech, there is no clear risk, in order to sell their own cars, it is simply shameless.

Therefore, if you look at the analysis, there is actually no so-called legal problem at all.

Reprinted from the @zhzz, the views in the text are only for sharing and exchange, do not represent the position of this public account, such as copyright and other issues, please inform, we will deal with it in a timely manner.

-- END --