Although "data-driven" has long become the basic consensus of the autonomous driving industry, in fact, until now, data-driven deep learning algorithms are still mainly used in the perception module, but rarely used in regulation
Before getting into the main topic of the article, we need to briefly explain what the “regulation” mentioned in this article is.
I believe that everyone in the autonomous driving industry must know the major modules of perception, positioning, decision-making, planning, and control. Although everyone makes such distinctions in daily communication, it would be a bit confusing if they are still classified in this way in specific work. Problem.
The problem here lies in decision-making. Broadly speaking, decision-making includes prediction, and the division of prediction is very awkward - prediction itself is highly related to perception, but prediction is the pre-stage of decision-making. Only prediction first Decisions can then be made.
We can call all the established facts that have happened T0, and perception is the established fact that has already happened. For example, how the vehicle should move in the future and what others may do are all things that will happen in the future. We can call it T0-.
There is an advantage to using T0 as the watershed division: the perception link only needs to care about the established facts that have happened on the road, and any inferences based on this established facts can be directly put into the decision-making part.
Such classification can more closely combine the prediction results with the decision-making part, thereby artificially "weakening" the relationship between prediction and perception, and "strengthening" the relationship between prediction and decision-making. Reflected in actual work, this division can better determine what should be predicted and how to better use the prediction results to serve decision-making. (The above content is summarized by the author after communicating with Dr. Yang Wenli, CEO of Leadjun Technology)
Based on the above logic, there is usually a PNC group within an autonomous driving company. The main task of the PNC group is to be responsible for regulation and control. Work. After everyone has worked in the PNC group for a long time, they will collectively refer to prediction, decision-making, planning, and control as "regulation and control."
Then this "regulation and control" is what is called "regulation and control" in this article.
Next, let’s combine “data-driven” and “regulation”.
Although "data-driven" has long become a basic consensus in the autonomous driving industry, in fact, until now, data-driven deep learning algorithms are still mainly used in the perception module, but are rarely used in regulation and control. ——Currently, regulation-related algorithms in the industry are mainly rule-driven.
Although currently published information on autonomous driving control algorithms is mostly for rule-based algorithms, when I open the recruitment software, I often see that when autonomous driving companies recruit control algorithm engineers, they require candidates to understand Deep learning algorithms, which to some extent also shows that many companies are currently "marching" towards hybrid decision-making algorithms.
The picture comes from the recruitment requirements of an autonomous driving company
There are currently few articles that fully and carefully explain the use of deep learning algorithms in autonomous driving control application. Therefore, this article will discuss the limitations of rules and the potential advantages of deep learning algorithms, why deep learning algorithms cannot yet be applied to large-scale control algorithms at this stage, and whether future control algorithms will be based on rules or deep learning. Let’s elaborate on three aspects.
1. Rule-based algorithms have limitations
Currently, regulation-related The algorithm is mainly based on rules, but with the improvement of autonomous driving levels and the continuous expansion of application scenarios, the limitations of the rule-based control algorithm will be exposed in actual tests.
The shortcomings of rule-based regulation and control algorithms are mainly caused by the three categories of corner cases: non-deterministic scenarios, strong interactive scenarios, and strongly subjective scenarios:
(1) Non-deterministic scenarios
Non-deterministic scenarios are relative to deterministic scenarios. The simplest deterministic scenario is to stop at a red light and go when the green light is on, and you can change lanes on the white dotted line. It is not possible to change lanes on the solid white line. Scenarios other than this can be regarded as "non-deterministic scenarios".
For example, in most cases, cars drive on roads with lane markings. However, in one case, the current road consists of three trunk roads plus one auxiliary road, and there are only lane lines and no green belts between the auxiliary roads and the trunk road. This is This means that vehicles on the auxiliary road may merge into the main road under any circumstances. This is a typical non-deterministic scenario. Because most drivers don't care about solid and dotted lines at all, you don't know when a car will suddenly appear around you and how likely it is to change lanes directly.
Because the trajectories of other vehicles on the road are probabilistic, rule-based algorithms are difficult to process.
(2) Strong interaction scene
The most typical scene is the intersection. The routes of multiple objects in a scene will interact and influence each other, so interaction There is a certain amount of game involved.
In a very complex road scene, constantly predicting the surrounding cars and then adjusting the plan based on the interaction results cannot be clearly described by human rules.
(3) Strong subjective scene
Simply put, under different road conditions, people will unconsciously adjust the distance between their own car and the car in front according to the environment. Because there may be a set of rules driving the human brain, but these rules are very abstract and cannot be specified. They can only understand the ineffable, such as the timing of changing lanes, reactions to other vehicles at intersections, etc.
In some scenarios, although rule algorithms can be used to imitate some behaviors and driving styles of car owners, people's subjective habits are difficult to solve with rules, and even if they can be solved, it will require a lot of manpower.
Using rule-based algorithms to deal with corner cases, the most difficult part is the prediction process. Because, in addition to predicting the trajectory of its own vehicle, self-driving vehicles also need to predict the trajectories of "other road users", and there is often more than one "other road user" on the road, and they will interact with each other, so predicting their The trajectory is difficult.
When making predictions, human drivers will make decisions based on "experience" and "intuition" by observing the historical trajectory of the car and the driver's driving style. That is, they can choose their own car based on whether other vehicles are conservative. strategy, however, it is difficult for rule-based algorithms to achieve prediction effects similar to humans.
In addition, when the vehicle encounters some corner cases and needs to brake, it is difficult for the rule-based control system to ensure that the curve is smooth at all times.
For example, if there is a car in front of you that is driving at a very low speed, and the autonomous driving system issues an instruction to "change lanes to the right to overtake", but just halfway through the lane change, suddenly there is a car next to it. The car was overtaking rapidly in the right lane, and the decision-making system immediately canceled the lane change decision and regenerated a smooth curve to allow the vehicle to return to the original lane.
The difficulty here is how to balance elements such as comfort, safety, efficiency, and the achievability of the power model, and ensure that the trajectory planning path generated at this time is smooth?
Still taking the above example as an example, if the path generated at this time cannot be guaranteed to be smooth, then the vehicle will drive very "hard" and become very uncomfortable while driving. If the driver/passenger can see an emergency happening around him at this time, and the system helps him avoid the accident in time, he will think that the system is very reliable. Even if he brakes suddenly, the passenger will think that the automatic driving system "helped me avoid an accident." ACCIDENT".
But the reality is that in most cases, the system has discovered the problem long ago and made a decision to directly "nip it in the bud." But the problem is that most drivers/passengers will not know that a "seemingly ordinary" lane change has just "almost caused an accident", and they may not even notice that there is a car coming from behind. Therefore, if the control system cannot generate a smooth curve at this time, the driver/passenger will not only think that the system "helped me avoid an accident", but will blame the system for being "too rigid" and "swinging around" when changing lanes. Go".
Although this type of problem can also be solved using rule-based algorithms, the solution process is very troublesome.
In addition, although in theory, some long-tail scenarios can also be solved with rule-based algorithms, in actual work, it is difficult for engineers to ensure that these complex codes can solve the problem effectively. At the same time, it does not take up too much computing resources. These codes can be easily run on an offline system, but once in the car, can the on-board chip support such a huge consumption of computing power? Solving these problems greatly tests the coding level and engineering capabilities of engineers.
Moreover, some engineers may have written thousands of lines of code, and they have forgotten what they have written. There are more and more rules and more and more confusing logic. When a new requirement is added at some point in the "future", the system will become difficult to maintain.
2. Deep learning algorithms can improve the anthropomorphism of the system in appropriate scenarios
The above mentioned many limitations of rule algorithms, but in fact the The limitations don't stop there. If we look at all the problems together, we can collectively call them the "low anthropomorphism" of rule algorithms.
The so-called anthropomorphism refers to whether the system looks like an "old driver" during driving. A control algorithm expert from Hao Mo Zhixing said: "The core ability of the anthropomorphic veteran driver is to adapt to changes, and using a rule-based system requires setting a lot of rules to distinguish scenarios, and setting a large number of environmental parameters and cost functions. Constantly searching for the optimal solution is difficult to achieve in real time mathematically."
At present, self-driving vehicles that can be tested on the road are more like a "novice driver" in most cases, while "experienced drivers" The biggest difference between "novice drivers" and "experienced drivers" is that "experienced drivers" have stronger predictive capabilities and are better able to adapt to long-tail scenarios.
Comparing the two, although most "novice drivers" are also very safe, "novice drivers" will be "stupid" when driving on the road, while "experienced drivers" can operate various operations while ensuring safety. It will be smoother and more efficient. This difference is called "parole experience."
Essentially, "novice drivers" rely more on "driving school-based teaching materials" on the road, and their decision-making and planning methods are more like if else (rule algorithms), while "experienced drivers" rely on experience when making decisions And subjective judgment, the way of decision-making and planning is relatively abstract, more like a "deep learning algorithm".
In summary, by introducing deep learning algorithms in the regulation and control process, learning as much as possible from the experience of human drivers, and automatically refining features to find patterns, we can improve the "anthropomorphism" of the regulation and control algorithm.
The data engine mentioned by Tesla at this AI Day will improve certainty by inputting data into neural networks to better solve real-world prediction problems.
The person in charge of Qingzhou Zhihang’s planning and control algorithm believes that prediction (judging the future trajectories of other vehicles) is the most suitable and first step to apply deep learning algorithms on a large scale. He said: "Qingzhou Zhihang's prediction algorithm uses deep learning."
Specifically in terms of scenarios, he believes that those scenarios that are highly dependent on human subjective judgment need to be guided by deep learning algorithms.
"For example, you can use traditional classification AI models (which can be deep learning or traditional machine learning classification models) for decision making, deep learning for trajectory generation to guide path planning, and end-to-end Reinforcement learning gives the current action that should be executed to guide motion planning (planning of the final trajectory)."
This expert also believes that structured high-speed scenes may be easier to solve by traditional algorithms rather than structured Urban scenes may have more room to take advantage of AI models.
Zhitu uses reinforcement learning algorithms in the decision-making process of high-speed ramp intersection scenarios.
Dr. Yang Wenli, CEO of Lingjun Technology, believes that the importance and applicability of deep learning algorithms are not exactly the same in different modules of autonomous driving.
“From the process of perception-fusion-prediction-decision-planning-control, the earlier the part, the better the effect of the deep learning algorithm.”
Dr. Yang Wenli said: "The perceptual system has a large amount of data processing, but it allows errors and has a certain tolerance for accuracy and recall (because subsequent fusion and decision-making can correct perceptual errors), so deep learning methods are more suitable.
“In the decision-making part, rules and learning each have their own strengths, so we use a hybrid architecture, use deep learning methods to improve performance, and use rule methods to ensure safety.
"The control system has a small amount of data, strong requirements for reliability, and absolutely no room for error. The use of deep learning algorithms not only cannot improve performance, but may also introduce uncertainties and unreliable factors, so it is not suitable for use. Deep learning algorithm. Moreover, the underlying technology that controls this has a history of 200 years. The mathematical proof is perfect, and there is no need to use deep learning."
Moreover, in the case of errors in the planning algorithm , the control algorithm also needs to be safe. In the PID of the classic control algorithm, in order to ensure safety, the control layer can refuse to execute the "wrong instructions" issued by the planning layer. For example, if the plan issues an instruction to drive the vehicle to 120 km/h, but the vehicle's own hardware cannot support the speed of 120 km/h, at this time, the control algorithm can do a check and directly refuse to execute the instruction to ensure that the vehicle can operate normally. Driving without overturning.
If the control algorithm wants to cover all errors, it must use a rule-based algorithm to ensure reliability.
Based on this, when it comes to "deep learning algorithms are applied to regulation and control", what everyone calls "regulation and control" actually mainly refers to prediction, decision-making and planning, and does not include control.
In addition to solving practical technical problems, deep learning algorithms also have a huge advantage that cannot be ignored, which is to make full use of GPU or neural network dedicated processors to reduce CPU computing power consumption.
In order to realize certain functions, the complex rule codes used in the early stages of autonomous driving development will consume a lot of CPU computing power. However, the design of on-board chips generally tends to increase the computing power of deep learning, and the computing power of the CPU part is limited. Engineers must find ways to save. By replacing a large number of complex rule-based algorithms with deep learning algorithms, a lot of CPU computing power can be saved.
Qingzhou Zhihang planning and control algorithm experts believe that urban NOA mass production vehicles generally use embedded chips, and the CPUs are based on ARM architecture. The CPU computing power resources cannot be compared with L4 Demo vehicles based on X86 architecture. Therefore, if you want to achieve a driving experience close to L4, it is necessary for suppliers to try to introduce deep learning algorithms in the regulation and control process and put the computing load on deep learning dedicated processors.
Since deep learning algorithms sound so "cool", why are control algorithms used in practice? What about still focusing on rules rather than deep learning algorithms?
This is because the deep learning algorithm also has many shortcomings. It is these shortcomings that cause the deep learning algorithm to be less effective than the rules in the short term.
1. It is difficult to determine whether the driving behavior of human drivers is "worth learning"
In an ideal situation, the more data used for training, the better the deep learning algorithm. The better the effect, but in some cases, as the amount of data increases, the effect of the algorithm may decrease. Why is this?
The deep learning algorithm must first learn the driving behavior of a large number of human drivers before it can be applied to the regulation algorithm. However, the problem is that different drivers have different driving styles, and the same driver’s driving style in different situations is also different. Not completely consistent, so how to ensure that the driving behavior of human drivers is good data worth learning?
For example, for some lane changing behaviors at intersections, we cannot change lanes by pressing the solid line when driving normally, but many drivers often like to change lanes by pressing the solid line.
Pony.ai’s control algorithm expert said: “Human drivers also have many bad habits when driving. If they just blindly pursue anthropomorphism and ignore the three main optimization goals of the autonomous driving system: safety, comfort and efficiency, That would be putting the cart before the horse. In fact, even if an AI model is used to make up for the shortcomings of traditional algorithms, it should be improved around these three dimensions."
Therefore, in many cases, only pre-determined Only when the data is "cleaned" can the deep learning algorithm perform better. However, how to ensure efficiency and quality for large-scale real-time data cleaning? This is a problem.
In response to this problem, some companies have artificially set some rules to evaluate whether the driver drives well or not, and then filter out the ones that look bad, but this is Rules are used to "constrain" deep learning; moreover, this approach is too simple and crude, and may be like "throwing out the baby with the bath water" and accidentally deleting some useful driving behavior data.
2. Deep learning algorithms are more difficult to cope with demand changes than rules
The algorithm of a certain autonomous driving company stipulates that when driving on the road, the safe distance from the vehicle in front should be Not less than five meters. One day, the product manager put forward a new requirement: when driving normally on the road, the safe distance from the vehicle in front was changed from no less than five meters to no less than ten meters.
This situation will expose a problem of deep learning algorithms: if engineers want to change the model according to requirements, they need to retrain the model, which is very costly. But if you use a rule-based algorithm, it is very simple at this time, and the problem can be solved with a few lines of code.
Moreover, if an accident occurs, it is difficult for the deep learning algorithm to quickly find out where the problem lies in the shortest possible time like the rule algorithm.
3. The effect of the deep learning algorithm may not be as good as the rules
Qingzhou Zhihang regulation and control algorithm experts said: If the data infrastructure is imperfect, When an AI model is introduced, the performance of the AI model is difficult to meet expectations.
For example, the ACC following distance can be determined through simple test calibration, but it may be difficult to achieve the same effect through an AI model. The main reason is that the data infrastructure of many companies is imperfect and there is no way to find diverse data in different car-following scenarios, making it difficult to train a model that meets or exceeds the performance of the rules.
Haomo Zhixing's control algorithm expert also said: "The current AI model effect has not completely defeated the rules. It tries to personify people but the effect is not good enough because of the model architecture and input and output definitions of this type of cognitive network. The data preparation is not yet complete, and it has not reached the level of perception model. Everyone is still exploring (but I think there will be a major breakthrough within a year)."
4. Deep learning The algorithm still needs rules to tell the truth
In the process of driving a vehicle on the road, errors in the decision-making algorithm cannot be completely avoided. Because even if a person is driving, no matter how good the experienced driver is, it is impossible to guarantee that his decision-making will always be completely correct, and the same is true for decision-making algorithms based on deep learning models.
Therefore, when designing the system, developers should consider the possibility that "the decision-making algorithm will make mistakes" and be "psychologically prepared" to "accept" the potential fact that the decision-making algorithm will make mistakes.
In this regard, Lingjun Technology adopts a hybrid decision-making architecture. The company’s CEO Dr. Yang Wenli said that “deep learning algorithms are good at learning from human driving experience, so they are mainly used to improve performance and are designed by programmers. The rules have been fully tested and demonstrated to ensure safety and have a higher priority."
Looking further down, the planning layer can also reject "wrong instructions" from decision-making. For example, the input speed and size of obstacles should be within a certain range (for example, there cannot be a car with a speed of 100m/s like "taking off"); another example is that the decision-making layer gives the instruction "change lanes to the left", but this When there are cars on the left, because there is no complete solution to the space planning when changing lanes at this time, the planning can directly "refuse to execute the instruction".
The planning algorithm mentioned here that rejects "wrong instructions" from the previous link can only be based on rules.
Based on the above, we can see that in those scenarios where deep learning algorithms can be used, the rules for playing the role of "cover-up" are still essential.
5. The introduction of deep learning algorithms means that a large amount of code needs to be replaced
There is also a very challenging problem for autonomous driving companies: if hybrid regulation is to be used Algorithms will inevitably face the problem of a large amount of code that needs to be replaced, and doing so will inevitably require a "pile of people."
Qingzhou planning and control algorithm experts said that the traditional rule-based system did not consider adding interfaces related to deep learning when designing, and the code base has been relatively solidified after years of development. It is very costly to reconstruct the code to introduce deep learning. Gao,
"But Qingzhou Zhihang took into account the widespread use of AI when it began to build the decision planning algorithm stack in 2019, and considered the situation after the introduction of the model in advance at each interface level, and made targeted structural adjustment.”
According to a senior expert who once served as the person in charge of the regulation algorithm at a leading Robotaxi company, after the introduction of deep learning in regulation, the workload of code rewriting is very large, because it involves online and offline , simulation and other systems transformation.
The author has heard complaints from users many times in the Internet industry, why programmers of a certain software always choose to patch rather than reconstruct the bloated system. This example is somewhat similar to the problem faced when replacing rule-based governance algorithms with hybrid governance algorithms - the benefits of refactoring are low.
You have rewritten. The improvement in system operation level is limited, but you have to pay a huge cost for it. This kind of "hard work may not necessarily be a thank you" thing is "unbearable" for any enterprise. "heaviness". This concern also makes autonomous driving companies have a mentality of "just want to watch from a distance, not want to play with" hybrid control algorithms.
In fact, after communicating with the control algorithm directors/engineers of several autonomous driving companies, the author found that unlike the "scene" when deep learning neural networks were applied to perception in 2014-2015, currently , everyone is generally “not in a hurry” about the application of deep learning algorithms in regulation and control.
It can be said that everyone still carries out pre-research and practical applications with a "dispensable" mentality.
To be more specific, autonomous driving companies know that "hybrid control algorithms" using rule-based deep learning algorithms are the future, but currently there are not many scenarios where hybrid control algorithms can be used, so why not Recruit a few people and "let them do their own research slowly."
Therefore, at this stage, the attitude of most autonomous driving companies towards hybrid planning and control algorithms is that they only use hybrid planning and control algorithms when dealing with very few scenarios such as ramps and intersections; in the past, rule-based algorithms were used The corner case that has been solved by the algorithm will not be moved as long as no "intolerable" problems are found.
There are also planning and control algorithm experts who said very bluntly that whether or not to use a hybrid planning and control algorithm depends on the demand. This is an engineering issue.
Based on the above logic, the current attitude of autonomous driving companies towards hybrid control algorithms is “dispensable”, and the scope of application of hybrid control algorithms in the future also needs to be “questioned.”
Let’s continue the discussion based on the logic of the previous section.
Many experts believe that no matter how mature deep learning algorithms and hybrid frameworks are, rule codes will always account for the majority of control algorithms, and deep learning algorithms only play a supporting role.
The key reason they gave is: deep learning algorithms have the characteristics of "unexplainability". Using deep learning algorithms to make predictions and decisions follows probability, rather than strict causality. Therefore, it will inevitably make people "uneasy" if it is allowed to assume the "first responsibility" for regulation.
Dr. Yang Wenli, CEO of Rulingjun Technology, said: "Deep learning is a black box. The 'scenario' understood by humans may be different from the 'scenario' understood by deep learning. There was an example of diss artificial intelligence before, in the image If you change a few pixels, the artificial intelligence will recognize the mistake; if you put a few black tapes on the traffic signs, Tesla will not be able to recognize it. Unless it is 'consistent' at the pixel level, what people think is the 'same scene' will not be recognized in the same way. From the perspective of deep learning, it is most likely a 'different scenario'.
"The deep learning algorithm is based on the fitting of input and output. It can obtain better anthropomorphism from the input and output characteristics, but in fact it is difficult to Understand the internal logic of driving. "To address the possibility of errors in decision-making using deep learning algorithms, we adopted a hybrid decision-making architecture."
The CTO of a commercial vehicle driverless company and a control algorithm expert from a new car-making force also hold similar views.
At a technology exchange event at the end of September, Horizon CTO Dr. Huang Chang also talked about this issue, but Huang Chang believed that in hybrid regulation algorithms, deep learning will be the mainstream, and rules will only As a "necessary aid" - to avoid obvious and semantically explainable errors. ”
Huang Chang explained that based on his experience, for a rule-based system that has been refined for many years and is well-tuned, if deep learning is only introduced into a certain local module, the improvement will be very significant. Limited, or even no improvement. "Because other modules are still designed based on rules. Under this large framework, if you replace a module with a deep learning algorithm, the cost of adapting other modules to this new module will be very high. The entire verification cycle is very long, so long that you don’t have the patience to complete it. ”
The person in charge of the control algorithm of a Robotaxi company and the Qingzhou control algorithm expert also believe that the future control algorithm will be based on deep learning and supplemented by rules.
Qingzhou Planning and Control Algorithm Experts said that in Qingzhou Zhihang, currently in the prediction process, deep learning algorithm related operations (including pre-processing, model reasoning and post-processing) account for close to 95% of the overall module running time. In decision-making and planning, the proportion of deep learning algorithm operations has also reached about 30%.
“It’s not easy to quantify, but in general, deep learning algorithms should account for 50%-60% of the regulation algorithms.”
He believes that, in the end, regulation will 80%-90% of the computing time of the control algorithm will be spent on the deep learning algorithm. Moreover, unlike what Dr. Yang Wenli mentioned earlier that “there is no need to use AI in the underlying control algorithm,” he believes that the control algorithm is rule-based, but deep learning models can also be used to optimize control parameters and increase the adaptability of the algorithm. sex.
He pointed out: "It is understandable if there were previous disputes in the industry about 'AI-based' or 'rules-based', but the information released by Tesla on its most recent AI DAY has shown , Tesla's extensive use of deep learning control algorithms performs very well, which reflects from the side that Musk's "first principles" can also be applied to deep learning-based control algorithms in a sense: If humans As a complex neural network, the brain can learn to drive through a large amount of experience, and the control algorithm can also use deep learning to obtain better performance and performance."
There is a big problem with deep learning algorithms The controversy is its "unexplainability". So, if a certain scene changes from "unknown" to "known", do we need to add another rule to the algorithm to "determine" it?
Regarding the author’s question, a control algorithm expert from a Robotaxi company believes that it is “feasible” to determine the uncertainty of AI’s scene recognition through rules; however, Huang Chang and Qingzhou Control Algorithm experts believe it is “unfeasible” and “unnecessary.”
Huang Chang said: "We cannot force this system to be 'explainable' - because the so-called 'explainable' means using rules to constrain it, or even relying entirely on rules, like an expert system, the result is , a system that runs smoothly in city A will still require a large number of engineers to debug when running in city B; moreover, it will make those corner cases that are complex and cannot be described by artificial rules "unsolvable". ”
It can be seen that forcing the system to be “interpretable” returns to the “limitations of rules” mentioned in the first section of this article.
Huang Chang believes that when introducing deep learning algorithms into regulation, we don’t have to worry too much about whether the system is “explainable”.
"I am communicating with you. Between us, I can fully understand you and deduce your entire set of logic, but I don't know some of your underlying intuition and some hidden traits. ( Even you may not know it yourself), but this does not affect our ability to build trust and communicate smoothly. In the same way, we do not need to limit our exploration of the application boundaries of deep learning algorithms by whether they are 'explainable'.
“I read a post some time ago, saying that the difference between deep learning and classical statistics is that classical statistics still tries to use rule models to make the entire system ‘interpretable’, but deep learning Breaking through this, it even goes to extremes to optimize the ultimate goal.
“People will make mistakes, rule systems will make mistakes, and data systems will also make mistakes, but in a broad statistical sense, when the data scale is large enough, deep learning algorithms will be an order of magnitude better than the rule system. Therefore, we can use it with confidence.
"On this basis, visualize things like the environment model so that people can understand it, and then introduce a small number of necessary rules to constrain the depth in the last stage of planning Learn the algorithm and that's enough. "
Huang Chang believes that in the era of Algorithm 2.0, AI has the ability to complete self-adaptation. Therefore, autonomous driving companies may need to design some "algorithms on top of the algorithm" so that it can help the basic algorithm become more efficient. Iterate and adapt well in application scenarios.
Qingzhou control algorithm experts believe that as the author mentioned, "When a scenario changes from unknown to known, use rules to give the response plan to ' The idea of "defining it" is difficult to work in practice, because the scenes in reality are often dynamically adjusted - they may be clear at first, but as they change, they become "blurred", so it is very difficult to It is difficult to describe it clearly using rules.
Regarding the "unexplainability" of deep learning algorithms, his response is: "Many people will say that deep learning is 'unexplainable', but in fact, if you look carefully Design your deep learning algorithm so that it can achieve the effect of 'interpretability', and even this explainability may be stronger than the rules. ”
He gave an example: a self-driving vehicle turns right at an intersection and needs to avoid a car going straight. At this time, it is difficult to use rules to exhaustively determine the speed of the straight-moving vehicle and the distance between the two vehicles. What measures should be taken by the autonomous driving system at what time? But if a deep learning algorithm is used, as long as the positions and status of the self-driving car and potential obstacles (through vehicles) are given, as well as some relevant environmental and historical information, the trained neural network You can calculate the probabilities of "you let me" or "I let you" respectively. This is the same idea as when people make decisions when driving. "Probability is easier to explain than rules."
Speaking of this, the author thought of a topic that does not seem to be directly related to the topic of this article, but is deeply related - what is human "intuition" ?
I have often seen people who do business management research say that entrepreneurs often rely not on research but on intuition at critical moments when making decisions on major matters. Few people can explain clearly what this intuition is, but the author feels that this "intuition" is actually a data-driven "deep learning algorithm."
The so-called "intuitive accuracy" is not a talent. Behind it is the accumulation of information and the cognition formed on this basis. Similar to big data analysis, there is correlation between conclusions and information, but there is no causal relationship, or the causal relationship is not direct or obvious enough, and it is difficult for people to explain it clearly with logic, so they call it "intuition."
In this sense, we do not need to force deep learning algorithms to be "interpretable".
However, Hao Mo Zhixing’s control algorithm experts believe that even if a large number of AI models are used for cognition, it must be scenario-based and must have explicit and interpretable “intentions and meanings”. "This is our principle."
The expert said: "The strength of deep learning is that it can adapt to changing circumstances and can quickly provide a solution in various complex environments. However, its black box characteristics sometimes lead to wrong solutions, and rules-based The algorithm is actually all about 'I know, I know' and 'I know, I don't know'. The possibility of it giving a correct solution to an 'unknown' complex scene is lower than the deep learning method, but 'I can't do it correctly'. The answer does not affect my judgment of whether it is right or wrong after seeing the answer. Therefore, AI is the main force in answering questions, and the rules are responsible for verification. This is the method we use."
Pony.ai's regulation Algorithm experts believe that at this stage, it is difficult for deep learning algorithms to ensure security, so for a long time in the future, traditional algorithms and deep learning will have to be combined.
The regulatory algorithm expert said: "We believe that L2 is likely to be able to reach an AI-dominated state, because AI does not need to deal with all corner cases - after all, there are still drivers who can provide the final safety guarantee. , Tesla’s sharing on AI Day also illustrates this point.
“But the L4 algorithm is unlikely to be AI-based for a long time in the future. The main reason is its There are issues with interpretability and the ability to handle corner cases with security risks; however, with the advancement of technology, we believe that there will be more and more AI parts in the algorithm.
"In fact, traditional planning and control algorithms are by no means a pile of rules. Good traditional planning and control algorithms are more abstract mathematical models that summarize and summarize scenarios - using different mathematical tools to construct planning problems. Model and solve through optimization methods, it has the ability to characterize the essence of the problem and fundamentally ensure security within the theoretical scope.
"Therefore, the introduction of deep learning will not involve a lot of code rewriting , it should be more about how to combine traditional algorithms and deep learning, and use traditional algorithms to ensure driving safety. It can be said that traditional control algorithms and AI models are not a simple substitution relationship. Algorithm development is also step-by-step iterative. There is no need to spend a lot of money to replace the rule code with an AI model all at once, but in a step-by-step manner. ”
An architect from an OEM made a suggestion: run rule-based algorithms in the main system and deep learning algorithms in shadow mode, and wait until the training effect of the deep learning algorithm exceeds the rules. At that time, the deep learning algorithm is deployed into the main system.
Attached are two cases about the application of deep learning algorithms in autonomous driving control
1. Zhijia : Reduce fuel consumption through deep learning algorithms
Zhijia’s application of data-driven decision-making has also experienced a transition from rule-based decision-making to rule-data hybrid decision-making. Cui Dixiao, chief scientist of Zhijia Technology Zeng said that they believe that the timing of lane changes is very suitable for using deep learning algorithms to assist decision-making.
Because trunk logistics has very high requirements on timeliness and fuel consumption, on the one hand, inappropriate lane changes will cause drivers to use automatic The proportion of lane-changing functions is reduced. On the other hand, the rule-based lane-changing algorithm is conservative. If the "lane-changing" command is only triggered when the speed of the vehicle in front is too low, it will lead to more subsequent braking, thereby bringing down the overall timeliness. and fuel consumption. In order to reduce fuel consumption, the autonomous driving system collects data to observe how drivers make decisions, and then studies more human-like lane-changing decisions.
This plan has gone through two stages.
The first stage is when the lane changing decision-making is changed from the driver’s “self-triggered” form to “system recommendation”. At this stage, engineers will continue to study why the driver made a lane-changing decision at a certain point, and then formulate rules for lane-changing recommendations. After the system gives the lane change suggestion, if the driver allows the lane change, the system considers the decision at this time to be appropriate. If the driver thinks that the lane change should not be made at this time, the driver does not need to adopt the lane change suggestion.
In the second stage, engineers will also collect manual lane changing data in different scenarios, and combine the data collected in the first stage with the system lane changing decisions and the driver's lane changing decisions in inconsistent scenarios (such as the driver triggering a lane change) , the system is not triggered; or the system triggers lane change, but the driver does not adopt it) for labeling and model training, a lane change decision model is obtained, the original planning and control algorithm of the system is improved, and the fuel-saving performance is further improved.
2. Qingzhou Zhihang: Use deep learning algorithms to improve the “anthropomorphism” of the system
According to Qingzhou Zhihang’s prediction and control experts, at this stage, The Qingzhou control algorithm reserves a large number of interfaces for deep learning and actively develops deep learning control algorithms. The ultimate goal of Qingzhou Zhihang's overall planning and control algorithm architecture is to achieve AI First (deep learning algorithm priority), let the model do most of the behavioral output, and then use rules to complement it (the rules mainly deal with safety-related issues).
Qingzhou Zhihang believes that regulatory and control capabilities will be an important manifestation of the functional differentiation of urban NOA. At present, Qingzhou Zhihang has begun to try to use AI as much as possible in the regulation and control aspects of the urban NOA plan, so that the system's decision-making can be "close to humans" in most cases and "beyond humans" in extreme cases.
For example, Qingzhou already has specialized deep learning algorithms to handle certain key planning scenarios that will be encountered when using urban NOA, such as lane change timing, ramp merge timing, and intersection gaming.
In order to ensure the predictability and controllability of the deep learning algorithm, Qingzhou Zhihang will set verification on the input and output to ensure whether the input data meets the current design requirements. At the input level, rules will be used to select applicable scenarios for deep learning algorithms to ensure the rationality of the model input and limit the model to run within the scope of the designed problem. At the output level, rules will also be used to verify the output to ensure The output achieves the preset effect. At the same time, if you encounter a scene that the model cannot handle well, you can also automatically record time points and scene-related information, and then update the model through data closed loop to make the model have stronger generalization ability and adapt to more scenes.
At present, in Qingzhou Zhihang, the technology of generating the "possible future trajectory" of the vehicle through deep learning (such as imitation learning or reinforcement learning) is relatively mature, and is trained by collecting a large number of expert driving trajectories. Model, now the trajectories generated by the model do not need to be modified in 90% of cases; but in 10% of cases, the trajectories generated by deep learning may not be effective or even pose safety risks. In this case, a framework is needed to Correct/modify it. Qingzhou uses its domestically pioneered space-time joint algorithm (an algorithm based on numerical optimization) to fine-tune the trajectories generated by deep learning.
Qingzhou planning and control algorithm experts say that the space-time joint planning algorithm is more suitable for use in conjunction with deep learning algorithms than the traditional horizontal and vertical separation. Because the trajectory generated by the deep learning algorithm itself is actually a combined space-time trajectory. This trajectory not only contains position, speed, and direction information, but also time information. Therefore, the spatio-temporal joint planning algorithm can be better integrated with the results of the model, and thus can seamlessly add security protection to the deep learning algorithm.
The above is the detailed content of Ten Thousand Words Interpretation of the Application of Deep Learning Algorithms in Autonomous Driving Regulation and Control. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
