This article will continue to analyze the core technologies involved in autonomous driving. For the content of the first "Perception" and the second "Decision-making" content, please click on the original historical text to view.
3. Implementation
If the environmental perception system is equivalent to the driver's eyes, and the decision-making and planning system is equivalent to the driver's brain, then the executive control system is equivalent to the driver's hands and feet.
Specifically, the automatic driving control execution system means that after the system makes a decision and plan, it replaces the driver to control the vehicle and feeds back to the underlying module to perform tasks.
It can be said that the execution control system is the basis for the driving of autonomous vehicles. The various control systems of the vehicle need to be connected to the decision-making system through a bus, and can accurately control the degree of acceleration, braking, and steering according to the bus instructions issued by the decision-making system. Driving actions such as light control to achieve autonomous driving of the vehicle.
This article will introduce in detail the core technology, mainstream control algorithms and technical solutions of the automatic driving control execution module.
Ξ 1. Core technology
The core technology of automatic driving control execution mainly includes the longitudinal control and lateral control technology of the vehicle.
Longitudinal control, that is, the vehicle's drive and brake control, refers to the realization of precise follow-up of the desired vehicle speed through the coordination of the accelerator and brake. Lateral control, that is, through the adjustment of the steering wheel angle and the control of the tire force, the path tracking of the autonomous vehicle is realized.
1) Longitudinal control
Vehicle longitudinal control refers to the control in the direction of the driving speed, that is, the automatic control of the vehicle speed and the distance between the vehicle and the front and rear vehicles or obstacles.
Autonomous vehicles use the comprehensive control method of accelerator and brake to track the predetermined vehicle speed. Various motor-engine-transmission models, vehicle operation models and braking process models are combined with different control algorithms to form a variety of In the longitudinal control mode, the typical longitudinal control system structure is shown in the figure below:
As one of the most important control systems of autonomous vehicles, the automatic driving longitudinal control system is one of the effective ways to solve traffic jams and reduce the incidence of traffic accidents.
The longitudinal control system responds quickly to dangerous scenes, and the collision avoidance control is accurate and effective, which can avoid traffic accidents and casualties to the greatest extent. In addition, on the premise of ensuring driving safety, the longitudinal control system can also shorten the distance between the workshops, effectively increase the road traffic rate, and reduce the environmental pollution caused by traffic jams.
2) Horizontal control
Vehicle lateral control refers to control perpendicular to the direction of movement, that is, steering control. The goal of the lateral control system is to control the car to automatically maintain the desired driving route, and to have good riding comfort and stability under different vehicle speeds, loads, wind resistance, and road conditions.
The structure of a typical lateral control system is shown in the following figure: 
Vehicle lateral control can be roughly divided into two basic design methods: a method based on driver simulation and a control method based on vehicle dynamics model.
The method based on driver simulation can be divided into two in detail. One is to design the controller using a simpler dynamic model and driver manipulation rules, and the other is to train the controller to obtain the control algorithm by using the data of the driver's manipulation process .
Based on the method of vehicle dynamics model, it is necessary to establish a more accurate vehicle lateral motion model. A typical model is the single-track model, which considers that the left and right sides of the car have the same characteristics.
Ξ 2. Control algorithm
Automatic driving control methods can be divided into two types, namely traditional control methods and intelligent control methods.
1) Traditional control method
The traditional control methods mainly include PID control, fuzzy control, optimal control, sliding mode control (predictive model control MPC) and so on.
PID control, also known as proportional integral derivative control, is one of the earliest control strategies developed. In simple terms, the principle is to form a control deviation based on the given value and the actual output value, and the deviation is proportional, integral, and differential to form a control variable through a linear combination to control the controlled object.
Due to its simple algorithm, good robustness and high reliability, about 90% of the control loops still have PID structure. 
Fuzzy control, the full name of fuzzy logic control, is a computer digital control technology based on fuzzy set theory, fuzzy language variables and fuzzy logic inference.
Compared with classical control theory, the biggest feature of fuzzy logic control strategy is that it does not require accurate mathematical formulas to establish an accurate mathematical model of the controlled object, so it can greatly simplify the complexity of system design and mathematical modeling, and improve system modeling. And the efficiency of simulation control. However, the design of fuzzy control lacks systemicity, and there are certain problems in the control of complex systems.
Optimal control focuses on the study of basic conditions and comprehensive methods to optimize the performance indicators of the control system. It can be summarized as: For a controlled dynamic system or movement process, find an optimal control plan from a kind of allowed control plan, so that the movement of the system is transferred from an initial state to a specified target state. At the same time, its performance index value is the best.
Sliding mode control is also called variable structure control, which is essentially a special kind of nonlinear control. The difference between this control strategy and other controls is that the "structure" of the system is not fixed, and it can change purposefully and continuously according to the current state of the system during the dynamic process, forcing the system to move in accordance with a predetermined "sliding mode" state trajectory.
Since the sliding mode can be designed and has nothing to do with the object parameters and disturbances, the sliding control has the advantages of fast response, corresponding parameter changes and disturbances, no online identification of the system, and simple physical realization.
However, sliding control is not without its drawbacks. In practical applications, when the state trajectory reaches the sliding mode surface, it is difficult to strictly slide along the sliding mode surface to the equilibrium point. Instead, it travels back and forth on both sides to approach the equilibrium point, which will cause vibration and affect normal applications.
2) Intelligent control method
The biggest difference between the intelligent control method and the traditional control method is that the intelligent control method pays more attention to the application of the control object model and the application of comprehensive information learning. Common intelligent control methods mainly include model-based control, neural network control and deep learning methods.
Model-based control is generally called model predictive control. Its current control action is obtained by solving a finite time domain open-loop optimal control problem at each sampling instant.
The basic principle can be summarized as follows: At each sampling moment, according to the current measurement information currently obtained, an open-loop optimization problem in a finite time domain is solved online, and the first element of the obtained control sequence is applied to the controlled object , At a sampling moment, repeat the above process, then refresh the optimization problem with the new measurement value and re-solve it.
The advantage of this control method is that the accuracy of the model is not high, the modeling is convenient, and because the non-minimized description model is used, the system has good robustness and stability. 
Neural network control can regard the control problem as a pattern recognition problem, and the recognized pattern is mapped to the "change" signal of the "behavior" signal. The most significant feature of neural control is the ability to learn. It is achieved by continuously modifying the connection weights between neurons and discretely storing them in the connection network. It has a good effect on the control of nonlinear systems and systems that are difficult to model.
The deep learning method can obtain deep-level feature representation, avoid the complexity of manually selecting features and the dimensional disaster of high-dimensional data, and has great advantages in feature extraction and model fitting.
Since the autonomous driving system needs to minimize human participation, the ability of deep learning to automatically learn state characteristics makes deep learning more advantageous in the research of autonomous driving systems.
Ξ 3. Technical solution
According to the mapping process from the driving environment to the driving action, automatic driving control technology can be divided into two different schemes: indirect control and direct control.
1) Indirect control method based on planning-tracking
The indirect control scheme for autonomous driving can be simply summarized as: According to the current vehicle behavior requirements, plan a space-feasible and time-controllable collision-free safe motion trajectory under the conditions of the vehicle's own kinematics and dynamics constraints, and then design an appropriate The control law tracks the generated target trajectory to realize autonomous driving.
The principle is shown in the figure below: 
2) Direct control method based on artificial intelligence
Because the driving environment of autonomous vehicles has the characteristics of uncertainty, non-repeatability and unpredictability, it is difficult to establish an accurate mathematical model to design the control law, so traditional control strategies can no longer meet the requirements of autonomous driving control.
In this context, the direct control method based on artificial intelligence has become the mainstream form of the current automatic driving control system. 
The decision-making control model based on artificial intelligence essentially simulates the human brain's perception of the external environment information and the car body's own information, and at the same time, the process of obtaining continuous and stable output from driving experience and online learning mechanism.
This control mode can effectively improve the self-driving car's adaptability to different scenarios, and represents the mainstream development direction of the automatic driving control execution system in the future.
References for this article:
1. "2017 China Artificial Intelligence Series White Paper-Intelligent Driving"-China Artificial Intelligence Society
2. "2018 Artificial Intelligence Autonomous Driving Research Report"-Aminer