An extreme driving planning control method for an automatic driving racing car

Abstract

The application provides a limit driving planning control method for an automatic driving racing car, a multi-stage reinforcement learning course with gradually increased difficulty of a backward reinforcement learning module, a vehicle trajectory planning of an entire racing track is completed by using an expert guide strategy and a reinforcement learning strategy in turn, and an application range of the reinforcement learning strategy is gradually increased in a training process; a risk perception variation module combines a curiosity driving mechanism and a variation algorithm, performs variation operation on a key risk state and a reinforcement learning action network, promotes the reinforcement learning algorithm to learn limit driving knowledge contained in the key risk state, and improves robustness and safety in a critical risk state; and a model predictive control module generates a local optimal control command based on a nonlinear model predictive control algorithm, tracks a target trajectory provided by the backward reinforcement learning module, makes the vehicle dynamics constraint satisfied and potential risks predictable, and can realize motion control of the automatic driving racing car.

Description

Technical Field

[0001] This invention relates to autonomous driving systems, and more specifically to a method for extreme driving planning and control of autonomous racing cars. Background Technology

[0002] In recent years, with the continuous advancement of artificial intelligence and sensing technology, autonomous driving technology has seen significant development and widespread application. Autonomous racing is one attempt to apply it to motorsports. Autonomous racing refers to racing cars controlled by an autonomous driving system without human intervention. The field of autonomous racing is developing rapidly, bringing entirely new experiences and challenges to motorsports, and is of great significance in promoting the development of autonomous driving technology, improving driving safety and reliability, and optimizing transportation systems.

[0003] In motorsports, professional drivers precisely manipulate their cars to reach performance limits, achieving maximum speed and minimum lap times while ensuring safety. Clearly, designing the planning and control system for autonomous racing cars is far more challenging than designing autonomous driving systems for everyday driving, as it requires a comprehensive consideration of the vehicle's extreme characteristics, complex track conditions, and a precise balance between driving performance and vehicle safety. Through testing and optimization under extreme conditions, the performance of autonomous driving systems can be improved, ensuring their effectiveness in various complex driving scenarios. Simultaneously, autonomous racing cars can provide an opportunity for innovation and advancement in autonomous driving technology. Applying this technology to everyday driving scenarios allows autonomous vehicles to possess the high-level driving skills of professional drivers, thereby fully expanding the control potential and safety thresholds of autonomous vehicles under extreme driving conditions. This is of great significance for further reducing traffic accident rates and losses, and accelerating the practical application of advanced autonomous driving technologies. Summary of the Invention

[0004] In view of this, the present invention provides a limit driving planning and control method for autonomous racing cars, which can combine expert experience data, reinforcement learning algorithms and classical control to drive autonomous racing cars to reach their performance limits while ensuring safety, thereby achieving maximum speed and minimum lap time.

[0005] A method for extreme driving planning and control of autonomous racing cars includes the following steps:

[0006] Step 1: Divide the entire track T into N parts, denoted as follows;

[0007] T = [T1,T2,…,T] i ,…,T N (1)

[0008] In the formula, T idenotes the i-th track phase; in each training round, the backward reinforcement learning module adopts a reinforcement learning algorithm to use the expert-guided strategy π g previously obtained in the first h phases for exploration learning using the reinforcement learning algorithm in the remaining N-h phases until the overall performance M π reaches a preset performance threshold or reaches a specified number of training rounds, it will enter the next course learning phase h = h-1. h gradually decreases from N-1 to 0 by iteration, and the exploration range of reinforcement learning gradually retreats from the last track to the entire track;

[0009] wherein the state space S of the reinforcement learning algorithm is set as:

[0010]

[0011] wherein h denotes the iteration phase of course learning; s F is the curve length along the reference line direction, d F is the normal distance from the reference line; β, ψ and respectively represent the vehicle's center of mass side slip angle, heading angle and yaw rate. V represents the total vehicle speed, V s and respectively represent the current state and the speed component of the expert-guided strategy at s F = s on the tangent of the reference path; a g represents the total acceleration of the vehicle, with the unit being gravity acceleration g, and M is the current vehicle state mode;

[0012] The action space A of the reinforcement learning algorithm is set as:

[0013] A = [d F_com , ΔV s_com ] T (3)

[0014] wherein d F_com represents the normal distance from the current target position to the reference path, represents the V s residual term compared with the reference path, wherein V s_com represents the V s speed component command value finally delivered to the model predictive control module;

[0015] The reward function R of the reinforcement learning algorithm is composed of three reward types, namely, instantaneous reward R i , phase reward R p and terminal state reward R l , and is expressed as:

[0016] R = R i + R p+R l (4);

[0017] Step 2: Perform mutation operations on the key risk states and reinforcement learning action networks, specifically including:

[0018] Step 21: Identify the risk status

[0019] During reinforcement learning training, quadruples (s) containing key risk states will be used. t ,a t ,r t ,s t+1 The data is stored in another experience pool D′ and processed through an auxiliary cost function J. c To identify risk status, when J c >J c0 ,s t Defined as a risk state, where J c0 The safety threshold of the cost function;

[0020] Step 22: Intrinsic Reward Design

[0021] When a batch of samples is sampled from the experience pool D and the risk state s is identified t At that time, by monitoring the key motion states of the vehicle Superimposed Gaussian noise is used to perform a mutation operation, resulting in M ​​enhanced states g(s). t Based on the reward function R and cost function J c Define every two enhanced states and The variation index c between ij Its definition is:

[0022]

[0023] Select the corresponding c ij The two enhanced states with the highest value and And introduce a related c ij The highest value among the M enhanced states Intrinsic rewards driven by curiosity are directly proportional to R. int Its definition is:

[0024]

[0025] In the formula, λ is the temperature coefficient, η is the decay weight, and σ is the number of training rounds; for J c ≤J c0 The general state, intrinsic reward R int Set to 0; compare the external reward R and the internal reward R in equation (4). int Add them together to get the total reward of the reinforcement learning algorithm.

[0026] Step 23: Network weight variation

[0027] To analyze the sensitivity to risk states, a batch of samples N′ is sampled from the experience pool D′ every set number of rounds, assuming each sample consists of n quadruplets including the risk state; the network weight φ for the current action is calculated. a Risk sensitivity Its definition is:

[0028]

[0029] In the formula, |s| and |a| represent the dimensions of the state space and action space of the reinforcement learning algorithm, respectively; subsequently, Gaussian noise is used as the basis for the equation. and risk sensitivity Setting perturbations for the action network weights Its definition is:

[0030]

[0031] Step 24: Network Weight Update

[0032] Based on the generated n perturbations The corresponding n mutated action network weights are generated:

[0033]

[0034] Based on the performance of the above n mutated similar action networks under their respective risk states, the optimal action network weights are selected, denoted as φ. best Replace the current action network weights φ with them. a And used for subsequent iterative updates;

[0035] The backward reinforcement learning module obtains the target trajectory, including the reference path and reference speed, based on the updated action network weights.

[0036] Step 3: Based on the nonlinear model predictive control algorithm, track the reference path and reference speed provided by the backward reinforcement learning module, and generate locally optimal control commands to satisfy vehicle dynamics constraints and predict potential risks. This nonlinear model predictive control algorithm is expressed as follows:

[0037]

[0038] stx k+1 =f(x) k ,u k (19)

[0039]

[0040] In the formula, k represents the k-th step, and C is the control step size, which is consistent with the prediction step size. The local optimal control sequence is defined within C control steps. and This represents the state range and control variable range of a nonlinear model predictive control algorithm. For the target longitudinal displacement; W = {w r ,w v ,w u} represent the cost function weights for path tracking error, speed tracking error, and the rate of change of the control variable Δu, respectively.

[0041] Better, instantaneous reward R i The definition is as follows:

[0042] R i =k v ·V s +k a ·G a +R is (5)

[0043] in,

[0044]

[0045] In the formula, G a R represents the current vehicle performance utilization rate. iS k represents the reward based on the solution obtained from the model predictive control module. v k a and k S These are the weight parameters for each item;

[0046] Stage Reward R p The definition is as follows:

[0047]

[0048] in, and t j Let k represent the entry speed, exit speed, and passage time of the j-th track phase, respectively. p1 k p2 and k p3 These are the respective weight parameters;

[0049] Termination state reward R l Represented as:

[0050]

[0051] Where, k r ko t are negative constants, representing the reward function values ​​corresponding to the vehicle rolling over and going out of bounds, respectively; total t′ is the time it takes for the vehicle to complete a full lap of the track in this round. total k is the time taken to complete the entire lap of the track under expert guidance strategy. t A negative constant represents the weighting coefficient corresponding to the overall cycle.

[0052] Preferably, in step 21, J c Including the cost of rollover risk J cr and the cost of instability risk J ci :

[0053] J c =J cr +J ci (9)

[0054] in,

[0055]

[0056]

[0057] In the formula, The lateral load transfer ratio is used as an indicator of vehicle rollover risk; F zi Let LTR be the vertical load on each wheel, i = 1, 2, 3, 4; when LTR > 1, the vehicle will roll over, and its safety threshold LTR is... t Set to 0.8; J ci Used to measure excessive yaw rate under the vehicle's adhesion limit. The risk of vehicle instability caused by the sideslip angle β; k r and k i This represents the weight parameters for each item.

[0058] Preferably, in step 24, the optimal action network weights, which have higher reward values ​​and lower cost function values ​​compared to the original network weights, are selected as φ. best .

[0059] The present invention has the following beneficial effects:

[0060] This invention provides a method for extreme driving planning and control of autonomous racing cars, comprising: 1) a backward reinforcement learning module: designing a multi-stage reinforcement learning course with progressively increasing difficulty, sequentially utilizing expert-guided strategies and reinforcement learning strategies to complete vehicle trajectory planning for the entire track, and gradually backward during training to improve the application scope of the reinforcement learning strategy; 2) a risk perception mutation module: combining a curiosity-driven mechanism and mutation algorithm, performing mutation operations on key risk states and reinforcement learning action networks to promote the reinforcement learning algorithm to learn the extreme driving knowledge contained in key risk states, thereby improving robustness and safety under critical risk states; 3) a model predictive control module: generating locally optimal control commands based on a nonlinear model predictive control algorithm, tracking the target trajectory provided by the backward reinforcement learning module, so as to satisfy vehicle dynamics constraints and predict potential risks. In practical applications, the autonomous driving system of this invention can be used in conjunction with various sensors and actuators to achieve motion planning and control of autonomous racing cars. For example, it can be used in conjunction with sensors such as LiDAR, cameras, and GPS to acquire environmental and vehicle status information; it can also be used in conjunction with actuators such as brakes, throttles, and steering gears to achieve motion control of autonomous racing cars. Attached Figure Description

[0061] Figure 1 This is a flowchart of the extreme driving planning and control method for autonomous racing cars according to the present invention.

[0062] Figure 2 The example diagram shows a comparison of the test results after training and the trajectory and motion state of the expert-guided strategy during a key curve of a certain track.

[0063] Figure 3 This is a comparison diagram of the test results after training and the motion state of the expert-guided strategy on the entire track in the example. Detailed Implementation

[0064] This invention designs a method for extreme driving planning and control of autonomous racing cars, such as... Figure 1 As shown, it includes the following steps:

[0065] Step 1: The backtracking reinforcement learning module employs a curriculum-based reinforcement learning algorithm. It utilizes an expert-guided strategy to create an initial curriculum, which is then taken over and explored by the reinforcement learning algorithm. The learning process is accelerated by gradually increasing the task difficulty. For example... Figure 1 As shown, the entire track T is first divided into N parts based on the type of straightaways and curves, represented as follows:

[0066] T = [T1,T2,…,T] i ,…,T N (1)

[0067] In the formula, T i Let represent the i-th track stage. In each training round, the agent of the reinforcement learning algorithm uses an expert-guided strategy π for the first h stages. g In the remaining Nh stages, reinforcement learning algorithms are used for exploratory learning until the overall performance M is achieved. π Reaching the preset performance threshold Upon reaching a specified number of training rounds, the next learning phase h = h-1 will begin. h gradually decreases iteratively from N-1 to 0, representing a shift in the scope of reinforcement learning's exploration from the last track to the entire track. The overall strategy can be represented as...

[0068] In this invention, the expert-guided strategy can be generated based on driver operation data, prior knowledge, or other heuristic methods, which will not be specifically described here. The following will illustrate the specific settings of the state space, action space, and reward function of the reinforcement learning algorithm using a particular embodiment.

[0069] In this embodiment, the state space S of the reinforcement learning algorithm is set as follows:

[0070]

[0071] In the formula, h represents the iteration stage of course learning. To simplify the trajectory planning, the reference trajectory of the expert-guided strategy is used as the reference line to convert the vehicle position into a position in the Serret-Frenet coordinate system, s F Let d be the curve length along the reference line direction. F β, ψ, and ψ represent the normal distances from the reference line. These represent the vehicle's sideslip angle, heading angle, and yaw rate, respectively. V represents the total vehicle speed. s and These represent the current state and the expert guidance strategy in s, respectively. F =s is the velocity component tangent to the reference path. a g M represents the total acceleration of the vehicle, expressed in gravitational acceleration g. M is the current vehicle status mode, including: 1-normal driving, 2-out of bounds, 3-rollover, 4-reaching the destination.

[0072] The action space A of the reinforcement learning algorithm is set as follows:

[0073] A = [d F_com ,ΔV s_com ] T (3)

[0074] In the formula, d F_com This represents the normal distance from the current target position to the reference path. V represents the value compared to the reference path. s The residual term, where V s_com This represents the V that is ultimately passed to the model prediction control module. s Velocity component command value.

[0075] The reward function R of the reinforcement learning algorithm is composed of the instantaneous reward R0. i Stage Rewards R p and termination state reward R l It consists of three reward types, represented as

[0076] R = R i +R p +R l (4)

[0077] Instantaneous reward R i Considering vehicle speed, performance utilization, and the solution obtained from the model predictive control module, the goal is to incentivize the autonomous racing car to fully utilize its performance to achieve the shortest lap time. i The definition is as follows:

[0078] R i =k v ·V s +k a ·G a +R is (5)

[0079] in,

[0080]

[0081] In the formula, G a R represents the current vehicle performance utilization rate. iS k represents the reward based on the solution obtained from the model predictive control module. v k a and k s These are the weight parameters for each item.

[0082] Stage Reward R p The agent is given feedback after each stage of the track. Since the training objective is to minimize the overall lap time, each stage's objective includes not only minimizing the time but also considering entry and exit speeds to ensure that adjacent track sections are also traversed as quickly as possible. p The definition is as follows:

[0083]

[0084] in and t j Let k represent the entry speed, exit speed, and passage time of the j-th track phase, respectively. p1 kp2 and k p3 These are the weight parameters for each item.

[0085] When the autonomous driving race car reaches a terminated state, the training round ends, and a termination reward R will be given based on different termination state patterns. l There are three possible outcomes for this termination state: rollover, crossing the boundary, and reaching the destination. l Represented as:

[0086]

[0087] Where, k r k o t are negative constants, representing the reward function values ​​corresponding to the vehicle rolling over and going out of bounds, respectively; total t′ is the time it takes for the vehicle to complete a full lap of the track in this round. total k is the time taken to complete the entire lap of the track under expert guidance strategy. t A negative constant represents the weight coefficient corresponding to the overall cycle. Step 2: The risk perception mutation module performs mutation operations on key risk states and the reinforcement learning action network to promote the reinforcement learning algorithm to learn the extreme driving knowledge contained in the key risk states and improve training efficiency, while avoiding the reinforcement learning algorithm from generating overly conservative policies. This module mainly includes the following four steps.

[0088] Step 21: Identify the risk status

[0089] During reinforcement learning training, quadruples (s) containing key risk states will be used. t ,a t ,r t ,s t+1 Stored in another experience pool D′; via auxiliary cost function J c To identify risk status, when J c >J c0 ,s t Defined as a risk state, where J c0 The safety threshold of the cost function. c The cost of rollover risk J cr and the cost of instability risk J ci Composition, defined as:

[0090] J c =J cr +J ci (9)

[0091] in,

[0092]

[0093]

[0094] In the formula, The lateral load transfer ratio is used as an indicator of vehicle rollover risk. F zi (i = 1, 2, 3, 4) represent the vertical loads on each wheel. When LTR > 1, the vehicle will roll over; its safety threshold LTR is... t Set to 0.8. J ci Used to measure excessive yaw rate under the vehicle's adhesion limit. The risk of vehicle instability caused by the sideslip angle β. r and k i This represents the weighting parameters for each item. Step 22: Intrinsic Reward Design

[0095] While risky states may lead to training failures, they may approximate the optimal policy and contain unknown, extreme driving knowledge. Therefore, this invention introduces an intrinsic reward mechanism driven by curiosity to improve the sampling efficiency of these high-value samples.

[0096] When a batch of samples is sampled from the experience pool D and the risk state s is identified t At that time, by monitoring the key motion states of the vehicle Superimposed Gaussian noise is used to perform a mutation operation, resulting in M ​​enhanced states g(s). t Other states can be solved based on the vehicle model. To quantify the intrinsic incentive value of curiosity corresponding to this risk state, this invention is based on the reward function R and the cost function J. c Define every two enhanced states and The variation index c between ij Its definition is:

[0097]

[0098] Select the corresponding c ij The two enhanced states with the highest value and And introduce a related c ij maximum value Intrinsic rewards driven by curiosity are directly proportional to R. int Its definition is

[0099]

[0100] In the formula, λ is the temperature coefficient, η is the decay weight, and σ is the number of training rounds. For J c ≤J c0 The general state, intrinsic reward R int Set it to 0. Combine the external reward R and the internal reward R in equation (4).int Add them together to get the total reward of the reinforcement learning algorithm.

[0101]

[0102] Step 23: Network weight variation

[0103] To analyze the sensitivity of risk states, each τ GA In each round, a batch of samples N′ is sampled from the experience pool D′, consisting of n quadruplets including risk states. Based on the dynamic response characteristics of the risk states in N′ to the weights of the reinforcement learning action network, the current action network weights φ are calculated. a Risk sensitivity Its definition is:

[0104]

[0105] In the formula, |s| and |a| represent the dimensions of the state space and action space of the reinforcement learning algorithm, respectively.

[0106] Subsequently, Gaussian noise was introduced. The perturbation ∈, as the weight of the action network, is based on risk sensitivity. Normalize the perturbation ∈, where perturbation ∈ is defined as

[0107]

[0108] Step 24: Network Weight Update

[0109] Based on the perturbation ∈, n mutated action network weight individuals P = {p1,…,p} are generated. n The weights of each network are represented as follows:

[0110]

[0111] The performance of similar action networks after the above n mutation operations under their respective risk states, that is, the average external reward R of this series of action networks over a certain period under different initial risk states in N′. p and cost function J c_p The fitness function is defined as f(p) = (R... p J cp ),

[0112] Based on the fitness function, select the weights with higher reward values ​​R relative to the original network weights. p and lower cost function value J cp The best individual is denoted as φ. best Replace the current action network weight φ with it. a This data will be used for subsequent iterations and updates.

[0113] The backward reinforcement learning module obtains the target trajectory, including the reference path and reference speed, based on the updated action network weights.

[0114] Step 3: The model predictive control module tracks the reference path and reference speed provided by the backward reinforcement learning module based on a nonlinear model predictive control algorithm, and generates locally optimal control commands to satisfy vehicle dynamics constraints and predict potential risks. This nonlinear model predictive control algorithm is expressed as follows:

[0115]

[0116] stx k+1 =f(x) k ,u k (19)

[0117]

[0118] In the formula, k represents the k-th step, and C is the control step size, which is consistent with the prediction step size. It is a locally optimal control sequence within C control steps. and This represents the state range and control variable range of a nonlinear model predictive control algorithm. For the target longitudinal displacement, it can be determined based on the actual normal distance deviating from the reference normal. and target speed V s_com The result is obtained through calculation. W = {w r ,w v ,w u} represent the cost function weights for path tracking error, speed tracking error, and the rate of change of the control variable Δu, respectively.

[0119] In summary, the extreme driving planning and control system for autonomous racing cars provided by this invention was used for training and testing of embodiments in a simulation environment. Figure 2 This is a comparison of the test results after training with the trajectory and motion state of the expert-guided strategy during a key bend in a certain track. Figure 3 The diagram shows a comparison between the test results after training and the motion state of the expert-guided strategy across the entire track. In this embodiment, the expert-guided strategy took 111.0 seconds to traverse the entire track, while the trained driving strategy achieved a lap time of 105.4 seconds, a reduction of 5.0%, thus validating the effectiveness of the proposed system.

Claims

1. A method for extreme driving planning and control of autonomous racing cars, characterized in that, Includes the following steps: Step 1: Divide the entire track T into N parts, denoted as follows; T=[T1,T2,…,T i ,…,T N ] (1) In the formula, T i This represents the i-th track stage; in each training round, the backward reinforcement learning module uses a reinforcement learning algorithm to apply a pre-obtained expert guidance strategy π to the first h stages. g In the remaining Nh stages, reinforcement learning algorithms are used for exploratory learning until the overall performance M is achieved. π Reaching the preset performance threshold Or, once the specified number of training rounds is reached, the next learning phase h = h-1 will begin; h gradually decreases from N-1 to 0, and the scope of reinforcement learning exploration gradually retreats from the last track to the entire track; The state space S of the reinforcement learning algorithm is set as follows: In the formula, h represents the iterative stage of course learning; s F Let d be the curve length along the reference line direction. F For the normal distance from the reference line; β, ψ and These represent the vehicle's sideslip angle, heading angle, and yaw rate, respectively; V represents the total vehicle speed. s and These represent the current state and the expert guidance strategy in s, respectively. F =The velocity component s is tangent to the reference path; a g This indicates the total acceleration of the vehicle, expressed in gravitational acceleration g, where M represents the current vehicle state mode. The action space A of the reinforcement learning algorithm is set as follows: A=[d F_com ,ΔV s_com ] T (3) In the formula, d F_com This represents the normal distance from the current target position to the reference path. V represents the value compared to the reference path. s The residual term, where V s_com This represents the V that is ultimately passed to the model prediction control module. s Velocity component command value; The reward function R of the reinforcement learning algorithm is composed of the instantaneous reward R0. w Stage Rewards R p and termination state reward R l It consists of three reward types, represented as follows: R=R w +R p +R l (4) Step 2: Perform mutation operations on the key risk states and reinforcement learning action networks, specifically including: Step 21: Identify the risk status During reinforcement learning training, quadruples (s) containing key risk states will be used. t ,a t ,r t ,s t+1 ) is stored in another experience pool D′, and then processed through the cost function J. c To identify risk status, when J c >J c0 ,s t Defined as a risk state, where J c0 The safety threshold of the cost function; Step 22: Intrinsic Reward Design When a batch of samples is sampled from the experience pool D and the risk state s is identified t At that time, by monitoring the key motion states of the vehicle Superimposed Gaussian noise is used to perform a mutation operation, resulting in M ​​enhanced states g(s). t Based on the reward function R and cost function J c Define every two enhanced states g i* (s t ) and g j* (s t The variation index c between ) ij Its definition is: Select the corresponding c ij The two enhanced states g with the highest values i* (s t ) and g j* (s t ), and introduce a relation to c ij The highest value c among the M enhanced states i*j* Intrinsic rewards driven by curiosity are directly proportional to R. int Its definition is: R int =λc i*j* exp(-ησ) (13) In the formula, λ is the temperature coefficient, η is the decay weight, and σ is the number of training rounds; for J c ≤J c0 The state, intrinsic reward R int Set to 0; compare the external reward R and the internal reward R in equation (4). int Add them together to get the total reward of the reinforcement learning algorithm. Step 23: Network weight variation To analyze the sensitivity to risk states, a batch of samples N′ is sampled from the experience pool D′ every set number of rounds, assuming each sample consists of n quadruplets including the risk state; the network weight φ for the current action is calculated. a Risk sensitivity Its definition is: In the formula, |s| and |a| represent the dimensions of the state space and action space of the reinforcement learning algorithm, respectively; subsequently, Gaussian noise is used as the basis for the equation. and risk sensitivity Setting perturbations for the action network weights Its definition is: Step 24: Network Weight Update Based on the generated n perturbations The corresponding n mutated action network weights are generated: Based on the performance of the above n mutated similar action networks under their respective risk states, the optimal action network weights are selected, denoted as φ. best Replace the current action network weights φ with them. a And used for subsequent iterative updates; The backward reinforcement learning module obtains the target trajectory, including the reference path and reference speed, based on the updated action network weights. Step 3: Based on the nonlinear model predictive control algorithm, track the reference path and reference speed provided by the backward reinforcement learning module, and generate locally optimal control commands to satisfy vehicle dynamics constraints and predict potential risks. This nonlinear model predictive control algorithm is expressed as follows: s.t.x k+1 =f(x k ,u k ) (19) In the formula, k represents the k-th step, and C is the control step size, which is consistent with the prediction step size; The local optimal control sequence is defined within C control steps. and This represents the state range and control variable range of a nonlinear model predictive control algorithm. For the target longitudinal displacement; W = {w r ,w v ,w u } represent the cost function weights for path tracking error, speed tracking error, and the rate of change of the control variable Δu, respectively.

2. The extreme driving planning and control method for autonomous racing cars as described in claim 1, characterized in that, instantaneously... Reward R w The definition is as follows: R w =k v ·V s +k a ·G a +R is (5) in, In the formula, G a R represents the current vehicle performance utilization rate. iS k represents the reward based on the solution obtained from the model predictive control module. v k a and k S These are the weight parameters for each item; Stage Reward R p The definition is as follows: in, and t j Let k represent the entry speed, exit speed, and passage time of the j-th track phase, respectively. p1 k p2 and k p3 These are the respective weight parameters; Termination state reward R l Represented as: Where, k r k o t are negative constants, representing the reward function values ​​corresponding to the vehicle rolling over and going out of bounds, respectively; total t′ is the time it takes for the vehicle to complete a full lap of the track in this round. total k is the time taken to complete the entire lap of the track under expert guidance strategy. t A negative constant represents the weighting coefficient corresponding to the overall cycle.

3. The extreme driving planning and control method for autonomous racing cars as described in claim 1, characterized in that, In step 21, J c Including the cost of rollover risk J cr and the cost of instability risk J ci : J c =J cr +J ci (9) Among them, In the formula, The lateral load transfer ratio is used as an indicator of vehicle rollover risk; F zm Let m represent the vertical load on each wheel, where m = 1, 2, 3, 4; when LTR > 1, the vehicle will roll over, and its safety threshold LTR is... t Set to 0.8; J ci Used to measure excessive yaw rate under the vehicle's adhesion limit. The risk of vehicle instability caused by the sideslip angle β; k r and k i This represents the weight parameters for each item.

4. The extreme driving planning and control method for autonomous racing cars as described in claim 1, characterized in that, In step 24, the optimal action network weights, which have higher reward values ​​and lower cost function values ​​compared to the original network weights, are selected as φ. best .

Images