A method and device for autonomous driving of two-wheeled vehicles based on near-end strategy optimization

Abstract

This application relates to the field of control technology, providing a method and apparatus for road condition decision-making for autonomous driving of two-wheeled vehicles based on near-end strategy optimization. The method includes: acquiring the state information and environmental perception data of the two-wheeled vehicle; acquiring boundary reference information of the currently passable area based on the state information and environmental perception data; calculating the action probability distribution of the two-wheeled vehicle under the current traffic state based on the boundary reference information, state information, and environmental perception data; sampling the optimal control parameters of the two-wheeled vehicle from the action probability distribution by optimizing the objective function; adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and continuously collecting and updating the state information and environmental perception data of the two-wheeled vehicle in real time; re-executing the autonomous driving road condition decision-making process for the two-wheeled vehicle until the driving stopping conditions of the two-wheeled vehicle are met. This application embodiment can adapt to complex and changing outdoor road traffic environments, achieving reliable operation of autonomous driving road condition decision-making for two-wheeled vehicles.

Description

Technical Field

[0001] This application relates to the field of control technology, and more specifically to a method and apparatus for road condition decision-making for autonomous driving of two-wheeled vehicles based on near-end strategy optimization. Background Technology

[0002] With the rapid development of artificial intelligence, the Internet of Things, and intelligent transportation systems (ITS) technologies, autonomous driving technology has gradually expanded from the traditional automotive field to two-wheeled vehicles, such as electric bicycles and shared bicycles. Bicycles, as a green, flexible, and economical short-distance urban mode of transportation, have advantages in alleviating traffic congestion and reducing carbon emissions. However, traditional bicycles rely heavily on manual operation, resulting in low safety and poor traffic efficiency in complex urban environments. Therefore, realizing autonomous driving capabilities for bicycles, especially with highly robust and real-time path planning functions, has become an important research direction in the field of intelligent micro-mobility.

[0003] In the autonomous driving technology system, path planning is one of the core modules. Its task is to generate a safe, efficient, and traffic-compliant driving trajectory from the starting position to the target position for the vehicle based on perceived environmental information. Currently, path planning methods for four-wheeled vehicles are based on global path planning using graph search algorithms such as A* and Dijkstra, and local obstacle avoidance and trajectory optimization methods using strategies such as Dynamic Window (DWA) and Model Predictive Control (MPC). However, these methods have the following technical problems when directly applied to two-wheeled bicycles: First, the bicycle dynamics model has nonholonomic constraints, underactuated characteristics, and strong nonlinearity, and its balance control and steering response mechanisms differ from those of four-wheeled vehicles. Second, urban road environments contain a large number of dynamic obstacles (such as pedestrians and other cyclists) and narrow passages, which places higher demands on the real-time performance, smoothness, and safety of path planning. In summary, there is an urgent need for a novel technical solution to realize path planning in bicycle autonomous driving systems and solve at least one of the technical problems existing in the current technology. Summary of the Invention

[0004] This application provides a method and apparatus for road condition decision-making for autonomous driving of two-wheeled vehicles based on near-end strategy optimization. It can adapt to complex and ever-changing outdoor road traffic environments and achieve reliable operation of road condition decision-making for autonomous driving of two-wheeled vehicles.

[0005] In a first aspect, embodiments of this application provide a method for road condition decision-making for autonomous driving of a two-wheeled vehicle based on near-end strategy optimization. The two-wheeled vehicle operates in an outdoor traffic environment containing dynamic obstacles and road boundaries. The method includes:

[0006] Acquire the status information and environmental perception data of the two-wheeled vehicle;

[0007] Based on state information and environmental perception data, obtain boundary reference information of the currently passable area;

[0008] Based on the boundary reference information, state information and environmental perception data, calculate the probability distribution of actions of the two-wheeled vehicle under the current traffic condition;

[0009] The optimal control parameters of the two-wheeled vehicle are obtained by sampling from the action probability distribution by optimizing the objective function; wherein, the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment;

[0010] Adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. Jump to the step of obtaining the boundary reference information of the currently passable area based on the status information and environmental perception data, and re-execute the two-wheeled vehicle autonomous driving road condition decision-making process until the driving stopping conditions of the two-wheeled vehicle are met.

[0011] Secondly, embodiments of this application provide a two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization, which has the function of implementing the two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization provided in the first aspect above. The function can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware.

[0012] In one embodiment, the two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization is used when the two-wheeled vehicle operates in an outdoor traffic environment containing dynamic obstacles and road boundaries. The device includes:

[0013] The acquisition module is configured to acquire the status information and environmental perception data of the two-wheeled vehicle; and based on the status information and environmental perception data, acquire the boundary reference information of the currently passable area.

[0014] The decision module is configured to calculate the probability distribution of actions of the two-wheeled vehicle under the current traffic state based on the boundary reference information, state information, and environmental perception data; and to sample the optimal control parameters of the two-wheeled vehicle from the probability distribution by optimizing the objective function; wherein the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment;

[0015] The control module is configured to adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. It then jumps to the step executed by the decision module to obtain the boundary reference information of the currently passable area based on the status information and environmental perception data, so as to re-execute the two-wheeled vehicle autonomous driving road condition decision process until the driving stopping conditions of the two-wheeled vehicle are met.

[0016] Thirdly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a computer, cause the computer to perform the two-wheeled vehicle autonomous driving road condition decision-making method based on proximal policy optimization as described in the first aspect.

[0017] Fourthly, embodiments of this application provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the two-wheeled vehicle autonomous driving road condition decision-making method based on proximal policy optimization as described in the first aspect.

[0018] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute the two-wheeled vehicle autonomous driving road condition decision-making method based on proximal strategy optimization provided in the first aspect.

[0019] Compared to existing technologies, this application's embodiments introduce a near-end policy optimization (PPO) reinforcement learning method into the road condition decision-making process of two-wheeled vehicle autonomous driving, and explicitly incorporate boundary reference information of passable areas during the decision-making process, achieving joint constraint modeling of dynamic obstacles and road boundaries in complex outdoor traffic environments. Since this application's embodiments make continuous action space decisions based on policy probability distributions and iteratively optimize the objective function, which includes safety constraints, path tracking error, and control smoothness terms, through multi-round simulation interaction, rather than relying on fixed rules, a single cost function, or offline experience-based control parameter setting in existing technologies, this application's embodiments can adaptively output optimal control parameters that meet safety and stability requirements even when traffic conditions are continuously changing. This effectively reduces decision jitter and collision risks caused by environmental uncertainty or perceived noise. Therefore, this application's embodiments, through boundary constraint perception and reinforcement learning policy optimization, can improve the robustness and real-time performance of two-wheeled vehicle decisions in dynamic traffic scenarios. Since the objective function is obtained through multiple rounds of interactive learning in various traffic scenarios, and takes into account safety, tracking accuracy and control smoothness, the action decision results obtained in this application embodiment are more stable, continuous and in line with actual driving needs. This helps to achieve safe driving and comfortable control effects, and can adapt to complex and changing outdoor road traffic environments, enabling reliable operation of autonomous driving road condition decision-making for two-wheeled vehicles. Attached Figure Description

[0020] The objectives, features, and advantages of the embodiments of this application will become readily understood by referring to the accompanying drawings and the detailed description of the embodiments. Wherein:

[0021] Figure 1 This is a flowchart illustrating the road condition decision-making method for autonomous driving of two-wheeled vehicles based on near-end strategy optimization in the embodiments of this application;

[0022] Figure 2 This is a schematic diagram of the structure of a two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization according to an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of the structure of a terminal device according to an embodiment of this application. Detailed Implementation

[0024] Currently, existing technologies have yielded relatively mature research results on path planning methods for four-wheeled vehicles, such as global path planning based on graph search algorithms like A* and Dijkstra, and local obstacle avoidance and trajectory optimization methods employing strategies like Dynamic Window (DWA) and Model Predictive Control (MPC). However, when these methods are directly applied to two-wheeled bicycles, they present the following technical challenges: Firstly, bicycle dynamics models exhibit nonholonomic constraints, underactuation, and strong nonlinearity, and their balance control and steering response mechanisms differ from those of four-wheeled vehicles. Secondly, urban road environments contain numerous dynamic obstacles (such as pedestrians and other cyclists) and narrow passages, placing higher demands on the real-time performance, smoothness, and safety of path planning. Therefore, a novel technical solution is urgently needed to realize path planning for autonomous bicycle systems, addressing at least one of the technical problems existing in current technologies.

[0025] Compared to existing technologies, this application provides a method and apparatus for road condition decision-making for autonomous driving of two-wheeled vehicles based on near-end strategy optimization. Specifically, the two-wheeled vehicle operates in an outdoor traffic environment containing dynamic obstacles and road boundaries. Based on this, the following steps are taken: acquiring the state information and environmental perception data of the two-wheeled vehicle; acquiring boundary reference information of the currently passable area based on the state information and environmental perception data; calculating the action probability distribution of the two-wheeled vehicle under the current traffic state based on the boundary reference information, state information, and environmental perception data; sampling the optimal control parameters of the two-wheeled vehicle from the action probability distribution by optimizing an objective function; wherein the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained through multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment; adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collecting and updating the state information and environmental perception data of the two-wheeled vehicle in real time; jumping to the step of acquiring the boundary reference information of the currently passable area based on the state information and environmental perception data; and re-executing the autonomous driving road condition decision-making process for the two-wheeled vehicle until the driving stopping conditions of the two-wheeled vehicle are met.

[0026] In the above embodiments, a near-end policy optimization (PPO) reinforcement learning algorithm is introduced into the road condition decision-making process of two-wheeled vehicle autonomous driving. Environmental perception data, vehicle state information, and boundary reference information of passable areas are deeply fused to construct a probabilistic policy-based action decision-making mechanism, achieving adaptive decision-making in complex and dynamic traffic environments. Since this embodiment uses an online decision-making method based on policy gradients and obtains optimized objective function parameters through multi-round interactive learning in a simulated environment for various traffic scenarios, rather than relying on fixed rules, preset thresholds, or manually designed control strategies in a single scenario as in existing technologies, this embodiment can continuously output stable and reasonable control decisions in the face of dynamic obstacle changes, irregular road boundaries, and uncertain traffic participant behavior, effectively improving the robustness and generalization ability of two-wheeled vehicles in real road environments.

[0027] Furthermore, in this embodiment, by introducing a multi-objective optimization objective function consisting of a safety constraint term, a path tracking error term, and a control smoothness term, the decision-making process not only focuses on whether the vehicle can pass through the current road segment, but also comprehensively considers driving safety, trajectory following accuracy, and the continuity and smoothness of control commands. Since the safety constraint term explicitly constrains the safe distance between the two-wheeled vehicle and dynamic obstacles and road boundaries, the path tracking error term ensures that the vehicle travels stably along the desired path within the passable area, and the control smoothness term suppresses drastic changes in control parameters, thereby avoiding vehicle shaking, sharp turns, or instability, the optimal control parameters obtained in this embodiment have higher stability and feasibility in actual execution.

[0028] Furthermore, since the embodiments of this application adopt a closed-loop iterative mechanism, after each execution of the optimal control parameters, the state information and environmental perception data of the two-wheeled vehicle are updated in real time, and the boundary reference information of the passable area is recalculated, thereby continuously correcting the action probability distribution and optimizing the decision. Therefore, the road condition decision results obtained by the embodiments of this application can maintain real-time performance and consistency in dynamic traffic environments, and can give full play to the self-learning and adaptive advantages of reinforcement learning strategies in complex environments. It can well cope with practical application needs such as road congestion, sudden obstacle appearance, and rapid changes in traffic scenarios.

[0029] In summary, the embodiments of this application, through a decision-making scheme based on near-end strategy optimization, can comprehensively improve the safety, stability, and environmental adaptability of autonomous driving for two-wheeled vehicles, and enhance the decision-making effect of autonomous driving in complex outdoor traffic environments.

[0030] It should be noted that the server involved in the embodiments of this application can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0031] The terminal devices involved in the embodiments of this application can be devices that provide voice and / or data connectivity to users, handheld devices with wireless connectivity, or other processing devices connected to a wireless modem. Examples include mobile phones (or "cellular" phones) and computers with mobile terminals, such as portable, pocket-sized, handheld, computer-embedded, or vehicle-mounted mobile devices that exchange voice and / or data with a wireless access network.

[0032] Reference Figure 2 , Figure 2This document provides a flowchart illustrating a road condition decision-making method for autonomous driving of two-wheeled vehicles based on near-end strategy optimization, as provided in an embodiment of this application. This method can be executed by an autonomous driving road condition decision-making device for two-wheeled vehicles based on near-end strategy optimization. It can be applied to outdoor traffic scenarios for autonomous driving of two-wheeled vehicles, such as complex road environments containing dynamic obstacles, road boundaries, and uncertain traffic participants. By acquiring the state information of the two-wheeled vehicle and environmental perception data, a model is created for the currently passable area. Based on a near-end strategy optimization algorithm, the probability distribution of the two-wheeled vehicle's actions under the current traffic state is calculated. Optimal control parameters that satisfy safety constraints, path tracking requirements, and control smoothness are sampled from this probability distribution. These optimal control parameters are then used to replace the control commands generated by traditional rule-based or single optimization models and input into the two-wheeled vehicle control system, enabling autonomous decision-making and stable driving of the two-wheeled vehicle under complex road conditions. The method includes steps 101-105:

[0033] Step 101: Obtain the status information and environmental perception data of the two-wheeled vehicle.

[0034] Two-wheeled vehicles can be understood as vehicles or autonomous driving platforms with two driving wheels, capable of operating in outdoor traffic environments, and supporting speed and / or steering adjustments through control parameters. Examples include electric bicycles, electric mopeds, and electric motorcycles, which possess motor drives, electronic control units, and relatively complete sensing and execution interfaces, making them suitable for deploying reinforcement learning-based autonomous driving or intelligent decision-making methods. Fuel-powered two-wheeled motorcycles, including ordinary motorcycles and light motorcycles, achieve automatic speed and steering control through electronic throttle, steer-by-wire, or assisted steering actuators, suitable for intelligent driving decision-making applications in complex traffic environments. Two-wheeled vehicles with automatic balancing or assisted stabilization structures, such as two-wheeled vehicles with gyro stabilization systems and self-balancing two-wheeled vehicles (non-standing), maintain vehicle stability during autonomous driving through control algorithms and actuators. Two-wheeled autonomous driving platforms used for testing or operation in closed scenarios include two-wheeled robotic vehicles for research or industrial testing and autonomous driving two-wheeled experimental platforms. Two-wheeled vehicles with intelligent assisted driving functions include intelligent motorcycles or intelligent electric vehicles with automatic cruise control, lane keeping, obstacle avoidance, or autonomous decision-making capabilities.

[0035] In this step, the onboard sensors of the two-wheeled vehicle acquire basic input data for autonomous driving decisions. This state information includes, but is not limited to, the following kinematic or dynamic parameters: the vehicle's speed, acceleration, tilt angle, current position coordinates, vehicle attitude, driving speed, acceleration, steering angle, and wheel speed. The environmental perception data includes the position, distance, direction of movement, and speed of surrounding obstacles. For example, this includes the position and motion state of surrounding obstacles, road boundary information, and behavioral characteristics of other traffic participants obtained through LiDAR, cameras, millimeter-wave radar, or multi-sensor fusion methods. After time synchronization and preprocessing, the aforementioned state information and environmental perception data serve as input for subsequent road condition decisions.

[0036] Step 102: Based on the status information and environmental perception data, obtain the boundary reference information of the currently passable area.

[0037] In this step, based on the two-wheeled vehicle's current position, driving direction, and environmental perception data, the surrounding road space is analyzed to identify road boundaries, obstacle outlines, and dynamic safety buffer zones. Based on this, a spatial representation of the two-wheeled vehicle's traversable area at the current moment is constructed. The boundary reference information includes the left and right boundary curves of the traversable area, the width of the traversable road, and obstacle restriction areas, which are used to impose spatial constraints on subsequent action decision-making processes, preventing the generation of control actions that exceed the traversable area.

[0038] As an optional embodiment, in step 102, obtaining the boundary reference information of the currently passable area based on state information and environmental perception data includes: obtaining the coordinate values ​​of multiple boundary reference points and the safety margin corresponding to each boundary reference point. Each boundary reference point includes at least the boundary reference point closest to the left boundary of the road, the boundary reference point closest to the right boundary of the road, and the boundary reference point with the largest safety margin.

[0039] For example, in step 102, the road boundaries and obstacle boundaries within the current driving area of ​​the two-wheeled vehicle are identified based on environmental perception data. Multiple boundary reference points are selected on the road and obstacle boundaries, and the spatial coordinate values ​​corresponding to each boundary reference point are obtained. Simultaneously, combining the state information of the two-wheeled vehicle and the environmental perception data, a corresponding safety margin is calculated for each boundary reference point.

[0040] It is worth noting that the safety margin is used to characterize the minimum safe distance between a two-wheeled vehicle and its corresponding boundary reference point under the current traffic conditions. Its calculation process comprehensively considers at least the relative position and motion of obstacles, road geometry, and the motion of the two-wheeled vehicle itself, allowing the safety margin to be dynamically adjusted according to changes in the traffic environment and vehicle status. By comparing and selecting the safety margins corresponding to each boundary reference point, several representative boundary reference points are determined.

[0041] The boundary reference points include at least: the boundary reference point closest to the left boundary of the road, the boundary reference point closest to the right boundary of the road, and the boundary reference point with the largest safety margin. The boundary reference points closest to the left and right boundaries of the road are used to define the left and right boundaries of the lateral passable area for two-wheeled vehicles, while the boundary reference point with the largest safety margin reflects the boundary constraint position that has the greatest impact on driving safety under the current traffic conditions.

[0042] The boundary reference information constructed in the above manner can dynamically, safely, and accurately describe the passable area of ​​two-wheeled vehicles while ensuring physical feasibility, providing a reliable spatial constraint basis for the calculation of subsequent action probability distribution.

[0043] Step 103: Based on the boundary reference information, state information and environmental perception data, calculate the probability distribution of actions of the two-wheeled vehicle under the current traffic condition.

[0044] In this step, the state information, environmental perception data, and boundary reference information of the passable area are collectively constructed into a state vector of the current traffic state, and input into a policy network built based on a near-end policy optimization algorithm. The policy network outputs the probability distribution of each candidate control action in the continuous action space according to the current traffic state. The control actions may include target speed, acceleration, steering angle, or combinations thereof, thereby representing the set of feasible decisions for the two-wheeled vehicle in the current traffic state in probabilistic form.

[0045] As an optional embodiment, step 103, which involves calculating the action probability distribution of the two-wheeled vehicle under the current traffic condition based on the boundary reference information, state information, and environmental perception data, includes: calculating a safe distance index for the currently passable area based on the safety margin in the boundary reference information; calculating a current path deviation index based on the two-wheeled vehicle position coordinates in the state information and obstacle dynamic information in the environmental perception data; fusing the safe distance index and the path deviation index to generate constraint weight coefficients; combining the constraint weight coefficients with the speed and acceleration parameters in the state information through a preset constraint mapping relationship to obtain a parameter vector for the action probability distribution; and generating a physically feasible action probability distribution based on the parameter vector, wherein the output space of the action probability distribution is constrained within preset boundaries of the speed adjustment range and the steering angle range.

[0046] Specifically, in step 103, the probability distribution of the two-wheeled vehicle's actions under the current traffic condition is calculated based on the boundary reference information, state information, and environmental perception data. First, based on the safety margin corresponding to each boundary reference point in the boundary reference information, a safe distance index for the passable area of ​​the two-wheeled vehicle under the current traffic condition is calculated. This safe distance index characterizes the overall safety margin between the two-wheeled vehicle and road boundaries and dynamic obstacles. When the safety margin is small, the corresponding safe distance index decreases to reflect the increased risk level of the current driving environment. Second, based on the two-wheeled vehicle's position coordinates and driving direction in the state information, and the obstacle positions, speeds, and directions of motion in the environmental perception data, a path deviation index is calculated relative to the desired driving path for the two-wheeled vehicle's current trajectory. This path deviation index measures the degree of deviation between the actual driving state of the two-wheeled vehicle and the centerline of the target driving area. Subsequently, the safe distance index and the path deviation index are fused to generate constraint weight coefficients for adjusting the action output. Specifically, when the safe distance index decreases or the path deviation increases, the corresponding constraint weight coefficients increase to enhance the constraint strength on the control action. When the safety distance index increases and the path deviation is small, the corresponding constraint weight coefficient decreases to improve the flexibility of action selection. Next, through a preset constraint mapping relationship, the constraint weight coefficient is combined with the speed and acceleration parameters in the state information to obtain a parameter vector describing the action probability distribution. This parameter vector is used to adjust the mean, variance, or boundary range of each candidate control action in the continuous action space, ensuring that the action output matches the current traffic state and safety constraints. Finally, based on the parameter vector, an action probability distribution that satisfies the dynamic constraints and physical feasibility of the two-wheeled vehicle is generated. The output space of this action probability distribution is limited to a preset speed adjustment range and steering angle range to avoid generating control commands beyond the vehicle's executable capabilities, thereby representing the feasible decision set of the two-wheeled vehicle in the current traffic state in probabilistic form.

[0047] Step 104: By optimizing the objective function, the optimal control parameters of the two-wheeled vehicle are sampled from the action probability distribution. The parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term. These parameters are obtained through multi-round interactive learning and iterative optimization of the two-wheeled vehicle's behavior under various traffic scenarios in a simulated environment.

[0048] In this step, the action probability distribution is constrained and optimized based on the near-end policy optimization algorithm, and the optimal control parameters that satisfy the preset objective function are sampled from it.

[0049] It can be explained that the proximal policy optimization algorithm is a policy optimization method for solving continuous decision-making problems, improving policy learning efficiency while ensuring policy update stability. The proximal policy optimization algorithm constructs a parameterized policy model, directly outputs the probability distribution of actions based on the environment state, and uses maximizing expected reward as the optimization objective, continuously updating policy parameters through repeated interactions with the environment. Unlike traditional value function-based optimization methods, the proximal policy optimization algorithm focuses on iteratively improving the policy itself, naturally adapting to continuous action spaces, and is suitable for applications such as autonomous driving that require high control precision and real-time performance. In its implementation, the proximal policy optimization algorithm introduces a policy update constraint mechanism to limit the difference between the new and old policies, avoiding policy performance oscillations or degradation due to excessively large single update magnitudes. This algorithm typically constructs an objective function with a pruning mechanism to constrain the range of policy probability changes during policy improvement, ensuring the policy gradually approaches the optimal solution while maintaining stability in the learning process. Furthermore, the proximal policy optimization algorithm can combine state value evaluation results to perform weighted updates of the policy gradient, thereby improving sample utilization efficiency while ensuring convergence. Due to its characteristics of convergence stability, relatively simple implementation, and low sensitivity to hyperparameters, the proximal policy optimization algorithm is particularly suitable for continuous control problems in complex dynamic environments. In autonomous driving applications, the proximal policy optimization algorithm can learn driving strategies through repeated interactions in various traffic scenarios, achieving a better balance between safety, smoothness, and environmental adaptability in the generated control actions. This provides a reliable algorithmic foundation for autonomous decision-making by vehicles in complex road conditions.

[0050] The objective function comprehensively considers safety constraints, path tracking error, and control smoothness. The safety constraints limit the minimum safe distance between the two-wheeled vehicle and dynamic obstacles and road boundaries. The path tracking error measures the deviation between the actual trajectory and the desired trajectory of the two-wheeled vehicle. The control smoothness term limits the variation of control parameters between adjacent time steps. The parameters of the objective function are obtained through multi-round interactive learning and iterative optimization of two-wheeled vehicle behavior under various traffic scenarios in a simulated environment, thereby improving the generalization ability and stability of the strategy under complex road conditions.

[0051] Understandably, in complex traffic environments, balancing driving safety, path tracking accuracy, and control process smoothness is crucial to guiding the strategy model to generate control parameters that meet actual driving needs. Safety constraints represent the safety relationship between the two-wheeled vehicle and surrounding dynamic obstacles and road boundaries. By constraining the relative distance between the two-wheeled vehicle and obstacles / road boundaries, the safety constraints significantly increase the cost when the actual distance is less than a preset safety distance threshold. This suppresses control actions that may lead to collisions or boundary violations, prompting the strategy to prioritize driving schemes with larger safety margins during decision-making, thereby improving the driving safety of two-wheeled vehicles in dynamic traffic environments.

[0052] The path tracking error term measures the degree of deviation between the actual trajectory of the two-wheeled vehicle and the desired trajectory. This term is calculated by comprehensively considering indicators such as lateral offset, heading angle error, or trajectory point distance. When the two-wheeled vehicle deviates significantly from the desired path, the corresponding cost increases, thereby guiding the strategy to generate control parameters that can gradually correct the driving direction and reduce trajectory deviation, enabling the two-wheeled vehicle to maintain effective tracking of the target path under obstacle avoidance and safety constraints.

[0053] The control smoothness term limits the variation of control parameters between adjacent time points, focusing on constraining the continuity of control quantities such as speed, acceleration, and steering angle over time. When control parameters change drastically between adjacent decision points, the cost of this term increases, thereby suppressing overly aggressive or discontinuous control behavior, avoiding problems such as vehicle instability, ride discomfort, or excessive load on actuators, and improving the smoothness and stability of the two-wheeled vehicle's driving process.

[0054] Furthermore, the weight parameters in the objective function are not fixed but are obtained by constructing various typical traffic scenarios in a simulated environment, conducting multi-round interactive learning on the decision-making behavior of two-wheeled vehicles, and continuously adjusting and optimizing them during the iteration process. In this way, the objective function can adapt to different road structures, traffic densities, and obstacle movement characteristics, improving the generalization ability and robustness of the strategy under complex road conditions, thereby achieving more stable and reliable autonomous driving decision-making results in practical applications.

[0055] Specifically, the parameters of the objective function were obtained through multi-round interactive learning and iterative optimization of two-wheeled vehicle behavior under various traffic scenarios in a simulated environment. For example, in a simulated urban street environment, 1000 interaction data were collected, revealing a safety incident rate as high as 18%. Initial weights were set to safety 0.5, path 0.3, and smoothness 0.2, but frequent instances of insufficient emergency braking due to pedestrians suddenly crossing the road occurred. Through 500 rounds of optimization, the safety weight was gradually increased to 0.65, while the path weight was reduced to 0.25. After optimization, the system prioritizes ensuring safe distances in pedestrian-dense areas; for example, when a pedestrian appears more than 5 meters away, the safety constraint score is higher, and the system automatically selects a deceleration command instead of pursuing path accuracy. Ultimately, the safety incident rate decreased to 6%, and the average path tracking deviation increased from 0.4 meters to 0.5 meters, but the safety-first principle significantly improved overall reliability.

[0056] In a simulated highway scenario, 2000 interactive tests were conducted. Excessive path tracking deviation negatively impacted driving comfort. Initial weights were: safety 0.6, path 0.3, and smoothness 0.1. However, on long straight sections, the vehicle's excessive focus on safety caused a 0.6-meter deviation from the expected path. Through 800 rounds of optimization, the safety weight was fine-tuned to 0.55, the path weight increased to 0.35, and the smoothness weight remained at 0.1. After optimization, greater emphasis was placed on path accuracy on highway sections. For example, on obstacle-free straight sections, the path tracking error term had a higher weight, allowing the vehicle to follow the planned path more smoothly. Ultimately, the path deviation was reduced to 0.2 meters, the safety incident rate remained below 3%, and unnecessary emergency braking was avoided.

[0057] In a simulated winding mountain road scenario, 1500 interactions were tested, revealing that large fluctuations in steering control led to trajectory instability. Initial weights were set to safety 0.5, path 0.4, and smoothness 0.1, but sudden changes in steering angle at sharp bends caused vehicle swaying. After optimization, steering smoothness is prioritized in bends; for example, when adjusting from a 10-degree to a 12-degree steering angle, the rate of change is calculated and limited to within 1 degree to prevent tire slippage. Ultimately, trajectory fluctuations were reduced, path deviation decreased from 0.35 meters to 0.25 meters, while the safe distance remained consistently greater than 0.8 meters.

[0058] In 3000 interactive tests simulating sudden obstacle scenarios, it was found that the initial weights of safety (0.5), path (0.3), and smoothness (0.2) resulted in a slow response. When an obstacle suddenly appeared 10 meters away, the system failed to decelerate in time, and the minimum safe distance dropped to 0.6 meters. Through 1000 rounds of optimization, the safety weight was significantly increased to 0.7, the path weight was reduced to 0.25, and the smoothness weight was reduced to 0.05. After optimization, safety constraints are prioritized for calculation within 0.5 seconds of obstacle appearance. For example, if the minimum safe distance is predicted to be 0.7 meters, the maximum safety penalty is immediately triggered, forcing the vehicle to decelerate. Ultimately, the safe distance compliance rate improved, and although path deviation increased, the safety risk was completely eliminated, meeting the primary requirements of emergency scenarios.

[0059] As an optional embodiment, in step 104, the optimal control parameters of the two-wheeled vehicle are sampled from the action probability distribution by optimizing the objective function. This includes: randomly sampling multiple candidate control parameters from the action probability distribution, wherein the candidate control parameters include speed adjustment parameters and steering angle parameters; wherein the value range of the speed adjustment parameters is constrained within a preset speed adjustment boundary, and the value range of the steering angle parameters is constrained within a preset steering angle boundary; for each candidate control parameter, a safety constraint term, a path tracking error term, and a control smoothness term are calculated respectively; a weighted sum of the safety constraint term, path tracking error term, and control smoothness term is calculated based on preset weight coefficients, and the weighted sum is used as the objective function value; the candidate control parameter with the largest objective function value is selected as the optimal control parameter.

[0060] Understandably, in step 104, multiple candidate control parameters are first randomly sampled from the action probability distribution according to a preset sampling strategy. These candidate control parameters include at least speed adjustment parameters and steering angle parameters. To ensure the physical executability of the sampling results, the value range of the speed adjustment parameter is limited to a preset speed adjustment boundary, and the value range of the steering angle parameter is limited to a preset steering angle boundary, thereby avoiding the generation of control commands exceeding the dynamic capabilities of the two-wheeled vehicle. Secondly, for each candidate control parameter, based on the current traffic state, the corresponding safety constraint term, path tracking error term, and control smoothness term are calculated. The safety constraint term is used to evaluate the minimum safe distance between the two-wheeled vehicle and road boundaries and dynamic obstacles under the action of the candidate control parameter. The path tracking error term is used to evaluate the degree of deviation between the predicted driving trajectory and the desired driving path after the two-wheeled vehicle executes the candidate control parameter; the control smoothness term is used to evaluate the magnitude of change between the candidate control parameter and the actual executed control parameter at the previous moment. Subsequently, based on preset weighting coefficients, the safety constraint term, path tracking error term, and control smoothness term are weighted and summed, and the resulting weighted sum is used as the objective function value of the corresponding candidate control parameter. By adjusting the magnitude of each weighting coefficient, a trade-off can be struck between safety, tracking accuracy, and control smoothness to adapt to different driving scenario requirements. Finally, the objective function values ​​corresponding to all candidate control parameters are compared, and the candidate control parameter with the largest objective function value is selected as the optimal control parameter for the current moment. Through the above method, under the premise of ensuring safety constraints and physical feasibility, the control command with the best overall performance is selected from the probability distribution to guide the autonomous driving behavior of the two-wheeled vehicle in the current traffic condition.

[0061] For example, in a two-wheeled vehicle autonomous driving system, multiple candidate control commands are first randomly selected from the action probability distribution output by the neural network. For instance, 10 different sets of control parameters are generated, each containing speed adjustment values ​​and steering angle values. To ensure these parameters are executable, the system sets physical constraint boundaries, such as limiting the speed adjustment range to between -5 m / s and +3 m / s, and the steering angle range to between -30 degrees and +30 degrees. If a sampled result exceeds these boundaries, the system automatically truncates it to the nearest valid value. This constraint mechanism prevents the generation of extreme commands that a two-wheeled vehicle cannot achieve, such as requiring an instantaneous 90-degree turn or a complete stop from high speed within 0.1 seconds.

[0062] For each candidate control parameter, the system predicts the two-wheeled vehicle's trajectory over the next two seconds and calculates the minimum distance to surrounding obstacles. For example, when the candidate parameters are a 15-degree right turn and an acceleration of 1 meter per second, the system simulates the vehicle entering the right lane, calculating a minimum distance of 1.2 meters to the right guardrail and 3.5 meters to the slow-moving vehicle ahead. The value of the safety constraint is determined based on these minimum distances; a higher score is awarded when all predicted distances are greater than the safety threshold of 1 meter. A sharp drop in score occurs if the predicted trajectory indicates a distance of less than 0.5 meters to an obstacle. This calculation considers the dynamic movement trends of obstacles, not just their current static distances.

[0063] The path tracking error term measures the ability of candidate control parameters to make the vehicle follow the planned path. For example, when a two-wheeled vehicle needs to travel along a winding mountain road, the system predicts the vehicle's position at the next 5 time steps after executing a certain candidate parameter, and then calculates the lateral offset distance between these predicted points and the corresponding points on the desired path. If the candidate parameter is a 10-degree left turn while maintaining the current speed, the predicted trajectory may deviate from the desired path by an average of 0.3 meters. Another parameter, a 12-degree left turn with slight deceleration, results in a predicted trajectory deviation of only 0.1 meters. The system converts the average offset distance into a score; the smaller the deviation, the higher the score. Simultaneously, the system also considers the direction and angle matching degree, ensuring that the vehicle's position is not only close to the path, but its driving direction is also consistent with the tangent direction of the path.

[0064] The control smoothness term evaluates the degree of change between the current candidate parameters and the actual parameters executed at the previous moment. For example, if the vehicle previously turned 5 degrees right at a speed of 8 meters per second, one current candidate parameter might be a 15-degree right turn at a speed of 7 meters per second, and another might be a 7-degree right turn at a speed of 7.8 meters per second. The system will calculate the changes in steering angle and speed separately. A 10-degree change in steering angle corresponds to a 1-meter-per-second speed change, while a 2-degree change in steering angle corresponds to a 0.2-meter-per-second speed change. The smaller the change, the higher the smoothness score. This design avoids abrupt changes in control commands, ensuring smooth movement of the two-wheeled vehicle, improving ride comfort and vehicle stability, which is especially important under slippery road conditions.

[0065] Finally, the three evaluation results are combined into a single objective function value according to their weights. For example, the weight of safety constraint is set to 0.6, the weight of path tracking error is 0.3, and the weight of control smoothness is 0.1. For a candidate parameter, if its safety constraint score is 0.9, its path tracking score is 0.7, and its smoothness score is 0.8, then the objective function value is calculated as 0.9 multiplied by 0.6 plus 0.7 multiplied by 0.3 plus 0.8 multiplied by 0.1, which equals 0.83. The system performs the same calculation for all candidate parameters. Assuming that the highest score among the 10 candidates is 0.89, corresponding to the parameters of turning left 8 degrees and decelerating 0.5 meters per second, this parameter is selected as the optimal control command at the current moment. In emergency obstacle avoidance scenarios, the safety weight can be dynamically increased to 0.8, allowing safety factors to dominate the decision-making process and ensuring the safe driving of two-wheeled vehicles in complex traffic environments.

[0066] Optionally, the weighted sum of the safety constraint term, path tracking error term, and control smoothness term is calculated based on preset weight coefficients, including: linearly weighting the safety constraint term, path tracking error term, and control smoothness term using preset weight coefficients. The preset weight coefficients are obtained through multiple rounds of interactive learning and iterative optimization in a simulation environment, with the weight coefficient for the safety constraint term being greater than that for the path tracking error term, and the weight coefficient for the path tracking error term being greater than that for the control smoothness term.

[0067] For example, 1000 rounds of testing were conducted in a simulated driving environment, recording the number of safety events, path deviation, and control fluctuation data after each round. The initial weights were set to safety 0.5, path 0.3, and smoothness 0.2. After 200 rounds of optimization, the safety weight increased to 0.6, the path weight to 0.3, and the smoothness weight to 0.1, resulting in a decrease in the safety event rate. In 500 rounds of testing, the weights were automatically adjusted; when encountering complex intersections, the safety weight further increased to 0.65, the path weight remained at 0.3, and the smoothness weight decreased to 0.05. The optimization process was based on real-time driving data, with the weight values ​​dynamically converging as the environment changed, ensuring that the safety weight was highest on highways, while the smoothness weight could be appropriately increased on open roads.

[0068] In pedestrian-intensive areas, a safety weight of 0.65 dominates the decision-making process. For example, when a two-wheeled vehicle approaches a zebra crossing, the safety constraint item of candidate parameter A scores 0.9, which multiplied by the safety weight of 0.65 equals 0.585; the path tracking error item scores 0.6, which multiplied by the path weight of 0.3 equals 0.18; and the control smoothness item scores 0.7, which multiplied by the smoothness weight of 0.05 equals 0.035. Parameters with higher safety scores are prioritized, even if the path deviation increases by 0.15 meters. The safety weight is always higher than other items to ensure immediate deceleration when a pedestrian suddenly appears, rather than pursuing path precision, thus avoiding accidents caused by insufficient safety distance.

[0069] For each candidate control parameter, the safety constraint term is calculated sequentially by multiplying it by the safety weight, the path tracking error term by multiplying it by the path weight, and the control smoothness term by multiplying it by the control smoothness weight. These three results are then summed to obtain the objective function value. For example, the safety constraint term 0.8 multiplied by the safety weight 0.65 equals 0.52, the path tracking error term 0.7 multiplied by the path weight 0.3 equals 0.21, and the control smoothness term 0.9 multiplied by the control smoothness weight 0.05 equals 0.045. Summing 0.52, 0.21, and 0.045 yields 0.775. This weighting multiplication maximizes the impact of safety factors on the objective function and minimizes the impact of smoothness, ensuring that the safety score significantly increases the objective function value in emergency situations, thereby selecting the safest control command.

[0070] The calculation methods for each item are explained below with specific examples.

[0071] In the above embodiments, optionally, the calculation of the safety constraint term includes: extracting the safety margin value corresponding to the boundary reference point of the currently passable area from the boundary reference information; extracting the real-time position coordinates, movement speed vector, and movement direction angle of surrounding obstacles from the environmental perception data; calculating the straight-line distance based on the current position coordinates of the two-wheeled vehicle and the position coordinates of the obstacle; predicting the minimum safe distance between the obstacle and the two-wheeled vehicle within the next 0.5 seconds by combining the obstacle movement speed vector; comparing the minimum safe distance with a preset safety threshold of 0.8 meters; if the minimum safe distance is less than 0.8 meters, the safety constraint term is the product of a preset penalty coefficient and the difference in safe distance; otherwise, the safety constraint term is zero.

[0072] For example, a safety margin value is extracted from road boundary information. For instance, the safety margin for the right-hand road boundary is set to 0.2 meters. The position coordinates of the vehicle ahead are obtained from environmental perception data as 15 meters from the two-wheeled vehicle, with a moving speed vector of 1 meter per second towards the two-wheeled vehicle and a moving direction angle of 45 degrees. The current straight-line distance is calculated to be 15 meters, predicting that the vehicle ahead will move 7.5 meters closer within 0.5 seconds, resulting in a minimum safe distance of 7.5 meters. The minimum safe distance of 7.5 meters is compared to a preset safety threshold of 0.8 meters. Since 7.5 is greater than 0.8, the safety constraint term is zero. An important detail is that when the minimum safe distance is less than 0.8 meters, the safety constraint term equals a preset penalty coefficient multiplied by the difference in safe distance. For example, when the difference is 0.1 meters, the penalty coefficient is 10.0, and the safety constraint term is 1.0. This ensures that the system can significantly reduce the score of candidate parameters in dangerous scenarios.

[0073] In the above embodiments, optionally, the calculation of the path tracking error term includes: calculating the lateral deviation between the current two-wheeled vehicle position coordinates and the reference path information in the state information and the environmental perception data; and generating the path tracking error term as the product of a preset weighting coefficient and the square of the lateral deviation.

[0074] For example, the lateral deviation is calculated based on the current coordinates of the two-wheeled vehicle and the reference path information. For instance, if the reference path is a winding mountain road, and the actual position of the two-wheeled vehicle deviates 0.3 meters to the right of the path, the lateral deviation is 0.3 meters. Using a preset weighting coefficient of 0.5, the path tracking error term is calculated as the weighting coefficient multiplied by the square of the lateral deviation, i.e., 0.5 multiplied by the square of 0.3, resulting in 0.045. An important detail is that the lateral deviation is always taken as the shortest distance perpendicular to the reference path direction. The weighting coefficient of 0.5 is used to balance the impact of path tracking accuracy on the objective function, ensuring that the vehicle can smoothly follow complex curves without excessive deviation.

[0075] In the above embodiments, optionally, the calculation of the control smoothness term includes: obtaining the speed adjustment parameter and steering angle parameter in the control parameters of the previous time step; calculating the rate of change of the current speed adjustment parameter with respect to the speed adjustment parameter of the previous time step, and the rate of change of the current steering angle parameter with respect to the steering angle parameter of the previous time step; generating the control smoothness term as the product of a preset weighting coefficient and the sum of the squares of the rate of change of speed adjustment and the rate of change of steering angle.

[0076] For example, the speed adjustment parameter of the previous time step is 1.0 m / s, and the steering angle parameter is 3 degrees. The current candidate parameters are a speed adjustment of 0.8 m / s and a steering angle of 4 degrees. The calculated speed adjustment rate of change is 0.2 m / s, and the steering angle rate of change is 1 degree. The sum of squares of the rates of change is the square of 0.2 plus the square of 1, i.e., 0.04 plus 1 equals 1.04. Using a preset weighting coefficient of 0.1, the smoothness term is 0.1 multiplied by 1.04, resulting in 0.104. An important detail is that the rate of change is calculated based on the absolute difference, the sum of squares ensures that large changes are amplified and penalized, and the weighting coefficient of 0.1 makes the smoothness account for a small proportion in the objective function, avoiding excessive restriction on the flexibility of control commands.

[0077] Step 105: Adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. Jump to the step of obtaining the boundary reference information of the currently passable area based on the status information and environmental perception data, and re-execute the two-wheeled vehicle autonomous driving road condition decision-making process until the driving stopping conditions of the two-wheeled vehicle are met.

[0078] This step sends the optimal control parameters to the two-wheeled vehicle's execution control module to adjust the vehicle's speed, steering, and attitude. Simultaneously, it continuously collects and updates the vehicle's status information and environmental perception data, returning to step 102 to form a closed-loop autonomous driving road condition decision-making process until a preset driving stop condition is detected.

[0079] As an optional embodiment, in step 105, the speed adjustment parameter and steering angle parameter in the optimal control parameters are continuous values. Based on this, adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters includes: checking whether the speed adjustment parameter is within a preset speed adjustment boundary and whether the steering angle parameter is within a preset steering angle boundary; when the speed adjustment parameter exceeds the preset speed adjustment boundary, adjusting the speed adjustment parameter to the nearest value within the preset speed adjustment boundary; when the steering angle parameter exceeds the preset steering angle boundary, adjusting the steering angle parameter to the nearest value within the preset steering angle boundary; generating a corrected control command based on the adjusted speed adjustment parameter and steering angle parameter; and adjusting the motion parameters of the two-wheeled vehicle based on the corrected control command. Further optionally, the continuous values ​​are modeled using a bounded probability distribution to ensure that the output value is within a physically feasible range.

[0080] For example, if the optimal control parameter detects a speed adjustment value of 4.0 m / s, but the preset speed adjustment boundary upper limit is 3.0 m / s, the speed adjustment value will be automatically adjusted to 3.0 m / s since 4.0 exceeds 3.0. This ensures that the nearest valid value within the boundary is always selected, preventing sudden speed changes that could lead to vehicle loss of control. For instance, on a highway, if the output acceleration command is 4.5 m / s, it will be truncated to 3.0 m / s to ensure the vehicle remains within a safe dynamic range while minimizing interference with the current driving state.

[0081] The optimal control parameter for steering angle was detected as 35.0 degrees, but the preset upper limit for steering angle is 30.0 degrees. Since 35.0 degrees exceeds 30.0 degrees, the steering angle is automatically adjusted to 30.0 degrees. The adjustment is based on absolute value comparison to ensure that the steering command conforms to the vehicle's physical limits. For example, on sharp curves, if the output command is a 40-degree steering angle, it is limited to 30 degrees to prevent tire slippage or vehicle rollover, significantly improving stability, especially in rainy road conditions.

[0082] Based on the adjusted speed adjustment parameter of 3.0 meters per second and the steering angle parameter of 30.0 degrees, a revised control command is generated. For example, if the original optimal parameters were a speed adjustment of 4.0 meters per second and a steering angle of 35.0 degrees, the revised command would become a speed adjustment of 3.0 meters per second and a steering angle of 30.0 degrees. The revised command is directly executed by the vehicle, skipping the transmission of invalid parameters and reducing latency. Boundary compliance is verified simultaneously during command generation to ensure that the command can be processed immediately by the vehicle control unit, avoiding execution failures due to invalid parameters.

[0083] The revised control commands are sent to the vehicle's execution module to adjust the two-wheeled vehicle's motion parameters. For example, if the vehicle's current speed is 25.0 meters per second, after applying a speed correction of 3.0 meters per second, the new speed becomes 28.0 meters per second. A steering angle command of 30.0 degrees triggers the vehicle to turn right, and the actual driving trajectory matches the planned path. The adjustment process is achieved through a smooth transition, avoiding abrupt changes in speed or steering. For example, when accelerating from 25.0 meters per second to 28.0 meters per second, acceleration is applied in stages to ensure that passenger comfort is not affected while maintaining safe driving within the road boundaries.

[0084] Further optionally, before adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters, the method further includes: determining the current safety risk index based on the calculation results of the safety constraint terms; comparing the current safety risk index with a preset safety risk threshold; and when the current safety risk index is greater than the preset safety risk threshold, executing a preset safety control strategy to replace and update the optimal control parameters.

[0085] For example, a safety risk index is directly generated based on the calculation results of safety constraints. For instance, a safety constraint score of 0.7 is used as the current safety risk index. A higher safety constraint score indicates a greater safety risk; for example, a score of 0.7 represents high risk when the minimum safe distance is predicted to be 0.6 meters, while a score of 0.2 represents low risk at a distance of 1.5 meters. Without additional calculations, the safety constraint score is used directly as the risk index, ensuring the real-time nature and accuracy of risk assessment and avoiding delayed decisions. Then, the current safety risk index of 0.7 is compared with a preset safety risk threshold of 0.5. Since 0.7 is greater than 0.5, a high safety risk is determined. The threshold is dynamically adjusted according to the traffic scenario; for example, it is set to 0.4 on city streets and 0.6 on highways. The comparison process is completed within 0.05 seconds, ensuring immediate response to sudden risks and avoiding misjudgments due to fixed thresholds, such as quickly identifying high risk when a pedestrian suddenly crosses the road. When the safety risk index of 0.7 exceeds the threshold of 0.5, the preset safety control strategy is immediately executed. For example, if the original optimal control parameters were speed adjustment +1.0 m / s and steering angle 0 degrees, they can be replaced with speed adjustment -2.0 m / s and steering angle -10 degrees. The safety strategy has the highest priority and directly overrides the original optimal parameters. The replaced parameters automatically pass boundary checks; for example, the speed adjustment of -2.0 m / s falls within the preset boundary of -3.0 to 3.0 m / s. After executing the strategy, the objective function is recalculated to ensure that the new parameters meet safety requirements and to prevent loss of vehicle control due to strategy execution. For example, during emergency braking, the vehicle will decelerate smoothly rather than abruptly.

[0086] Further, optionally, the optimization step of the objective function can also be based on cycling demonstration data, comparing the control parameters generated by the objective function with the demonstration control parameters; by adjusting the preset weight coefficients, the sum of the differences in the control parameter values ​​of all demonstration samples is minimized.

[0087] For example, a behavioral cloning loss function is constructed based on cycling demonstration data; the deviation between the control parameters generated by the objective function and the cycling demonstration control parameters is calculated using the behavioral cloning loss function; and the objective function is fine-tuned and optimized by minimizing the deviation calculated by the behavioral cloning loss function.

[0088] In practical applications, the behavioral cloning loss function is a core component of imitation learning, used to measure the difference between the control parameters generated by the autonomous driving system and expert riding demonstration data. The behavioral cloning loss function is directly constructed based on expert demonstration data, transforming the problem into a supervised learning task. Five hundred riding demonstration samples were collected, each containing speed adjustment parameters and steering angle parameters for a two-wheeled vehicle under specific traffic conditions. For example, one sample in the demonstration data shows that when driving on a curve, the expert selected a speed adjustment value of 1.0 m / s and a steering angle of 10 degrees.

[0089] The specific calculation process of the behavioral cloning loss function is as follows: For each demonstration sample, the difference between the control parameters generated by the objective function and the demonstration control parameters is calculated. For example, the objective function outputs a speed adjustment value of 1.5 m / s and a steering angle of 8 degrees under a specific state, while the corresponding values ​​in the demonstration data are 1.0 m / s and 10 degrees. Subtracting 1.0 from the speed adjustment difference (1.5) equals 0.5 m / s, and subtracting 8 from the steering angle difference (10) equals 2 degrees, the absolute values ​​of these differences are then summed to obtain the deviation value for that sample. The deviation values ​​of all 500 samples are accumulated to obtain the total deviation value.

[0090] The behavioral cloning loss function is constructed by minimizing the sum of the differences in control parameters across all demonstration samples. Initially, the preset weight coefficients are set to 0.6 for safety, 0.3 for path, and 0.1 for smoothness, resulting in a total bias of 25.0 for all demonstration samples. By adjusting the weight coefficients, for example, fine-tuning the safety weight to 0.55, the path weight to 0.35, and the smoothness weight to 0.1, the bias of all samples is recalculated, reducing the total bias to 22.5. Iterative optimization continues until the total bias changes to less than 0.1, ultimately determining the weight coefficients to be 0.52 for safety, 0.33 for path, and 0.15 for smoothness, stabilizing the total bias at 20.8.

[0091] The fine-tuning of the objective function is driven by the calculation results of the behavior cloning loss function. When the control parameters in certain scenarios differ significantly from the demonstration data, the weight coefficients related to these scenarios are adjusted first. For example, in a cornering scenario, if the steering angle generated by the objective function is found to be generally 2 degrees smaller than that of the demonstration data, the weight coefficient of the path tracking error term will be automatically increased to make the objective function more inclined to match the steering angle of the demonstration data.

[0092] In practical applications, optimizing the behavioral cloning loss function makes autonomous driving systems more closely resemble human riding habits. For example, in scenarios involving continuous curves, fine-tuning and optimization increased the generated steering angle from an average of 8 degrees to 10 degrees, better matching the demonstration data. Simultaneously, it ensures all control parameters remain within physically feasible limits, preventing the generation of commands exceeding the vehicle's dynamic capabilities in pursuit of matching demonstration data. The optimization process of the behavioral cloning loss function does not sacrifice safety constraints, always ensuring that safety constraints have the highest weight. For example, on highways, the safety weight is maintained above 0.5, the path-following weight is between 0.3 and 0.4, and the smoothing weight is between 0.1 and 0.2. This weighting allocation mimics human riding habits as closely as possible while ensuring safety. When high-risk situations appear in the demonstration data, the safety weight is automatically increased, ensuring that safety is always the primary consideration. Through the optimization of the behavioral cloning loss function, autonomous driving systems can better adapt to different riding styles and driving habits, improving the naturalness and comfort of driving. The optimization process considers not only numerical differences but also traffic scenario characteristics. For example, when optimizing for urban streets, greater emphasis is placed on path tracking accuracy, while on highways, safety constraints are prioritized. This scenario-aware optimization enables dynamic adjustment of weighting coefficients based on the actual traffic environment, achieving more intelligent autonomous driving decisions.

[0093] In this embodiment, a road condition decision-making model for autonomous driving of two-wheeled vehicles is constructed based on a near-end policy optimization algorithm. Boundary reference information of passable areas is introduced during the decision-making process to constrain and optimize the action probability distribution. In complex and dynamic outdoor traffic environments, optimal control parameters that simultaneously meet the requirements of safety, path tracking accuracy, and control smoothness can be obtained. This reduces the risk of decision instability and collision probability caused by environmental uncertainty and dynamic obstacle changes, bringing users a safer, smoother, and more reliable autonomous driving experience. It also reduces the problem of frequent manual intervention required in complex road conditions by rule-based or single optimization models in related technologies.

[0094] The above describes a two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization in the embodiments of this application. The following describes the two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization that executes the above-described two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization.

[0095] See Figure 2 ,like Figure 2 The diagram shows a structural schematic of a two-wheeled vehicle autonomous driving road condition decision-making device 60 based on near-end strategy optimization. The two-wheeled vehicle operates in an outdoor traffic environment containing dynamic obstacles and road boundaries. The two-wheeled vehicle autonomous driving road condition decision-making device 60 based on near-end strategy optimization in this embodiment can achieve the above-mentioned... Figure 1 The steps of the two-wheeled vehicle autonomous driving road condition decision-making method based on near-end policy optimization are executed in the corresponding embodiments. The functions implemented by the two-wheeled vehicle autonomous driving road condition decision-making device 60 based on near-end policy optimization can be implemented by hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above functions, and the modules can be software and / or hardware. The device 60 includes the following modules:

[0096] The acquisition module 601 is configured to acquire the status information and environmental perception data of the two-wheeled vehicle; and based on the status information and environmental perception data, acquire the boundary reference information of the currently passable area.

[0097] The decision module 602 calculates the action probability distribution of the two-wheeled vehicle under the current traffic state based on the boundary reference information, state information, and environmental perception data; and obtains the optimal control parameters of the two-wheeled vehicle by sampling from the action probability distribution through optimizing the objective function; wherein, the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment;

[0098] The control module 603 is configured to adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and to collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. It then jumps to the step executed by the decision module to obtain the boundary reference information of the currently passable area based on the status information and environmental perception data, so as to re-execute the two-wheeled vehicle autonomous driving road condition decision process until the driving stopping conditions of the two-wheeled vehicle are met.

[0099] In some implementations, the state information includes the speed, acceleration, tilt angle, and current position coordinates of the two-wheeled vehicle, while the environmental perception data includes the position, distance, direction of movement, and speed of surrounding obstacles.

[0100] In some implementations, the acquisition module 601, based on state information and environmental perception data, acquires boundary reference information of the currently passable area, and is configured to: acquire the coordinate values ​​of multiple boundary reference points and the safety margin corresponding to each boundary reference point; wherein each boundary reference point includes at least the boundary reference point closest to the left boundary of the road, the boundary reference point closest to the right boundary of the road, and the boundary reference point with the largest safety margin.

[0101] In some implementations, the decision module 602, based on the boundary reference information, state information, and environmental perception data, calculates the action probability distribution corresponding to the two-wheeled vehicle under the current traffic condition. This is configured to: calculate a safe distance index for the currently passable area based on the safety margin in the boundary reference information; calculate a current path deviation index based on the two-wheeled vehicle position coordinates in the state information and obstacle dynamic information in the environmental perception data; fuse the safe distance index and the path deviation index to generate constraint weight coefficients; combine the constraint weight coefficients with the speed and acceleration parameters in the state information through a preset constraint mapping relationship to obtain a parameter vector for the action probability distribution; and generate a physically feasible action probability distribution based on the parameter vector, wherein the output space of the action probability distribution is constrained within preset boundaries of the speed adjustment range and the steering angle range.

[0102] In some implementations, the decision module 602, by optimizing the objective function, samples the optimal control parameters of the two-wheeled vehicle from the action probability distribution. This is configured to: randomly sample multiple candidate control parameters from the action probability distribution, including speed adjustment parameters and steering angle parameters; wherein the value range of the speed adjustment parameters is constrained within a preset speed adjustment boundary, and the value range of the steering angle parameters is constrained within a preset steering angle boundary; for each candidate control parameter, calculate a safety constraint term, a path tracking error term, and a control smoothness term; calculate a weighted sum of the safety constraint term, path tracking error term, and control smoothness term based on preset weight coefficients, and use the weighted sum as the objective function value; select the candidate control parameter with the largest objective function value as the optimal control parameter.

[0103] In some implementations, the decision module 602 is configured to calculate a weighted sum of the safety constraint term, the path tracking error term, and the control smoothness term based on preset weight coefficients, by performing a linear weighted sum of the safety constraint term, the path tracking error term, and the control smoothness term using preset weight coefficients; wherein the preset weight coefficients are obtained through multiple rounds of interactive learning and iterative optimization in a simulation environment, and the weight coefficient of the safety constraint term is greater than the weight coefficient of the path tracking error term, and the weight coefficient of the path tracking error term is greater than the weight coefficient of the control smoothness term.

[0104] In some implementations, the speed adjustment parameter and steering angle parameter in the optimal control parameters are continuous values; the control module 603, which adjusts the motion parameters of the two-wheeled vehicle according to the optimal control parameters, is configured to: check whether the speed adjustment parameter is within a preset speed adjustment boundary and whether the steering angle parameter is within a preset steering angle boundary; when the speed adjustment parameter exceeds the preset speed adjustment boundary, adjust the speed adjustment parameter to the nearest value within the preset speed adjustment boundary; when the steering angle parameter exceeds the preset steering angle boundary, adjust the steering angle parameter to the nearest value within the preset steering angle boundary; generate a corrected control command based on the adjusted speed adjustment parameter and steering angle parameter; and adjust the motion parameters of the two-wheeled vehicle based on the corrected control command.

[0105] In some implementations, before adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters, the control module 603 is further configured to: determine the current safety risk index based on the calculation results of the safety constraint terms; compare the current safety risk index with a preset safety risk threshold; and when the current safety risk index is greater than the preset safety risk threshold, execute a preset safety control strategy to replace and update the optimal control parameters.

[0106] In some implementations, the decision module 602 is further configured to: compare the control parameters generated by the objective function with the demonstration control parameters based on the cycling demonstration data; and minimize the sum of the differences in the control parameter values ​​of all demonstration samples by adjusting the preset weight coefficients.

[0107] In this embodiment, the two-wheeled vehicle autonomous driving road condition decision-making device 60 based on near-end strategy optimization constructs a two-wheeled vehicle autonomous driving road condition decision-making model based on a near-end strategy optimization algorithm. In the decision-making process, boundary reference information of the passable area is introduced to constrain and optimize the action probability distribution. In complex and dynamic outdoor traffic environments, the optimal control parameters that simultaneously meet the requirements of safety, path tracking accuracy, and control smoothness can be obtained. This reduces the risk of decision instability and collision probability caused by environmental uncertainty and dynamic obstacle changes, bringing users a safer, smoother, and more reliable autonomous driving experience. It also reduces the problem of frequent manual intervention required in complex road conditions by rule-based or single optimization models in related technologies.

[0108] The above describes the two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization in the embodiments of this application from the perspective of modular functional entities. The following describes the two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization in the embodiments of this application from the perspective of hardware processing.

[0109] This application also provides a terminal device, such as... Figure 3As shown, for ease of explanation, only the parts related to the embodiments of this application are shown. For specific technical details not disclosed, please refer to the method section of the embodiments of this application. The terminal device can be any terminal device including mobile phones, tablets, personal digital assistants (PDAs), point of sales (POS) devices, in-vehicle computers, etc. These devices can be connected to the two-wheeled vehicle, thereby realizing the two-wheeled vehicle through the technical method provided in the embodiments of this application. The two-wheeled vehicle operates in an outdoor traffic environment containing dynamic obstacles and road boundaries. Taking a mobile phone as an example, the mobile phone can connect to the two-wheeled vehicle to obtain its status information and environmental perception data. Based on the status information and environmental perception data, it obtains the boundary reference information of the currently passable area. Based on the boundary reference information, status information, and environmental perception data, it calculates the probability distribution of the two-wheeled vehicle's actions under the current traffic conditions. By optimizing the objective function, it samples the optimal control parameters of the two-wheeled vehicle from the probability distribution of actions. The parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term. The parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle's behavior under various traffic scenarios in a simulated environment. The motion parameters of the two-wheeled vehicle are adjusted according to the optimal control parameters, and the status information and environmental perception data of the two-wheeled vehicle are collected and updated in real time. The process then jumps to the step of obtaining the boundary reference information of the currently passable area based on the status information and environmental perception data, and the two-wheeled vehicle autonomous driving road condition decision-making process is re-executed until the driving stopping conditions of the two-wheeled vehicle are met.

[0110] Figure 3 The diagram shown is a structural schematic of the terminal device provided in the embodiments of this application. (Reference) Figure 3 The terminal device includes components such as a radio frequency (RF) circuit 1010, a memory 1020, an input unit 1030, a display unit 1040, a sensor 1050, an audio circuit 1060, a wireless fidelity (WiFi) module 1070, a processor 1080, and a power supply 1090. The input unit 1030 includes a touch panel 1031 and other input devices 1032. The display unit 1040 includes a display panel 1041. The audio circuit 1060 includes a speaker 1061 and a microphone 1062. Those skilled in the art will understand that... Figure 3 The two-wheeled vehicle control module shown does not constitute a limitation on the mobile phone and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. It should be noted that... Figure 2 The devices 60 shown can all have the following characteristics: Figure 3 The structure shown, when Figure 2The two-wheeled vehicle autonomous driving road condition decision-making device 60 shown, based on near-end strategy optimization, has the following characteristics: Figure 3 When the structure shown is used, Figure 3 The processor in the device can perform the same or similar functions as the decision module 602 and control module 603 provided in the aforementioned device embodiments, and the input unit 1030 can perform the same or similar functions as the acquisition module 601 provided in the aforementioned device embodiments. Figure 3 The memory stores the computer programs that the processor needs to call when executing the above-mentioned two-wheeled vehicle autonomous driving road condition decision-making method based on near-end policy optimization.

[0111] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] In the embodiments of this application, the functional modules can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium.

[0114] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0115] The computer program product includes one or more computer instructions. When the computer program is loaded and executed on a computer, it generates, in whole or in part, the processes or functions described in the embodiments of this application. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a server or data center that integrates one or more available media. The available medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium, or a semiconductor medium, etc.

[0116] The technical solutions provided in the embodiments of this application have been described in detail above. Specific examples have been used in the embodiments of this application to illustrate the principles and implementation methods of the embodiments of this application. The description of the above embodiments is only for the purpose of helping to understand the methods and core ideas of the embodiments of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the embodiments of this application. Therefore, the content of this specification should not be construed as a limitation on the embodiments of this application.

Claims

1. A two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization, characterized in that, The method involves operating a two-wheeled vehicle in an outdoor traffic environment that includes dynamic obstacles and road boundaries. Acquire the status information and environmental perception data of the two-wheeled vehicle; Based on state information and environmental perception data, obtain boundary reference information of the currently passable area; Based on the boundary reference information, state information, and environmental perception data, the motion probability distribution of the two-wheeled vehicle under the current traffic condition is calculated, including: calculating a safe distance index for the currently passable area based on the safety margin in the boundary reference information; calculating a current path deviation index based on the two-wheeled vehicle position coordinates in the state information and obstacle dynamic information in the environmental perception data; fusing the safe distance index and the path deviation index to generate constraint weight coefficients; combining the constraint weight coefficients with the speed and acceleration parameters in the state information through a preset constraint mapping relationship to obtain a parameter vector for the motion probability distribution; and generating a motion probability distribution that satisfies physical feasibility based on the parameter vector, wherein the output space of the motion probability distribution is constrained within preset boundaries of the speed adjustment range and the steering angle range. The optimal control parameters of the two-wheeled vehicle are obtained by sampling from the action probability distribution by optimizing the objective function; wherein, the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment; Adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. Jump to the step of obtaining the boundary reference information of the currently passable area based on the status information and environmental perception data, and re-execute the two-wheeled vehicle autonomous driving road condition decision-making process until the driving stopping conditions of the two-wheeled vehicle are met.

2. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 1, characterized in that, Status information includes the speed, acceleration, tilt angle, and current position coordinates of the two-wheeled vehicle, while environmental perception data includes the position, distance, direction of movement, and speed of surrounding obstacles.

3. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 2, characterized in that, The process of obtaining boundary reference information for the currently passable area based on state information and environmental perception data includes: Obtain the coordinates of multiple boundary reference points and the safety margin corresponding to each boundary reference point; Each boundary reference point includes at least the boundary reference point closest to the left boundary of the road, the boundary reference point closest to the right boundary of the road, and the boundary reference point with the largest safety margin.

4. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 1, characterized in that, The step of sampling the optimal control parameters of the two-wheeled vehicle from the action probability distribution by optimizing the objective function includes: Multiple candidate control parameters are randomly sampled from the action probability distribution. The candidate control parameters include speed adjustment parameters and steering angle parameters. The value range of the speed adjustment parameters is constrained within a preset speed adjustment boundary, and the value range of the steering angle parameters is constrained within a preset steering angle boundary. For each candidate control parameter, calculate the safety constraint term, path tracking error term, and control smoothness term respectively; The weighted sum of the safety constraint term, path tracking error term, and control smoothness term is calculated based on preset weight coefficients, and the weighted sum is used as the objective function value. The candidate control parameter with the largest objective function value is selected as the optimal control parameter.

5. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 4, characterized in that, The calculation of the weighted sum of the safety constraint term, path tracking error term, and control smoothness term based on preset weight coefficients includes: The safety constraint term, path tracking error term, and control smoothness term are linearly weighted and summed using preset weighting coefficients; The preset weight coefficients are obtained through multiple rounds of interactive learning and iterative optimization in the simulation environment. The weight coefficient of the safety constraint term is greater than the weight coefficient of the path tracking error term, and the weight coefficient of the path tracking error term is greater than the weight coefficient of the control smoothness term.

6. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 1, characterized in that, The speed adjustment parameter and steering angle parameter in the optimal control parameters are continuous values; adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters includes: Check whether the speed adjustment parameters are within the preset speed adjustment boundaries and whether the steering angle parameters are within the preset steering angle boundaries; When the speed adjustment parameter exceeds the preset speed adjustment boundary, adjust the speed adjustment parameter to the nearest value within the preset speed adjustment boundary; When the steering angle parameter exceeds the preset steering angle boundary, the steering angle parameter will be adjusted to the nearest value within the preset steering angle boundary. Based on the adjusted speed adjustment parameters and steering angle parameters, a corrected control command is generated. The motion parameters of the two-wheeled vehicle are adjusted based on the revised control commands.

7. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 1, characterized in that, Before adjusting the motion parameters of the two-wheeled vehicle according to the optimal control parameters, the method further includes: Based on the calculation results of the security constraints, the current security risk index is determined; Compare the current security risk index with the preset security risk threshold; When the current security risk index is greater than the preset security risk threshold, the preset security control strategy is executed, and the optimal control parameters are replaced and updated.

8. The two-wheeled vehicle autonomous driving road condition decision-making method based on near-end strategy optimization according to claim 1, characterized in that, The optimization steps of the objective function further include: Based on the cycling demonstration data, the control parameters generated by the objective function are numerically compared with the demonstration control parameters; By adjusting the preset weighting coefficients, the sum of the differences in the control parameter values ​​of all demonstration samples is minimized.

9. A two-wheeled vehicle autonomous driving road condition decision-making device based on near-end strategy optimization, characterized in that, The two-wheeled vehicle operates in an outdoor traffic environment that includes dynamic obstacles and road boundaries, and the device includes: The acquisition module is configured to acquire the status information and environmental perception data of the two-wheeled vehicle; and based on the status information and environmental perception data, acquire the boundary reference information of the currently passable area. The decision module is configured to calculate the probability distribution of actions of the two-wheeled vehicle under the current traffic state based on the boundary reference information, state information, and environmental perception data; and to sample the optimal control parameters of the two-wheeled vehicle from the probability distribution by optimizing the objective function; wherein the parameters of the objective function include a safety constraint term, a path tracking error term, and a control smoothness term, and the parameters of the objective function are obtained by multi-round interactive learning and iterative optimization of the two-wheeled vehicle behavior under various traffic scenarios in a simulated environment; The decision-making module, when calculating the action probability distribution of the two-wheeled vehicle under the current traffic condition based on the boundary reference information, state information, and environmental perception data, is configured to: calculate the safe distance index of the currently passable area based on the safety margin in the boundary reference information; calculate the current path deviation index based on the two-wheeled vehicle position coordinates in the state information and the obstacle dynamic information in the environmental perception data; fuse the safe distance index and the path deviation index to generate constraint weight coefficients; combine the constraint weight coefficients with the speed and acceleration parameters in the state information through a preset constraint mapping relationship to obtain the parameter vector of the action probability distribution; and generate an action probability distribution that satisfies physical feasibility based on the parameter vector, wherein the output space of the action probability distribution is constrained within preset boundaries of the speed adjustment range and the steering angle range. The control module is configured to adjust the motion parameters of the two-wheeled vehicle according to the optimal control parameters, and collect and update the status information and environmental perception data of the two-wheeled vehicle in real time. It then jumps to the step executed by the decision module to obtain the boundary reference information of the currently passable area based on the status information and environmental perception data, so as to re-execute the two-wheeled vehicle autonomous driving road condition decision process until the driving stopping conditions of the two-wheeled vehicle are met.

Images