Dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning

Abstract

The invention is suitable for the technical field of automatic driving, and provides a dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning, and the method comprises the steps: dividing obstacles into three classes: pedestrians, vehicles and static obstacles through a target detection algorithm, and determining the risk distribution characteristics of each class; secondly, a 11-dimensional high-dimensional space-time state space is constructed, and risk pre-judgment is achieved; a PPO reinforcement learning algorithm is introduced, a high-dimensional state is used as input, four-dimensional potential field deformation parameters for three types of obstacles are output, and a differentiated continuous potential field form is autonomously learned; constructing a covariance matrix based on Riemannian geometry, generating a non-abrupt-change anisotropic repulsion field, and superposing a virtual escape force to crack a local minimum value; and synthesizing the gravitational force, the total repulsive force and the escape force to obtain a total potential field force, and outputting a steering angle and acceleration instruction conforming to vehicle dynamics constraints. According to the strategy, a complete obstacle avoidance framework is constructed, and the technical limitation of a traditional scheme is effectively relieved.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving technology, and in particular relates to a dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning. Background Technology

[0002] With the rapid development of autonomous driving technology, obstacle avoidance safety of intelligent vehicles in complex environments such as urban roads and residential areas has become a core requirement. Urban road environments contain various interfering targets, including pedestrians, other vehicles, and static obstacles, and the motion characteristics and collision risks of different targets vary significantly.

[0003] Artificial potential field methods have become one of the mainstream local obstacle avoidance algorithms due to their high computational efficiency and real-time performance. The traditional model guides the vehicle towards the target point using a gravitational field and avoids collisions with obstacles using a repulsive field. However, traditional repulsive field designs employ a distance-dependent homogenization model, meaning that the repulsive force around the same obstacle in different directions is only related to distance, not direction. This model has two major drawbacks: first, it cannot effectively distinguish between the directional risks of pedestrians and vehicles; second, the repulsive force coefficient needs to be manually preset and cannot be dynamically adjusted according to the real-time environment, easily leading to local minima or target unreachability problems in complex multi-objective scenarios.

[0004] Reinforcement learning technology can learn optimal decision-making strategies through environmental interaction. Among them, proximal policy optimization algorithms are widely used in autonomous driving decision optimization due to their support for continuous action spaces and strong training stability. However, existing RL+APF fusion solutions either only adjust and optimize the overall potential field parameters through RL without designing differentiated repulsive fields for the directional risks of pedestrians and vehicles; or lack dedicated local minima-breaking mechanisms; and most solutions only design optimization strategies for specific scenarios, lacking the ability to adapt to general and complex scenarios such as continuous switching between straight and curved roads in urban areas and the coexistence of pedestrians, vehicles, and static obstacles. This results in the solution needing to be adjusted separately according to the scenario, making it difficult to meet the obstacle avoidance requirements of high-level autonomous driving covering all scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning, which aims to solve the technical problems existing in the prior art as identified in the background art.

[0006] This invention is implemented as follows: a dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning, the method comprising:

[0007] S1. Multi-source data fusion is performed using lidar, millimeter-wave radar, and cameras to collect real-time information on intelligent vehicle status, obstacle information, and risk characteristic parameters.

[0008] S2. Based on real-time collected intelligent vehicle status, obstacle information, and risk characteristic parameters, construct an 11-dimensional high-dimensional spatiotemporal state vector. ;

[0009] S3, convert the state vector The input is fed into a pre-trained PPO policy network, which outputs 4D Riemannian fluid field deformation parameters for different obstacle types.

[0010] S4. Construct a dynamic Riemannian fluid field based on the deformation parameters of the 4D Riemannian fluid field and the basic parameters corresponding to the obstacle types.

[0011] S5. Establish a gravitational potential field Calculate the gravitational force that guides the vehicle toward the target point. And virtual escape forces used to break local minimum deadlocks. The direction of the virtual escape force is perpendicular to the direction of gravity;

[0012] S6. Repeat the calculation of repulsive forces for all obstacles in the environment, and perform vector superposition to obtain the total repulsive force. The total repulsive force ,gravitational and virtual escape force Vector synthesis is performed to obtain the total potential force. ;

[0013] S7. Based on the total potential force Based on vehicle dynamics constraints, the steering angle of the intelligent vehicle within the current control cycle is calculated and output. and acceleration The command is given, and the vehicle is driven to perform obstacle avoidance maneuvers.

[0014] The beneficial effects of this invention are:

[0015] By constructing a dynamic flow potential field based on the Riemannian metric tensor, a repulsive potential field is realized that is continuous and differentiable in the whole space and has anisotropic shape. It can accurately match the differentiated collision risk distribution of pedestrians, vehicles and static obstacles, and effectively avoid the homogenization repulsive force problem and turning oscillation of traditional potential fields.

[0016] By introducing PPO reinforcement learning as the core of decision-making, data-driven and adaptive adjustment of potential field deformation parameters is realized, replacing the traditional manual preset or fuzzy adjustment mode, and significantly improving the system's adaptability and generalization in dynamic multi-objective environments.

[0017] A high-dimensional spatiotemporal state space was designed, incorporating time-dimensional features such as distance change rate and collision urgency. This enabled advanced perception and prediction of dynamic collision risks, solving the problem of delayed response in traditional solutions.

[0018] A virtual escape force perpendicular to the direction of gravity was designed as an endogenous mechanism, which can directly and effectively solve the "local minimum deadlock" problem in the traditional artificial potential field method from the perspective of mechanical structure, and ensure the vehicle's ability to get out of trouble in complex scenarios such as narrow road sections.

[0019] By configuring differentiated basic parameters and deformation parameters for different obstacle types, the solution achieves full coverage and precise protection for multiple types of obstacles. Moreover, the solution does not depend on specific scenario parameters and can be seamlessly adapted to various common and complex scenarios on urban roads, making it highly practical and versatile. Attached Figure Description

[0020] Figure 1 An overall obstacle avoidance flowchart provided for embodiments of the present invention;

[0021] Figure 2 This is a flowchart of the PPO reinforcement learning training sub-process provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] like Figure 1 As shown, a dynamic Riemannian manifold artificial potential field obstacle avoidance method based on PPO reinforcement learning is described, the method comprising:

[0024] S1: Environmental perception and status information acquisition

[0025] By fusing multi-source data—LiDAR (to acquire the 3D position of obstacles), millimeter-wave radar (to measure relative velocity in real time), and cameras (to identify obstacle types based on target detection algorithms)—the problem of blind spots in single-sensor perception is solved, and the following core information is output in real time:

[0026] Intelligent vehicle status: including current location (Horizontal and vertical coordinates in the global coordinate system) Longitudinal driving speed lateral acceleration yaw rate and target location (Given by the global path planning module).

[0027] Obstacle information: including obstacle type ( =1 represents a pedestrian. =2 represents social vehicles, =3 represents static obstacles, and real-time location. (in global coordinate system) ), obstacle movement speed (Dynamic obstacles are represented by actual driving speed, static obstacles by speed of 0), and the straight-line distance between the vehicle and the obstacle. (Core safety assessment criteria) and distance change rate (Reflects the trend of approaching / moving away from obstacles).

[0028] Risk characteristic parameter: Collision urgency (The reciprocal of the collision time quantifies the urgency of the risk, quickly distinguishing between routine and emergency scenarios), local obstacle density. (Number of obstacles within a 10m radius centered on the vehicle, normalized to) (Different strategies for adapting to crowded and open spaces).

[0029] in:

[0030] Straight-line distance between vehicle and obstacle Euclidean distance between the vehicle and the obstacle (unit: ), is the most basic safety assessment indicator, and its calculation formula is:

[0031] ;

[0032] in, Let x and y be the x and y coordinates of the vehicle in the global coordinate system. The coordinates of the obstacle are the horizontal and vertical coordinates, which directly reflect the spatial distance between the vehicle and the obstacle.

[0033] Rate of change of distance (unit: Its core function is to predict risk trends, and the calculation formula is:

[0034] ;

[0035] in, Let X represent the components of the vehicle's velocity along the X and Y axes in the global coordinate system. This represents the component of the obstacle's velocity; This indicates that an obstacle is approaching the vehicle and a rapid response is required. This indicates that the obstacle is away and the normal driving strategy can be maintained.

[0036] The angular deviation between the vehicle's heading and the target point's direction (unit: This reflects the degree of deviation between the vehicle's current direction of travel and the target direction, ensuring that the vehicle does not deviate from the target path while avoiding obstacles. The calculation formula is:

[0037] ;

[0038] in, The x and y coordinates of the target point, The direction angle of the target point. This is the vehicle's current heading angle. The greater the deviation, the greater the steering angle that needs to be adjusted.

[0039] The longitudinal speed of the vehicle (unit: The obstacle avoidance action is directly collected by the vehicle's speed sensor. When driving at high speed, the obstacle avoidance action needs to be triggered earlier (due to the longer braking distance), while at low speed, it can be appropriately delayed to adapt to the obstacle avoidance needs of different vehicle speeds.

[0040] The lateral acceleration of the vehicle (unit: The calculation formula is: Used to constrain steering strength—exceeding the vehicle's maximum lateral acceleration (typically 1000 rpm). This can easily lead to rollover, so it needs to be included in the state vector for strategy optimization.

[0041] The yaw rate of the vehicle (unit: The data is directly collected by the gyroscope to quantify the speed of vehicle steering, avoiding excessive steering that could cause trajectory oscillation and improving driving comfort.

[0042] The speed of the obstacle (unit: The calculation formula is: Static obstacles The higher the speed of the dynamic obstacle, the higher the risk of collision, and the more conservative the strategy needs to be.

[0043] The relative angle between the vehicle's velocity vector and the obstacle's velocity vector (unit: Its core function is to "determine the movement intention of obstacles," and the calculation formula is:

[0044] ;

[0045] in, To prevent extremely small constants with a denominator of zero; Approaching 0° indicates traveling in the same direction, approaching 180° indicates traveling in opposite directions, and approaching 90° indicates crossing the road. Different intentions correspond to different obstacle avoidance strategies.

[0046] Collision urgency (unit: The value is the reciprocal of the collision time (TTC), which addresses the instability in network training caused by the excessively large value of the traditional TTC when the obstacle is far away. The calculation formula is as follows:

[0047] ;

[0048] The higher the risk (e.g., an obstacle is rapidly approaching). The higher the value, the faster the emergency obstacle avoidance strategy will be triggered; when the obstacle moves away, Without adding additional risk weights, the network is focused on key risks.

[0049] S2: Constructing a high-dimensional spatiotemporal state vector

[0050] To enable dynamic risk prediction in advance, the information collected in the first step is constructed into an 11-dimensional state vector. As input to the PPO network:

[0051] ;

[0052] This vector comprehensively covers four types of information: static distance, dynamic trend, vehicle status, and risk quantification.

[0053] S3: PPO Strategy Network Decision and Deformation Parameter Output

[0054] like Figure 2 As shown, the state vector Input is fed into a pre-trained PPO policy network. The network output is specific to the current obstacle type. 4D action vector That is, the deformation parameters of the Riemannian flow field:

[0055] ;

[0056] in, It is the longitudinal expansion coefficient, which controls the extension of the potential field along the principal axis of risk; The horizontal contraction coefficient controls the width in the vertical direction; This is the rotational bias angle, used to fine-tune the direction of the principal axis of the potential field; The escape disturbance factor determines the magnitude of the virtual escape force. For pedestrians, vehicles, and static obstacles, each parameter has a pre-defined reasonable value range.

[0057] Action vectors For a continuous space, four dimensions of Riemannian flow potential field deformation parameters are output for different obstacle types. The value range and physical meaning of each parameter are strictly matched with the obstacle risk characteristics to ensure that the potential field deformation is adjusted as needed, as detailed below:

[0058] For pedestrians ( Action vector The constraints have been verified through extensive simulations, balancing safety and flexibility.

[0059] (Longitudinal expansion coefficient, which controls the extension length of the potential field along the direction of pedestrian movement to avoid excessive stretching that leads to obstacle avoidance redundancy.)

[0060] (The lateral contraction coefficient controls the width of the potential field perpendicular to the direction of pedestrian movement, adapting to the narrow body shape of pedestrians and reducing the waste of lateral obstacle avoidance space.)

[0061] (Rotation offset angle, range ±15°, to avoid large offsets that may lead to missed risk assessments, and to adapt to the randomness of pedestrian movement).

[0062] (Escape force amplitude: pedestrian scenarios have low risk, so the escape force does not need to be too large to avoid interfering with the normal obstacle avoidance trajectory).

[0063] For vehicles ( Action vector The constraint range is adapted to the large size and high-risk characteristics of vehicles:

[0064] (The risk range in the vehicle's driving direction is large, the upper limit of the longitudinal expansion coefficient is higher, and the risk ahead can be perceived earlier.)

[0065] 0 (The vehicle is wide, and the lower limit of the lateral contraction coefficient is lower, ensuring sufficient lateral obstacle avoidance space and avoiding scrapes).

[0066] (Rotation offset angle range ±22.5°, adaptable to large-angle movements such as lane changes and turns, accurately matching the driving direction);

[0067] (Vehicle scenarios are high-risk, with greater escape force amplitude, quickly breaking deadlock scenarios such as "two vehicles sandwiching".)

[0068] For static obstacles ( Action vector The constraint range is adapted to the characteristics of static obstacles with no movement:

[0069] (Static obstacles have no movement and a narrow range of longitudinal expansion coefficients, avoiding over-perception).

[0070] (The potential field is nearly circular, and the lateral contraction coefficient is close to 1, which is suitable for the characteristics of uniform omnidirectional risk of static obstacles.)

[0071] (Static obstacles have no direction of motion, and the principal axis of the potential field does not need to be deflected, simplifying calculations while ensuring risk coverage.)

[0072] (Static obstacle scenarios have a low risk of deadlock, minimal escape force amplitude, and avoid meaningless trajectory disturbances.)

[0073] in, Its core function is to control the longitudinal extension of the potential field (along the direction of movement of the obstacle). By controlling the lateral contraction (perpendicular to the direction of obstacle movement), the two work together to achieve anisotropy (non-circular) of the potential field, accurately matching the risk distribution of obstacles; Adjust the direction of the principal axis of the potential field to ensure that the repulsive force focuses on protecting the high-risk direction; The magnitude of the virtual escape force is determined to specifically address the local minimum deadlock problem.

[0074] Design a reward function:

[0075] ;

[0076] Security rewards (Initial weight 0.5, highest priority)

[0077] The core objective is to avoid collisions. Rewards and penalties are differentiated based on whether the distance between the vehicle and obstacles meets a safety threshold, thereby strengthening network security awareness.

[0078] ;

[0079] in, Minimum safe distance for different obstacles (pedestrians) ,vehicle Static obstacles ), with reference to road traffic safety standards and vehicle braking performance settings; when When this occurs, a strong penalty (-200) is triggered, making it clear to the network that collision risk is absolutely prohibited; when At that time, the reward increases with the distance (the exponential function ensures that the reward is low at close range and high at long range), guiding the vehicle to actively move away from the obstacle, not just to the point of safety.

[0080] Deformation rationality reward (Initial weight 0.3, second priority)

[0081] The core objective is to avoid meaningless distortion of the potential field, constrain deformation parameters within a reasonable range, ensure that the potential field shape matches the obstacle type, and avoid wasting effort.

[0082] pedestrian: By using the difference in coefficient weights, the vertical expansion coefficient is constrained first to avoid excessive stretching of the potential field in pedestrian scenarios;

[0083] vehicle: To adapt to the high-risk characteristics of vehicles, the focus is on limiting the upper limit of longitudinal expansion while ensuring sufficient lateral contraction.

[0084] Static obstacles: The potential field is constrained to be nearly circular to avoid excessive deformation of the potential field in static obstacle scenarios (the weights are low, so excessive constraints are not required).

[0085] When the parameters are within a reasonable range A positive value encourages the network to adopt reasonable deformation; when the parameter exceeds the range, A negative value indicates an unreasonable deformation as a penalty.

[0086] Trajectory smoothness reward (Initial weight 0.2, auxiliary priority)

[0087] The core objective is to improve driving comfort by constraining the rate of change of the steering angle, thereby avoiding trajectory oscillations caused by frequent vehicle direction adjustments. The formula is:

[0088] ;

[0089] in, (Approximately 8.6°) is the maximum rate of change of steering angle that the vehicle can withstand (refer to the performance of a passenger car steering system); when the rate of change of steering angle is less than this value, the reward increases as the rate of change decreases, encouraging smooth steering; when the rate of change exceeds this value, a penalty (-8) is triggered to constrain the steering action to be non-abrupt and improve passenger comfort.

[0090] S4: Construction of Dynamic Riemannian Fluid Field and Calculation of Repulsive Force

[0091] This step constructs a continuously differentiable anisotropic repulsive potential field based on the deformation parameters output by PPO and the basic parameters of different obstacle types:

[0092] Based on the risk characteristics of obstacle types in Section 2.1, the basic parameters of the potential field for each type of obstacle are defined to ensure that the potential field shape accurately matches the obstacle risk, avoiding the problem of insufficient protection in high-risk scenarios and redundant protection in low-risk scenarios. The specific parameters are shown in the table below:

[0093] Obstacle types Basic safety radius ( ) Basic deformation gain Maximum radius of influence of potential field ( ) Basic repulsion coefficient pedestrian( ) 2.0 1.2 8 20 vehicle( ) 3.0 1.5 15 30 Static obstacles ( ) 2.5 1.0 10 25

[0094] in, The basic repulsion coefficient determines the overall strength of the repulsion force. The vehicle has the highest risk (large size, long braking distance) and the largest coefficient (30); the pedestrian has the second highest risk (random movement) and the medium coefficient (20); the static obstacle has the lowest risk (no movement) and the smallest coefficient (25). All parameters have been repeatedly verified in the CARLA simulation environment to balance safety and driving efficiency.

[0095] Determine the direction of the principal axis of the potential field: When there is a static obstacle, the velocity component is zero, and the main axis defaults to the vehicle's direction of travel.

[0096] The direction angle of the principal axis of the potential field in the vehicle coordinate system (unit: rad). The principal axis direction is the core direction of the longitudinal extension of the potential field (Note: In the vehicle coordinate system, the 0 radian direction corresponds to the front of the vehicle, ensuring that the potential field of static obstacles extends along the vehicle's driving path by default, providing longitudinal protection). It is also the concentrated area of ​​high-risk directions of obstacles.

[0097] , : The components of the obstacle's velocity vector along the X-axis (pointing towards the front of the vehicle) and Y-axis (left side of the vehicle) in the vehicle coordinate system, for static obstacles. ,at this time (The main axis direction is by default along the X-axis, i.e., the vehicle's direction of travel).

[0098] The rotational offset angle output by PPO is used to fine-tune the main axis direction to adapt to changes in the movement trend of obstacles (such as when a vehicle changes lanes, the main axis deflects with the direction of lane change to ensure that the repulsive force always points in the high-risk direction).

[0099] Constructing the Riemannian metric tensor (covariance matrix):

[0100] Construct rotation matrix Rotate the standard coordinate system (XY axis) to the direction of the principal axis of the potential field, ensuring that the principal axis of the potential field is aligned with the direction of high risk of obstacles. The formula is as follows: ;

[0101] The core function of this matrix is ​​to rotate the direction, so that the longitudinal extension direction of the potential field accurately matches the direction of obstacle movement (high-risk direction), and avoids deviation of the repulsive force direction.

[0102] Constructing the deformation matrix According to the PPO output , By adjusting the degree of longitudinal extension and lateral contraction of the potential field, anisotropy of the potential field can be achieved. The formula is as follows: ;

[0103] in, The longitudinal semi-axis length of the potential field (along the principal axis direction). The length of the transverse half-axis (perpendicular to the main axis direction); through and The difference causes the potential field to take on an elliptical shape—the potential field extends longer in the high-risk direction (vertical) and contracts narrower in the low-risk direction (lateral), accurately matching the risk distribution of obstacles.

[0104] Calculate the covariance matrix By combining the rotation matrix and the deformation matrix, the final Riemannian metric tensor is obtained, with the following formula: .

[0105] in, Rotation matrix The transpose of the rotation matrix (the inverse of the rotation matrix is ​​equal to its transpose, simplifying the calculation); As a real symmetric positive definite matrix, it uniquely determines the elliptical shape of the potential field in the two-dimensional plane, ensuring that the repulsive force at any position of the potential field can be obtained through gradient calculation without piecewise abrupt changes, thus guaranteeing the smoothness of the repulsive force from the mathematical foundation.

[0106] Constructing a repulsive force field:

[0107] A repulsive potential field is constructed using Mahalanobis distance (a distance metric that considers the elliptical shape of the potential field, distinct from traditional Euclidean distance). This ensures that the repulsive force dynamically changes with distance and the shape of the potential field, accurately protecting against high-risk directions while avoiding excessive interference from low-risk directions.

[0108] ;

[0109] ;

[0110] .

[0111] in, It is the inverse of the covariance matrix, reflecting the influence of the potential field shape on the distance metric; The function is exponential, ensuring that the potential energy decays rapidly with increasing "Maharanobis distance," thus preventing distant obstacles from excessively influencing vehicle decisions; when When the repulsive potential energy is 0, that is, the obstacle is outside the range of influence and no repulsive force is generated, thus avoiding meaningless computational load.

[0112] Repulsive force (negative gradient): The repulsive force is the force that repels potential energy along the negative gradient of the direction of fastest potential energy reduction, ensuring that the vehicle avoids obstacles along the safest path. The formula is:

[0113] ;

[0114] After expansion, the components are specifically X and Y axes:

[0115] ;

[0116] ;

[0117] The core advantage of this repulsive force is its adaptive direction: it depends not only on the direction of the line connecting the vehicle and the obstacle, but also on the potential field shape ( Dynamic adjustment can guide the vehicle to smoothly navigate along the tangent of the potential field ellipse, avoiding steering oscillations caused by traditional potential fields and improving driving comfort.

[0118] Calculation of total repulsive force of multiple obstacles

[0119] When multiple obstacles exist in the environment (e.g., "pedestrian on the left + vehicle in front + guardrail on the right"), the total repulsive force is the vector sum of the repulsive forces of all obstacles. The vector sum calculation considers both the magnitude and direction of the force, ensuring the vehicle covers the risks of all obstacles without overlooking any protection target. The formula is:

[0120] ;

[0121] ;

[0122] ;

[0123] in, These represent the number of pedestrians, vehicles, and static obstacles in the environment, respectively. For the first The repulsive force of a pedestrian on a vehicle has an X-axis component, and the same applies to the others. For example, in the scenario of "pedestrian on the left + vehicle in front", the repulsive force of the pedestrian on the left will generate a rightward component, and the repulsive force of the vehicle in front will generate a backward component. The total repulsive force will guide the vehicle to smoothly avoid the obstacle to the right and rear, while moving away from both types of obstacles, thus achieving multi-target collaborative obstacle avoidance.

[0124] S5: Gravity and Virtual Escape Force Calculation

[0125] Gravitational calculations: using a quadratic gravitational potential field:

[0126] ;

[0127] The corresponding gravitational force is:

[0128] ;

[0129] in, The gravity coefficient (which needs to be verified through extensive simulations to ensure that the coefficient can balance the attraction and repulsion: ensuring that the vehicle can be effectively pulled to the target point while avoiding excessive gravity that could cause a collision with obstacles). The x and y coordinates are the target point; the direction of gravity always points to the target point, and its magnitude increases with distance, ensuring that the vehicle can be quickly pulled back to the target path when it is far away from the target point, and smoothly adjusted when it is close to the target point without producing violent movements.

[0130] Virtual escape force calculation: This force is used to break local minimum deadlocks, and its direction is perpendicular to the direction of gravity. The calculation formula is as follows:

[0131] ;

[0132] in, Escape perturbation factor output by PPO (different obstacle types correspond to different ranges, accurately matching deadlock risk); and The unit vector component of the gravitational force direction is represented by the -y component and the x component. By combining the -y component and the x component, the escape force direction is ensured to be perpendicular to the gravitational force direction (e.g., when the gravitational force is forward, the escape force is to the right; when the gravitational force is to the right, the escape force is backward), effectively breaking the collinear balance between gravity and repulsion. To prevent zero constant, numerical anomalies are avoided when the vehicle reaches the target point (when gravity is zero).

[0133] The deadlock condition is: .

[0134] S6: Calculation of Total Potential Force

[0135] Total potential force It is the final resultant force guiding the vehicle's movement, combining the effects of gravity (pointing towards the target), total repulsive force (moving away from obstacles), and virtual escape force (breaking deadlock). Based on this and combined with vehicle dynamics constraints, it outputs executable control commands. Specific steps:

[0136] ;

[0137] in, The X and Y components of gravity. This is a component of the total repulsive force. The component of the virtual escape force; the direction of the resultant force determines the vehicle's obstacle avoidance direction, and its magnitude determines the magnitude of the vehicle's acceleration.

[0138] S7: Vehicle Control Command Generation and Execution

[0139] According to the total potential force Based on vehicle dynamics constraints, the final control commands are generated:

[0140] Steering angle calculation: The steering angle directly determines the vehicle's direction of travel and must be strictly constrained within the range that the vehicle chassis can withstand (to avoid oversteering leading to instability).

[0141]

[0142] in, It is the direction angle of the total potential force (i.e., the ideal steering angle of the vehicle). (Approximately 17.2°) is the steering angle adjustment range within a single control cycle (100ms); when the ideal steering angle is within the allowable range, it is used directly; when it exceeds the range, the maximum steering angle is taken (retaining the direction) to ensure that the steering action is safe and feasible.

[0143] Acceleration calculation: Acceleration determines the intensity of a vehicle's acceleration and deceleration. The total potential force needs to be projected onto the vehicle's current heading to distinguish between acceleration and braking forces.

[0144] ;

[0145] in:

[0146] The resultant force modulus;

[0147] The direction angle of the resultant force. This refers to the current angle at which the vehicle is facing.

[0148] The term implements projection: when the resultant force is in the same direction as the front of the car, it is positive (acceleration), and when it is opposite, it is negative (deceleration).

[0149] Force-acceleration conversion factor (converts the magnitude of force into units of acceleration to suit the mass characteristics of the vehicle).

[0150] .

[0151] in, For maximum comfort acceleration (referring to ergonomic principles to prevent passengers from leaning forward). To achieve maximum comfortable deceleration (avoiding passenger backward leaning while ensuring braking efficiency); when the initially calculated acceleration... When the value exceeds the range, the boundary value is used to ensure driving comfort and safety.

[0152] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0153] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0154] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for obstacle avoidance using a dynamic Riemannian manifold artificial potential field based on PPO reinforcement learning, characterized in that, The method includes: S1. Multi-source data fusion is performed using lidar, millimeter-wave radar, and cameras to collect real-time information on intelligent vehicle status, obstacle information, and risk characteristic parameters. S2. Based on real-time collected intelligent vehicle status, obstacle information, and risk characteristic parameters, construct an 11-dimensional high-dimensional spatiotemporal state vector. ; S3, convert the state vector The input is fed into a pre-trained PPO policy network, which outputs 4D Riemannian fluid field deformation parameters for different obstacle types. S4. Construct a dynamic Riemannian fluid field based on the deformation parameters of the 4D Riemannian fluid field and the basic parameters corresponding to the obstacle types. S5. Establish a gravitational potential field Calculate the gravitational force that guides the vehicle toward the target point. And virtual escape forces used to break local minimum deadlocks. The direction of the virtual escape force is perpendicular to the direction of gravity; S6. Repeat the calculation of repulsive forces for all obstacles in the environment, and perform vector superposition to obtain the total repulsive force. The total repulsive force ,gravitational and virtual escape force Vector synthesis is performed to obtain the total potential force. ; S7. Based on the total potential force Based on vehicle dynamics constraints, the steering angle of the intelligent vehicle within the current control cycle is calculated and output. and acceleration The command is given, and the vehicle is driven to perform obstacle avoidance maneuvers.

2. The method according to claim 1, characterized in that, S1 specifically includes: By fusing multi-source data from lidar, millimeter-wave radar, and cameras, the current status of intelligent vehicles, obstacle information, and risk characteristic parameters are collected. The intelligent vehicle status includes its current location. Longitudinal driving speed lateral acceleration yaw rate and target location ; The obstacle information includes obstacle type. Real-time location Speed ​​of movement Straight-line distance between the vehicle and the obstacle and the rate of change of distance ; The type of obstacle This includes pedestrians, vehicles, and static obstacles; The risk characteristic parameters include collision urgency. and local obstacle density .

3. The method according to claim 1, characterized in that, S2 specifically includes: The parameters obtained in S1, including the intelligent vehicle state, obstacle information, and risk characteristics, are used to construct an 11-dimensional state vector. : ； The state vector The input is fed into a pre-trained PPO policy network, which outputs an action specific to the current obstacle type. 4D Riemannian flow field deformation parameter action vector The action vector Including longitudinal expansion coefficient lateral shrinkage coefficient Rotational offset angle and escape disturbance factor ; It is the angular deviation between the vehicle's heading and the target direction, reflecting the degree of deviation between the vehicle's current direction of travel and the target direction; It is the relative angle between the velocity vectors of the vehicle and the obstacle.

4. The method according to claim 1, characterized in that, S3 specifically includes: Based on the deformation parameters, motion vector, and obstacle type Based on the corresponding fundamental parameters, a dynamic Riemannian flow situation field is constructed; The state vector The input is fed into a pre-trained PPO policy network, and the PPO policy network outputs an action specific to the current obstacle type. 4D action vector That is, the deformation parameters of the Riemannian flow field: ； in, It is the longitudinal expansion coefficient, which controls the extension of the potential field along the principal axis of risk; The horizontal contraction coefficient controls the width in the vertical direction; This is the rotational bias angle, used to fine-tune the direction of the principal axis of the potential field; The escape disturbance factor determines the magnitude of the virtual escape force.

5. The method according to claim 1, characterized in that, S4 specifically includes: Based on the obstacle speed and the aforementioned rotational offset angle Determine the direction of the principal axis of the potential field ; ； in: , : The components of the obstacle's velocity vector along the X and Y axes of the vehicle coordinate system, when the obstacle is static. ; The rotational offset angle output by the PPO is used to fine-tune the main axis direction to adapt to changes in the obstacle's movement trend. Construct rotation matrix and deformation matrix : ； ； in: Let be the length of the longitudinal semi-axis of the potential field. This is the length of the horizontal half-axis; Calculate the covariance matrix used to define the elliptical shape of the potential field. ; Based on the covariance matrix Basic repulsion coefficient and maximum radius of influence Calculate the repulsive potential field against the obstacle. and the resulting repulsive force : ； ； ； For all obstacles in the environment, the repulsive force is repeatedly calculated and then vectored to obtain the total repulsive force. ; ； ； ； in: These are the components of the total repulsive force along the X and Y axes; These represent the number of pedestrians, vehicles, and static obstacles in the environment, respectively. For the first The X-axis component of the repulsive force of a pedestrian on a vehicle; For the first The X-axis component of the repulsive force between vehicles; For the first The repulsive force of a static obstacle on the vehicle along the X-axis; For the first The Y-axis component of the repulsive force of a pedestrian on a vehicle; For the first The Y-axis component of the repulsive force between vehicles; For the first The Y-axis component of the repulsive force exerted by a static obstacle on the vehicle.

6. The method according to claim 1, characterized in that, S5 specifically includes: Establish a gravitational potential field : ； in, It is the gravitational coefficient; These are the x and y coordinates of the target point, respectively; x: represents the x-coordinate of the current location of the intelligent vehicle, and y: represents the y-coordinate of the current location of the intelligent vehicle. Calculate the gravitational force guiding the vehicle toward the target point. : ； Calculate the virtual escape force used to break local minimum deadlock. : ; in, The escape perturbation factor is the output of the PPO. and This is the unit vector component of the direction of gravity.

7. The method according to claim 1, characterized in that, S6 specifically includes: The total repulsive force ,gravitational and virtual escape force Vector synthesis is performed to obtain the total potential force. : ； These are the components of gravity along the X and Y axes. The virtual escape force is represented by its components on the X and Y axes.

8. The method according to claim 1, characterized in that, Specifically, S7 includes: According to the total potential force Calculate the direction angle of the total potential force. : ； , These are the components of the total potential force along the X and Y axes; Based on vehicle dynamics constraints, the heading angle Convert to steering angle The constraints are: ； Total potential force Projected onto the vehicle's current heading, calculate the initial acceleration. : ； in: The resultant force modulus; The direction angle of the resultant force. This refers to the current angle at which the vehicle is facing. Force-acceleration conversion coefficient; Based on comfort constraints, the initial acceleration By limiting the amplitude, the final acceleration is obtained. The constraints are: 。

Images