High parallelism reconfigurable multi-axle vehicle dynamics simulation method

Abstract

The application discloses a high-parallelism reconfigurable multi-axle vehicle dynamics simulation method, a reconfigurable parameter model is constructed by reading parameters; a six-degree-of-freedom coupled dynamics model is established; an initial state is solved based on static force balance; tire slip rate and braking force are calculated; terrain ground contact determination is carried out based on a double-layer elevation map; and vehicle pose and dynamics state are updated by multi-substep integration. The key lies in that a single-step evolution process of the dynamics model is converted into a batch state transition mapping and a computation graph is constructed, the computation graph is cyclically fused and compiled and exported, a high-parallelism simulation execution model supporting input tensors with variable batch size in runtime is generated, and batch control instructions are received, and parallel stepping is implemented on effective vehicles in combination with an activation mask matrix. The application improves adaptability under complex terrain through unified parameter model and double-layer terrain determination, and greatly improves the calculation efficiency and precision of multi-vehicle batch simulation through computation graph compilation and mask parallel processing.

Description

Technical Field

[0001] This application relates to the field of vehicle dynamics simulation technology, specifically to a highly parallel reconfigurable multi-axle vehicle dynamics simulation method. Background Technology

[0002] With the rapid development of intelligent connected vehicles and autonomous driving technologies, vehicle dynamics simulation, as a core component of algorithm development, testing and verification, and vehicle performance evaluation, is becoming increasingly important. Currently, most vehicle dynamics simulation models are customized for two-axle passenger cars or commercial vehicles with a fixed number of axles. Specifically, existing technologies have the following significant drawbacks and shortcomings:

[0003] First, multi-axle configurations suffer from extremely poor reusability and low modeling efficiency. Most existing vehicle dynamics simulation methods are designed for two-axle or fixed-configuration vehicles, with model structures and computational parameters typically fixed during the development phase. However, for heavy-duty transport vehicles, heavy commercial vehicles, and special engineering vehicles, there are significant differences in the number of axles (three, four, five, or even six), the number of tires, wheelbase distribution, center of gravity location, suspension stiffness parameters, and tire mounting coordinates. Traditional fixed-structure models require development engineers to perform geometric modeling, parameter tuning, and model verification from scratch for each vehicle model when encountering different vehicle types. This results in extremely poor model reusability, long development cycles, and enormous consumption of human and computational resources.

[0004] Second, planar dynamics models struggle to reflect the complex coupled responses of multi-axle vehicles. Current mainstream industrial vehicle simulations typically employ simplified planar dynamics models, focusing primarily on longitudinal, lateral, and yaw motions, while neglecting the vehicle's vertical, pitch, roll, and the dynamic transfer of tire normal loads. However, under complex conditions such as braking, sharp turns, traversing uneven surfaces, and bridges, the vertical load transfer and suspension coupling of multi-axle vehicles have a decisive impact on their pitch and roll attitudes due to the large number of axles, uneven mass distribution, and complex suspension systems. Ignoring these coupling factors often leads to simulation results that significantly deviate from the vehicle's actual physical motion, potentially causing unexpected loss of control during real-world testing of autonomous driving strategies.

[0005] Third, existing models suffer from massive computational demands, making it difficult to meet the requirements of batch parallelism and high real-time performance. To simulate high-fidelity dynamic responses in scenarios with high centers of gravity, multiple wheelbases, and multiple tire coupling, high-fidelity vehicle models typically contain a large number of differential and algebraic equations, resulting in extremely large computational demands for a single dynamic step. Especially in current mainstream autonomous driving algorithms' reinforcement learning training, batch testing, and scenario simulation, simulations of hundreds or thousands of vehicles often need to run simultaneously. Traditional vehicle-by-vehicle serial or simple multi-threaded parallel simulation methods are easily limited by CPU resource bottlenecks, making it difficult to meet the real-time requirements of training and testing. Current technologies lack effective means to transform complex multi-axle vehicle physical models into computational graph architectures compatible with large-scale GPU / TPU parallelism, thus failing to balance high fidelity and high real-time performance.

[0006] Fourth, there are deficiencies in the adaptation and grounding determination for complex terrain and multi-level roads. Traditional vehicle simulations mostly only support mapping and querying single-layer road surface elevation maps when dealing with terrain elevations. However, in real traffic scenarios, there are numerous three-dimensional traffic environments such as bridges, viaducts, ramps, and interchanges. When a vehicle enters these scenarios, each wheel may come into contact with different elevation layers. Existing conventional elevation map query methods cannot distinguish whether the current tire is in contact with the ground layer or the bridge surface layer, which can easily lead to errors in vehicle penetration of terrain or force calculations, resulting in instantaneous divergence in dynamic simulations. Summary of the Invention

[0007] The main objective of this application is to provide a highly parallel and reconfigurable multi-axle vehicle dynamics simulation method, including the following steps:

[0008] Step S1: Read the vehicle configuration parameters, parse the number of axles, tire mounting positions, vehicle mass, inertia tensor, suspension parameters, tire parameters and braking parameters, and construct a reconfigurable parameter model corresponding to the vehicle configuration.

[0009] Step S2: Based on the tire mounting position and vehicle center of gravity parameters in the reconfigurable parameter model, construct the vehicle's lateral and longitudinal dynamics model, vertical dynamics model, and rigid body attitude dynamics model to form a six-degree-of-freedom coupled dynamics model for a multi-axle vehicle.

[0010] Step S3: Based on the static balance relationship of the vehicle's weight, tire support force, and suspension support force, as well as the suspension-tire coupling relationship, solve for the steady-state initial state vector of the vehicle under the initial terrain conditions.

[0011] Step S4: Based on the tire longitudinal slip ratio, sideslip angle and normal load, calculate the basic longitudinal force and basic lateral force of each wheel, introduce the attachment circle constraint to uniformly scale the longitudinal and lateral forces, and use the tire transient hysteresis state to dynamically correct the tire force to obtain the actual tire force participating in the vehicle dynamics solution.

[0012] Step S5: Based on the friction dynamics model of the braking system, calculate the actual braking torque of each wheel edge, and input the braking torque, driving torque, and tire longitudinal force into the wheel rotation equation to update the wheel speed state of each wheel.

[0013] Step S6: Based on the elevation map and terrain coordinate mapping relationship, calculate the road surface height information of each wheel contact point. For a two-layer terrain scene that includes bridges or overpasses, read the elevation information of the ground layer and the bridge deck layer at the same time. Based on the distance relationship between the tire reference height and the height of the two terrain layers, determine the current contact layer and extract the road surface height of the corresponding layer.

[0014] Step S7: Couple the tire forces, suspension forces, gravity projections and rigid body moments of each wheel, and update the vehicle position, attitude, velocity and tire state using a multi-substep integration method to complete the vehicle single-step dynamics evolution.

[0015] Step S8: Convert the single-step dynamic evolution process into a batch state transition mapping containing multi-sub-step loops, construct the corresponding simulation calculation graph, perform loop fusion optimization and compilation export on the simulation calculation graph, and generate a compiled simulation execution model that supports variable batch size input tensors at runtime.

[0016] Step S9: Receive driving, braking, and steering commands from a batch of vehicles, construct a mask matrix containing the vehicle activation states, drive the compiled simulation execution model to perform parallel stepping only on the currently active valid vehicles based on the mask matrix, keep the previous state unchanged for inactive vehicles, and output the pose state, speed state, tire state, and tire force state of each vehicle.

[0017] In one embodiment, step S1 specifically includes:

[0018] Based on the configuration relationship that each axle includes left and right wheel sets, and each wheel set includes at least one tire, the total number of tires in the vehicle is determined, and the mounting position vector of each wheel in the vehicle coordinate system is generated.

[0019] According to the wheel position sequence, the vehicle mass, sprung mass, unsprung mass of each wheel, vehicle body inertia tensor, suspension stiffness matrix, suspension damping matrix, tire vertical stiffness matrix, and tire vertical damping matrix are organized into a unified parameter tensor to adapt to the unified parameter loading and state initialization of three-axle, four-axle, five-axle, and six-axle vehicles.

[0020] In one embodiment, in step S2, the six-degree-of-freedom coupled dynamics model includes a vehicle translational dynamics model, a vehicle rotational dynamics model, a posture kinematics model, and a suspension vertical coupling relationship;

[0021] The vehicle translational dynamics model incorporates longitudinal, lateral, and vertical velocities into the force balance relationship of the vehicle's center of mass; the vehicle rotational dynamics model establishes the equations for roll, pitch, and yaw motion based on the torque balance relationship of the vehicle around its center of mass; and the suspension vertical coupling relationship incorporates the vertical displacement of sprung mass, roll angle, and pitch angle into the calculation formulas for suspension deflection and suspension force of each wheel.

[0022] In one embodiment, step S3 specifically includes:

[0023] Based on the lateral and longitudinal lever arms of the normal loads of each wheel relative to the center of mass, the moment balance relationships in the roll and pitch directions are established respectively. Combined with tire compression, suspension deflection and terrain elevation, the vertical force balance equations, roll moment balance equations, pitch moment balance equations and tire-suspension coupling relationships of each wheel are combined into a system of linear equations, and the steady-state initial state vector is obtained by solving the system.

[0024] In one embodiment, step S4 specifically includes:

[0025] Based on the velocity decomposition results at the wheel center, the longitudinal slip ratio and sideslip angle of each wheel are obtained. The basic longitudinal force and basic lateral force of each wheel are calculated based on the tire stiffness parameters. The tire-road adhesion constraint is introduced to uniformly scale the longitudinal and lateral tire forces. Furthermore, the tire transient hysteresis state is introduced to obtain the actual tire forces participating in the solution of the vehicle dynamics.

[0026] In one embodiment, step S5 specifically includes:

[0027] The frictional torque at each wheel's brake disc is calculated based on the braking input and brake structure parameters. The braking friction coefficient is then introduced into the first-order dynamic hysteresis equation to simulate the hydraulic build-up, friction pair response, and transmission chain hysteresis process in the braking system. The braking torque, driving torque, and tire longitudinal force are then written into the wheel rotation equations for each wheel, and the angular acceleration of the wheel's equivalent moment of inertia is integrated.

[0028] In one embodiment, in step S6, the contact layer in the dual-terrain scene is determined as follows: by using the planar coordinates of each wheel's contact point in the ground coordinate system, and simultaneously querying the ground layer elevation map and the bridge deck layer elevation map, the difference between the tire reference height and the ground layer and the difference between the tire reference height and the bridge deck layer are obtained respectively; the layer corresponding to the smaller absolute value of the difference is selected as the current contact layer.

[0029] In one embodiment, in step S7, the simulation step size is divided into multiple sub-step sizes, and the vehicle state is iteratively integrated based on each sub-step size. For the vehicle attitude state, a quaternion update method is adopted, which multiplies the quaternion update matrix constructed by the current angular velocity with the quaternion of the previous moment to complete the attitude cross-sub-step update. After the integration of each sub-step, the updated quaternion is normalized to avoid numerical drift.

[0030] In one embodiment, step S8, specifically includes the cyclic fusion optimization:

[0031] The physical model computation nodes in steps S1 to S7, including tire force calculation nodes, suspension force calculation nodes, multi-substep integration nodes, and terrain lookup nodes, are abstracted into computation graph nodes through data dependencies. The same computation nodes in the multi-substep loop are merged into a single compiled computation kernel function by operator fusion.

[0032] Vehicle parameters, control inputs, and states are tensorized, allowing a single kernel function call to complete the single-step evolution of all activated vehicles.

[0033] In one embodiment, in step S9, the dimension of the mask matrix is ​​consistent with the dimension of the batch of vehicles; during parallel stepping, the Boolean value at the corresponding position in the mask matrix is ​​used to determine whether the vehicle is activated; for activated vehicles, the corresponding control commands and state data are passed into the compiled simulation execution model; for inactive vehicles, the computation graph branch of the vehicle is skipped, and the state data of the previous moment is directly assigned to the output result of the current moment.

[0034] Therefore, this application has the following beneficial effects:

[0035] First, this application achieves unified organization of vehicle axle count, wheel layout, mass parameters, suspension parameters, and tire parameters through a configuration-driven approach. This allows three-axle, four-axle, five-axle, and six-axle vehicles to complete parameter loading, state initialization, and dynamic solution within the same model framework, avoiding the problem of repetitive modeling for different vehicle types and improving the reusability and scalability of multi-axle vehicle dynamics models. Simultaneously, this application integrates lateral and longitudinal dynamics, vertical dynamics, rigid body attitude dynamics, tire dynamics, braking dynamics, and terrain grounding relationships into the solution process, enabling calculations of vehicle translation, rotation, normal load changes, and tire force changes within a unified framework. Furthermore, by constructing a highly parallel simulation execution model through multi-step integration, batch tensor organization, and loop fusion compilation, the vehicle dynamics simulation is expanded from vehicle-by-vehicle calculation to a parallel step-by-step approach oriented towards batch scenarios. While maintaining the integrity of the model calculation process, this improves the execution efficiency of multi-vehicle parallel simulation.

[0036] Second, as supplementary evidence of the inventive step of the claims in this application, it is also reflected in the following important aspects:

[0037] The expected benefits after the technical solution of this application is transformed are:

[0038] This application reduces the workload of repetitive modeling and maintenance between vehicles with different numbers of axles by using a reconfigurable parameter model and a unified simulation execution framework, which helps to shorten the development and adjustment process of multi-axle vehicle dynamics models.

[0039] The technical solution of this application fills a technological gap in the industry both domestically and internationally:

[0040] Existing vehicle dynamics simulation methods are mostly designed for vehicles with fixed configurations, lacking integrated processing for unified parameter organization, dual-layer terrain grounding determination, and batch parallel execution for multi-axle vehicles. This application incorporates the above-mentioned aspects into a single technical solution, forming a more complete multi-axle vehicle dynamics simulation process.

[0041] The technical solution of this application solves a technical problem that people have long desired to solve but have been unable to achieve successfully:

[0042] Multi-axle vehicle dynamics models need to consider factors such as vertical orientation, attitude, load transfer, and tire forces, while also adapting to the parallel computing requirements of batch scenarios. This application comprehensively addresses these issues through unified parameter organization, static equilibrium initialization, multi-substep integration, and batch state transition mapping. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a system flowchart of a highly parallel and reconfigurable multi-axle vehicle dynamics simulation method;

[0045] Figure 2 This is a schematic diagram of the multi-axle vehicle configuration and wheel layout in a highly parallel and reconfigurable multi-axle vehicle dynamics simulation method.

[0046] Figure 3 This is a system module diagram of a highly parallel, reconfigurable multi-axle vehicle dynamics simulation method. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0049] To address the shortcomings of existing technologies, this application provides a highly parallel, reconfigurable multi-axle vehicle dynamics simulation method, including steps S1-S9, as described above. Figure 1 , Figure 1 This is a system flowchart of a highly parallel, reconfigurable multi-axle vehicle dynamics simulation method.

[0050] Step S1: Read the vehicle configuration parameters, parse the number of axles, tire mounting positions, vehicle mass, inertia tensor, suspension parameters, tire parameters and braking parameters, and construct a reconfigurable parameter model corresponding to the vehicle configuration.

[0051] Step S2: Based on the tire mounting position and vehicle center of gravity parameters in the reconfigurable parameter model, construct the vehicle's lateral and longitudinal dynamics model, vertical dynamics model, and rigid body attitude dynamics model to form a six-degree-of-freedom coupled dynamics model for a multi-axle vehicle.

[0052] Step S3: Based on the static balance relationship of the vehicle's weight, tire support force, and suspension support force, as well as the suspension-tire coupling relationship, solve for the steady-state initial state vector of the vehicle under the initial terrain conditions.

[0053] Step S4: Based on the tire longitudinal slip ratio, sideslip angle and normal load, calculate the basic longitudinal force and basic lateral force of each wheel, introduce the attachment circle constraint to uniformly scale the longitudinal and lateral forces, and use the tire transient hysteresis state to dynamically correct the tire force to obtain the actual tire force participating in the vehicle dynamics solution.

[0054] Step S5: Based on the friction dynamics model of the braking system, calculate the actual braking torque of each wheel edge, and input the braking torque, driving torque, and tire longitudinal force into the wheel rotation equation to update the wheel speed state of each wheel.

[0055] Step S6: Based on the elevation map and terrain coordinate mapping relationship, calculate the road surface height information of each wheel contact point. For a two-layer terrain scene that includes bridges or overpasses, read the elevation information of the ground layer and the bridge deck layer at the same time. Based on the distance relationship between the tire reference height and the height of the two terrain layers, determine the current contact layer and extract the road surface height of the corresponding layer.

[0056] Step S7: Couple the tire forces, suspension forces, gravity projections and rigid body moments of each wheel, and update the vehicle position, attitude, velocity and tire state using a multi-substep integration method to complete the vehicle single-step dynamics evolution.

[0057] Step S8: Convert the single-step dynamic evolution process into a batch state transition mapping containing multi-sub-step loops, construct the corresponding simulation calculation graph, perform loop fusion optimization and compilation export on the simulation calculation graph, and generate a compiled simulation execution model that supports variable batch size input tensors at runtime.

[0058] Step S9: Receive driving, braking, and steering commands from a batch of vehicles, construct a mask matrix containing the vehicle activation states, drive the compiled simulation execution model to perform parallel stepping only on the currently active valid vehicles based on the mask matrix, keep the previous state unchanged for inactive vehicles, and output the pose state, speed state, tire state, and tire force state of each vehicle.

[0059] Specifically, in this embodiment, such as Figure 1 As shown, this application provides a highly parallel, reconfigurable multi-axle vehicle dynamics simulation method. In this embodiment, the method can be deployed on a computing server with a high-performance graphics processing unit (GPU) or tensor processor (TPU), utilizing just-in-time (JIT) compilation technology to transform the vehicle physics model into a high-performance computing kernel, thereby simultaneously driving hundreds or even thousands of vehicles to perform real-time physical evolution in a tensor-based manner. The method specifically includes:

[0060] Step S1: Construct a reconfigurable parametric model.

[0061] In this embodiment, a structured configuration file input by the user is first received. The number of vehicle axles n in the file is then parsed. a Based on the configuration relationship that each axle includes left and right wheel sets, and each wheel set includes at least one tire, the total number of tires is automatically determined to be n. w =2n a Simultaneously, the mounting position vectors of the relative centers of mass of each wheel, the vehicle mass, the sprung mass, and the unsprung mass of each axle are extracted. Then, the scalar and matrix parameters, including the vehicle inertia tensor, the suspension stiffness and damping of each wheel, and the vertical stiffness of the tires, are arranged in wheel position order and encapsulated into a unified parameter tensor in contiguous memory. This unified tensor structure allows the system to seamlessly switch between models of three-, four-, five-, and six-axle vehicles without modifying the underlying code, simply by changing the configuration file, achieving extremely high reconfigurability of parameter-driven configurations.

[0062] Step S2: Construct a six-degree-of-freedom coupled dynamic model.

[0063] Based on the constructed parametric tensors, a six-degree-of-freedom coupled dynamics model of a multi-axle vehicle is established in the physics engine. Specifically, in the vehicle coordinate system, rigid body translation equations are established according to the Newton-Euler equations to describe the vehicle's longitudinal, lateral, and vertical accelerations; and rigid body rotation equations are established according to the Euler rotation equations to describe the vehicle's roll, pitch, and yaw accelerations. Simultaneously, the vertical displacement, roll angle, and pitch angle of the sprung mass are dynamically incorporated into the geometric deformation calculation of each wheel suspension, enabling accurate calculation of suspension deflection and suspension support force. This strongly couples the sprung mass motion with the unsprung mass motion, improving the physical consistency of the multi-axle vehicle's attitude response under complex high center of gravity conditions.

[0064] Step S3: Initialize the static equilibrium state.

[0065] Because multi-axle vehicles have many constraints, improper initial state settings can easily lead to divergence of the initial integral. This step is designed to address this problem. Specifically, the initial terrain elevation information is read, and the vertical balance relationships of the vehicle's weight, the tire support forces of each wheel, and the suspension support forces, as well as the balance relationships of the rolling moment and pitch moment around the center of mass, are combined into a set of linear equations, i.e., constructing a matrix form Aξ=b. Using this matrix solver, the initial roll angle, pitch angle, and normal load preload of each wheel are obtained. These precise initial static balance values ​​are directly set as the initial state for the next dynamic simulation, ensuring physical stability at the start of the simulation and preventing non-physical phenomena such as "dropping" or "launching" of the vehicle at the moment of simulation start.

[0066] Step S4: Tire dynamics and transient hysteresis calculation.

[0067] This step performs high-precision calculations of tire forces. First, based on the wheel center velocity decomposition results, the longitudinal slip ratio and sideslip angle of each wheel are calculated using the effective rolling radius and angular velocity. The basic longitudinal and lateral forces are calculated using the tire's longitudinal slip stiffness and sideslip stiffness, and a friction ellipse (adhesion circle) constraint is introduced to uniformly scale the longitudinal and lateral forces, ensuring that the tire's net force does not exceed the road adhesion limit. To realistically reflect the tire's dynamic response under extreme conditions, this embodiment also introduces a first-order hysteresis element (i.e., transient hysteresis state) to dynamically correct the target tire force. Through a smooth transition, high-frequency oscillations caused by sudden changes in tire force are avoided, ensuring that the tire's output force at the edge of the adhesion limit is continuous and physically correct, thereby improving the simulation fidelity under extreme conditions.

[0068] Step S5: Modeling the friction dynamics of the braking system.

[0069] In actual vehicle braking, braking force cannot be instantaneously established due to delays in hydraulic or pneumatic systems. Therefore, this embodiment prioritizes mapping the brake pedal depth to the target friction coefficient of the braking system and introduces a braking dynamic response time constant. A first-order differential equation is used to simulate the hydraulic build-up, friction pair response, and transmission chain lag process in the brake. The calculated actual physical friction coefficient is multiplied by the brake structural parameters to obtain the wheel-side braking torque. Finally, the braking torque, driving torque, and tire longitudinal force are incorporated into the wheel rotation equation, and the wheel speeds are updated through numerical integration. This process not only accurately calculates the braking force but also reasonably considers the inertial lag of the braking system itself, significantly improving the simulation realism under braking conditions.

[0070] Step S6: Contact determination based on double-layer elevation map.

[0071] For high-fidelity simulations, terrain interaction is crucial. This step innovatively introduces a dual-layer contact determination mechanism based on conventional elevation map lookup. Specifically, the pixel coordinates of each wheel's contact point on the elevation map are calculated through coordinate mapping. For dual-layer terrain scenarios including bridges, ramps, or interchanges, the system simultaneously reads the elevation values ​​of the ground layer and the bridge deck layer at the same pixel location. Subsequently, the system calculates the difference between the tire's reference height and these two elevation layers, takes the absolute value, and selects the layer with the smallest absolute difference as the current contact layer for that tire, extracting its elevation value for subsequent tire compression calculations. This determination logic based on the closest physical distance effectively solves the problems of clipping or height judgment errors that may occur with multi-axle vehicles in three-dimensional traffic scenarios, ensuring the accuracy of traction calculations under complex terrain.

[0072] Step S7: Multi-step integration and state update.

[0073] Considering the high stiffness of multi-axle vehicle models, directly using a large step size for integration can easily lead to numerical divergence. This embodiment employs a multi-substep integration strategy, dividing a simulation step size Δt into N equal parts. sub The system operates on a substep size. Within each substep, the system recalculates tire forces, suspension forces, the projection of gravity, and rigid body moments, and updates vehicle velocity, angular velocity, and position. Specifically for vehicle attitude, the system uses quaternions for updates and performs quaternion normalization after each substep integration to prevent numerical drift during long-term integration, ensuring global singularity-free and high stability in vehicle attitude calculations under large-angle rotation scenarios.

[0074] Step S8: Computation graph fusion and compilation export.

[0075] To meet the computational demands of large-scale concurrent simulations, this step highly abstracts the physical computation process from the preceding steps. Based on temporal dependencies, all computational nodes within a multi-step loop (including tire forces, suspension forces, terrain queries, and dynamic integration) are constructed into a simulation computation graph. Using just-in-time (JIT) compilation technology, loop fusion and operator fusion optimizations are performed on the computation graph, merging the originally scattered mathematical operators into highly optimized GPU kernel functions. This fusion compilation process eliminates significant intermediate tensor memory read / write overhead, ultimately deriving a compiled simulation execution model that supports variable batch sizes of input tensors at runtime. This model can run directly on the target hardware, significantly reducing the computational latency for a single state update.

[0076] Step S9: Mask parallel stepping and batch result output.

[0077] During simulation execution, the system receives batch control commands (driving, braking, steering) containing B samples and constructs a Boolean activation mask matrix with the same dimension as the batch. When executing the compiled computation kernel, the system performs element-wise filtering based on the mask matrix: for currently active vehicles, the kernel function uses their state tensor and control inputs to perform complex dynamic evolution; for inactive vehicles (those that have completed their tasks or failed due to collisions), the system skips the cumbersome physical calculations and directly assigns the state vector from the previous time step to the next. This mask-driven strategy effectively avoids thread bundle divergence and fully utilizes the hardware's parallel throughput. After execution, the batch's pose state, velocity state, and tire force state are output as a unified tensor for direct use by external autonomous driving control algorithms, reinforcement learning training environments, or large-scale scenario simulations.

[0078] In one embodiment, step S1 specifically includes:

[0079] Based on the configuration relationship that each axle includes left and right wheel sets, and each wheel set includes at least one tire, the total number of tires in the vehicle is determined, and the mounting position vector of each wheel in the vehicle coordinate system is generated.

[0080] According to the wheel position sequence, the vehicle mass, sprung mass, unsprung mass of each wheel, vehicle body inertia tensor, suspension stiffness matrix, suspension damping matrix, tire vertical stiffness matrix, and tire vertical damping matrix are organized into a unified parameter tensor to adapt to the unified parameter loading and state initialization of three-axle, four-axle, five-axle, and six-axle vehicles.

[0081] Specifically, in this embodiment, the axle count information in the vehicle configuration parameter file is read, and the total number of tires in the vehicle is determined based on the configuration relationship that each axle includes left and right wheel sets, and each wheel set includes at least one tire. Then, the mounting position vector of each wheel in the vehicle coordinate system is generated, thereby completing the unified configuration unfolding for vehicles with different axle counts, which can be written as:

[0082]

[0083] In the formula: For the number of vehicle axles, This represents the total number of tires. For wheel position The mounting position vector relative to the vehicle's center of gravity. , , These are the longitudinal, lateral, and vertical coordinates of the wheel position in the vehicle coordinate system.

[0084] Following the wheel position sequence, the vehicle mass, sprung mass, unsprung mass of each wheel, and vehicle body inertia are written into a unified parameter tensor, enabling three-axle, four-axle, five-axle, and six-axle vehicles to be initialized under the same data structure. The corresponding data are:

[0085]

[0086] In the formula: For the overall vehicle quality, For the sprung mass, For wheel position The corresponding unsprung mass, Let be the inertia tensor of the vehicle body about its center of mass. , , These are the moments of inertia in the roll, pitch, and yaw directions, respectively.

[0087] By combining vehicle suspension and tire parameters, the suspension stiffness, damping, and tire vertical stiffness of each wheel are organized into a matrix form corresponding to each wheel position, thus obtaining a unified parameter model:

[0088]

[0089] In the formula: and These are the suspension stiffness matrix and the suspension damping matrix, respectively. and These are the tire vertical stiffness matrix and the tire vertical damping matrix, respectively; diagonal elements , , and Corresponding to wheel positions The suspension stiffness, suspension damping, tire vertical stiffness, and tire vertical damping at the location.

[0090] In one embodiment, in step S2, the six-degree-of-freedom coupled dynamics model includes a vehicle translational dynamics model, a vehicle rotational dynamics model, a posture kinematics model, and a suspension vertical coupling relationship;

[0091] The vehicle translational dynamics model incorporates longitudinal, lateral, and vertical velocities into the force balance relationship of the vehicle's center of mass; the vehicle rotational dynamics model establishes the equations for roll, pitch, and yaw motion based on the torque balance relationship of the vehicle around its center of mass; and the suspension vertical coupling relationship incorporates the vertical displacement of sprung mass, roll angle, and pitch angle into the calculation formulas for suspension deflection and suspension force of each wheel.

[0092] Specifically, in this embodiment, vehicle translational dynamics are established based on the vehicle coordinate system, and longitudinal velocity, lateral velocity, and vertical velocity are uniformly incorporated into the force balance relationship of the vehicle's center of mass, resulting in the rigid body translational equations for a multi-axle vehicle:

[0093]

[0094] In the formula: This is the velocity vector of the center of mass in the vehicle coordinate system; This is the angular velocity vector of the vehicle body; The resultant force acting on the vehicle body; This is the rotation matrix from the ground coordinate system to the vehicle coordinate system; This is the vector of gravitational acceleration.

[0095] Based on the torque balance relationship of the vehicle about its center of mass, rigid body rotation equations are established in the roll, pitch, and yaw directions. The vehicle attitude evolution satisfies:

[0096]

[0097] In the formula: It is the resultant torque formed at the vehicle's center of gravity by the forces of each wheel tire, suspension, and additional forces.

[0098] Combining rigid body dynamics and pose kinematics, the vehicle position and attitude updates can be written as follows:

[0099]

[0100] In the formula: This is the position vector of the vehicle's center of mass in the ground coordinate system; This is the rotation matrix from the vehicle coordinate system to the ground coordinate system; For the vehicle attitude quaternion; The quaternion update matrix is ​​constructed for angular velocity.

[0101] Furthermore, by incorporating the vehicle's vertical displacement, pitch angle, and roll angle into the suspension deformation calculation, the suspension deflection and suspension force of each wheel can be expressed as:

[0102]

[0103] In the formula: For wheel position index; and wheel position Corresponding to the deflection and suspension force of the suspension; This represents the vertical displacement of the vehicle body; This corresponds to the vertical displacement of the unsprung mass; This refers to the vehicle body roll angle; The vehicle body pitch angle.

[0104] In one embodiment, step S3 specifically includes:

[0105] Based on the lateral and longitudinal lever arms of the normal loads of each wheel relative to the center of mass, the moment balance relationships in the roll and pitch directions are established respectively. Combined with tire compression, suspension deflection and terrain elevation, the vertical force balance equations, roll moment balance equations, pitch moment balance equations and tire-suspension coupling relationships of each wheel are combined into a system of linear equations, and the steady-state initial state vector is obtained by solving the system.

[0106] Specifically, in this embodiment, the initial vertical equilibrium state of the vehicle is solved based on the balance relationship between the vehicle's weight, tire support force, and suspension support force. The vehicle's vertical equilibrium condition is as follows:

[0107]

[0108] In the formula: For wheel position Normal load at the location.

[0109] Based on the lateral and longitudinal lever arms of the normal loads of each wheel relative to the center of mass, the moment balance relationships in the roll and pitch directions are established respectively. From this, the initial roll and pitch angles of the vehicle should satisfy the following:

[0110]

[0111] In the formula: and wheel position The longitudinal and transverse coordinates relative to the centroid.

[0112] By incorporating tire compression, suspension deflection, and terrain elevation into the static equilibrium solution, the following matching relationship exists between tire support force and suspension support force:

[0113]

[0114] In the formula: and For wheel position Vertical tire support force and suspension support force at the location; This refers to the road surface height at the wheel's contact point.

[0115] By combining the vertical force balance, rolling moment balance, pitch moment balance of the entire vehicle, and the tire-suspension coupling relationship of each wheel, the initial equilibrium state vector of the vehicle can be obtained:

[0116]

[0117] In the formula: This is the coefficient matrix formed by static equilibrium relationships; It is a constant vector composed of vehicle weight and terrain height; This is the initial equilibrium state vector of the vehicle.

[0118] In one embodiment, step S4 specifically includes:

[0119] Based on the velocity decomposition results at the wheel center, the longitudinal slip ratio and sideslip angle of each wheel are obtained. The basic longitudinal force and basic lateral force of each wheel are calculated based on the tire stiffness parameters. The tire-road adhesion constraint is introduced to uniformly scale the longitudinal and lateral tire forces. Furthermore, the tire transient hysteresis state is introduced to obtain the actual tire forces participating in the solution of the vehicle dynamics.

[0120] Specifically, in this embodiment, the longitudinal slip ratio and sideslip angle of each wheel are defined based on the velocity decomposition results at the wheel center to describe the relative motion state between the tire and the road surface:

[0121]

[0122] In the formula: and For wheel position The longitudinal slip ratio and tire slip angle at the location; and This refers to the effective rolling radius and angular velocity of the corresponding wheel; and These are the longitudinal and lateral velocities of the wheel in the tire coordinate system, respectively. To avoid tiny positive numbers with a denominator of zero.

[0123] Based on the tire stiffness parameters and the aforementioned slip state, the longitudinal force and lateral force of each wheel base are calculated. The tire steady-state mechanical relationship can then be summarized as follows:

[0124]

[0125] In the formula: and wheel position The corresponding longitudinal slip stiffness and lateral slip stiffness; and These are the basic longitudinal force and basic lateral force of the wheel position, respectively.

[0126] By incorporating the adhesion limit condition into the tire force calculation process and uniformly scaling the longitudinal and lateral tire forces to ensure that the tire forces satisfy the tire-road adhesion circle constraint, we can obtain:

[0127]

[0128] In the formula: For wheel position The attachment scaling factor; This is the tire-to-road adhesion coefficient for that wheel position; and These are the longitudinal force and lateral force of the target after attachment constraints, respectively.

[0129] Furthermore, the transient hysteresis state of the tire is introduced to dynamically correct the target tire force, and finally the actual tire force participating in the solution of the vehicle dynamics is obtained:

[0130]

[0131] In the formula: and wheel position The dynamic state of being subjected to longitudinal and lateral forces; and These are the corresponding time constants; and The actual tire force that ultimately participates in the vehicle dynamics calculation.

[0132] Figure 2 This is a schematic diagram of the multi-axle vehicle configuration and wheel arrangement according to this application. The diagram defines the x-axis (longitudinal) and y-axis (lateral) with the vehicle's center of gravity as the origin, and shows the center velocity v of each wheel. i Side slip angle α i and tire lateral force F yi The distribution of these parameters intuitively reflects the spatial layout and kinematic decomposition of the physical parameters of multi-axle vehicles, providing geometric and kinematic support for dynamic modeling.

[0133] In one embodiment, step S5 specifically includes:

[0134] The frictional torque at each wheel's brake disc is calculated based on the braking input and brake structure parameters. The braking friction coefficient is then introduced into the first-order dynamic hysteresis equation to simulate the hydraulic build-up, friction pair response, and transmission chain hysteresis process in the braking system. The braking torque, driving torque, and tire longitudinal force are then written into the wheel rotation equations for each wheel, and the angular acceleration of the wheel's equivalent moment of inertia is integrated.

[0135] Specifically, in this embodiment, the braking input and brake structure parameters are read, and the friction torque at each wheel brake disc is calculated so that the braking command can be mapped to the wheel-side braking torque. The expression is as follows:

[0136]

[0137] In the formula: For wheel position The actual braking torque at the point; , , and These are the braking efficiency coefficient, braking friction coefficient, braking normal clamping force, and equivalent friction radius of the brake disc for the corresponding wheel position.

[0138] The braking torque, driving torque, and tire longitudinal force are incorporated into the wheel rotation equation, and the relationship between the wheel speeds is as follows:

[0139]

[0140] In the formula: For wheel position The equivalent moment of inertia at the point; This represents the driving torque for that wheel position.

[0141] To reflect the hydraulic build-up, friction pair response, and transmission chain hysteresis processes in the braking system, the braking friction coefficient is introduced into the dynamic equation, which can be further written as:

[0142]

[0143] In the formula: For wheel position The corresponding target braking friction coefficient; This is the dynamic response time constant of the braking system for that wheel position.

[0144] In one embodiment, in step S6, the contact layer in the dual-terrain scene is determined as follows: by using the planar coordinates of each wheel's contact point in the ground coordinate system, and simultaneously querying the ground layer elevation map and the bridge deck layer elevation map, the difference between the tire reference height and the ground layer and the difference between the tire reference height and the bridge deck layer are obtained respectively; the layer corresponding to the smaller absolute value of the difference is selected as the current contact layer.

[0145] Specifically, in this embodiment, the position of each wheel's contact point in the ground coordinate system is calculated based on the vehicle's center of gravity, vehicle attitude, and tire mounting position, and the corresponding elevation map pixel coordinates are generated, which can be defined as:

[0146]

[0147] In the formula: and wheel position The longitudinal and lateral coordinates of the grounding point in the ground coordinate system; and The coordinates of the origin of the elevation map; and These represent the spatial resolution of the elevation map in the vertical and horizontal directions, respectively. and These are pixel coordinates.

[0148] The elevation map is queried based on the pixel coordinates, the road surface height of each wheel contact point is read, and the terrain height change rate is calculated from the height difference between adjacent time points.

[0149]

[0150] In the formula: This is a function for querying elevation maps; For wheel position The road surface height corresponding to the grounding point; This is the simulated step size.

[0151] For two-layer terrain scenarios such as bridges, ramps, and interchanges, the elevations of both the ground layer and the bridge deck layer are included in the ground contact determination. The current contact layer is determined based on the distance relationship between the tire reference height and the elevations of the two terrain layers, which can be represented as:

[0152]

[0153] In the formula: and These are elevation maps for the ground level and the bridge deck level, respectively. For wheel position Reference height of the grounding point; This is the number of the layer currently in contact with this wheel position.

[0154] In one embodiment, in step S7, the simulation step size is divided into multiple sub-step sizes, and the vehicle state is iteratively integrated based on each sub-step size. For the vehicle attitude state, a quaternion update method is adopted, which multiplies the quaternion update matrix constructed by the current angular velocity with the quaternion of the previous moment to complete the attitude cross-sub-step update. After the integration of each sub-step, the updated quaternion is normalized to avoid numerical drift.

[0155] Specifically, in this embodiment, the tire forces, suspension forces, and gravity projections of each wheel are uniformly transformed into the vehicle coordinate system to obtain the resultant force and resultant torque required for single-step integration of the vehicle, which are as follows:

[0156]

[0157] In the formula: For wheel position The force transformation matrix from the coordinate system to the vehicle body coordinate system; For the first One tire force vector; This is the projection of gravity in the vehicle coordinate system; This is the position vector of the corresponding wheel position relative to the center of mass; The total torque generated by the suspension and additional systems.

[0158] Based on the simulation accuracy requirements, a simulation step size is divided into multiple sub-step sizes, and the vehicle state is iteratively integrated according to the sub-step sizes. The integration process of the high-stiffness multi-axle vehicle model can be written as follows:

[0159]

[0160] In the formula: For the step length; This represents the number of substeps within a single simulation step. For the first The first step is a large step. The state vector of each sub-step; For control input; For vehicle parameters; Input the terrain.

[0161] For the vehicle attitude state, the quaternion update process is combined with the integration process, and the integrated quaternion is normalized. Before and after normalization, we have:

[0162]

[0163] In the formula: For unnormalized posture quaternions; The normalized attitude quaternion.

[0164] In one embodiment, step S8, specifically includes the cyclic fusion optimization:

[0165] The physical model computation nodes in steps S1 to S7, including tire force calculation nodes, suspension force calculation nodes, multi-substep integration nodes, and terrain lookup nodes, are abstracted into computation graph nodes through data dependencies. The same computation nodes in the multi-substep loop are merged into a single compiled computation kernel function by operator fusion.

[0166] Vehicle parameters, control inputs, and states are tensorized, allowing a single kernel function call to complete the single-step evolution of all activated vehicles.

[0167] Specifically, in this embodiment, the single-vehicle dynamics model is extended into a batch tensor form, enabling the states, control variables, and parameters of different vehicles to be processed in parallel on the same computational graph, which can be uniformly denoted as:

[0168]

[0169] In the formula: This refers to the number of vehicles in a batch. For batch state tensors; To control the input tensors in batches; For batch parameter tensors; , and These are the state dimension, control dimension, and parameter dimension, respectively.

[0170] Based on the aforementioned batch tensor form, the single-step dynamics model and sub-step loop are reorganized into a unified batch state transition mapping. The batch update process is then as follows:

[0171]

[0172] In the formula: For inclusion Batch state transition operator updated step by step; For batch terrain input.

[0173] The next moment's state and the tire's force results are written into the output tensor, thus forming a simulation execution model that supports compilation export and runtime invocation:

[0174]

[0175] In the formula: For the simulation execution model in the first The output of the step, , and These are the output tensors for the normal load, longitudinal force, and lateral force of each wheel of the batch of vehicles.

[0176] In one embodiment, in step S9, the dimension of the mask matrix is ​​consistent with the dimension of the batch of vehicles; during parallel stepping, the Boolean value at the corresponding position in the mask matrix is ​​used to determine whether the vehicle is activated; for activated vehicles, the corresponding control commands and state data are passed into the compiled simulation execution model; for inactive vehicles, the computation graph branch of the vehicle is skipped, and the state data of the previous moment is directly assigned to the output result of the current moment.

[0177] Specifically, in this embodiment, the drive input, braking input, steering input, and vehicle validity input of a batch of vehicles are received and organized into a unified batch control tensor:

[0178]

[0179] In the formula: For batch drive torque input; For batch braking torque input; For batch redirection input; For batch settings or enable input.

[0180] Construct a mask matrix based on the vehicle activation state, perform parallel stepping only on currently active vehicles, and maintain the previous state for inactive vehicles, i.e.:

[0181]

[0182] In the formula: Activate the mask matrix for the vehicle; Update the batch dynamics operator.

[0183] The vehicle pose, velocity, and tire states after parallel stepping are output as a unified input for subsequent control, training, or scenario simulation tasks. The output can be written as:

[0184]

[0185] In the formula: Output the location of vehicles in batches; Output the attitude of a batch of vehicles; Output speed for a batch of vehicles; Outputs the angular velocity of a batch of vehicles.

[0186] A highly parallel and reconfigurable multi-axle vehicle dynamics simulation system includes:

[0187] The parameter parsing and configuration reconstruction module is used to read vehicle configuration parameters, parse the number of axles, tire mounting positions, vehicle mass, inertia tensor, suspension parameters, tire parameters and braking parameters, and construct a reconfigurable parameter model corresponding to the vehicle configuration.

[0188] The vehicle dynamics modeling module is used to construct the vehicle's lateral and longitudinal dynamics models, vertical dynamics models, and rigid body attitude dynamics models based on the vehicle's tire mounting positions and body center of gravity parameters; and to solve the vehicle's stable initial state under the current configuration and initial terrain conditions based on static equilibrium relationships and suspension-tire coupling relationships.

[0189] The tire-braking-terrain coupling calculation module is used to calculate the longitudinal and lateral forces of each tire based on the tire's longitudinal slip ratio, sideslip angle, and normal load; calculate the actual braking force of each wheel based on the friction dynamics model of the braking system; and calculate the road surface height information of each wheel's contact point based on the mapping relationship between the elevation map and the terrain coordinates, and complete the contact determination between the ground layer and the bridge deck layer in a double-layer elevation map scenario.

[0190] The dynamics integral update module is used to couple tire forces, suspension forces, gravity projections and rigid body moments, and update vehicle position, attitude, velocity and tire state using a multi-substep integration method to complete the vehicle's single-step dynamics evolution.

[0191] The high-parallelism simulation execution module is used to convert the single-step dynamic evolution process into a batch state transition mapping containing multi-sub-step loops, perform loop fusion optimization and compilation export on the computation graph, generate a compiled simulation execution model that supports variable batch size input tensors at runtime, and receive driving, braking and steering commands from a batch of vehicles, construct a mask matrix containing the vehicle activation states, perform parallel stepping on multiple vehicles simultaneously, and output the pose state, velocity state, tire state and tire force state of each vehicle.

[0192] Specifically, in this embodiment, as follows Figure 3 As shown in the illustration, this application also provides a highly parallel and reconfigurable multi-axle vehicle dynamics simulation system. In this embodiment, the system can be deployed on a server and achieves high-fidelity, high-concurrency simulation of multi-axle vehicles through the interaction between software-defined functional modules and underlying hardware acceleration. The system specifically includes the following five core functional modules:

[0193] 1. Parameter parsing and configuration reconstruction module

[0194] This module serves as the system's data entry point, primarily responsible for perceiving the vehicle's physical configuration. In its actual implementation, the module contains an object-oriented configuration file parsing engine. When the user inputs an XML or JSON configuration file containing information such as the number of axles, tire coordinates per axle, vehicle inertia, suspension, and braking characteristics into the system, the parsing engine uses regular expressions and DOM tree parsing techniques to transform the unstructured text data into parameter objects in memory.

[0195] Specifically, this module dynamically allocates wheel position arrays based on the physical constraints of the number of axles defined in the configuration file. Simultaneously, it concatenates and arranges the wheel position coordinates, stiffness parameters, damping parameters, vehicle mass, and inertia tensors from this array into a fixed-length one-dimensional parameter tensor. This tensor is directly mapped into the graphics processor's memory during system initialization. Subsequent computational kernel functions only need to index the vehicle's physical parameters for different axle counts through pointer offsets in the tensor. This design ensures seamless switching between multi-axle (three to six axles) vehicle models from the underlying data structure level, without requiring any reconstruction of the physical model code.

[0196] 2. Vehicle Dynamics Modeling Module

[0197] This module receives parameter tensors from the previous module and instantiates a six-DOF coupled vehicle model in the physics engine. Specifically, it employs a modular differential equation construction strategy. First, it constructs Newton-Euler translational and rotational differential equations at the origin of the vehicle coordinate system. Then, it incorporates the vertical displacement, roll angle, and pitch angle of the sprung mass into the suspension geometry to calculate the suspension deflection and support force at each wheel.

[0198] Of particular note is the embedded high-efficiency static equilibrium state solver in this module. Before the simulation officially starts, the solver establishes a set of linear equations based on the initial values ​​of vehicle weight and terrain elevation. These equations include the vertical equilibrium of the entire vehicle, the torque equilibrium around the X / Y axes, and the coupling between tire compression and suspension deformation. The equations cover the initial roll angle, pitch angle, and normal preload of each wheel. These results are directly fed into the dynamics integration module, thus avoiding divergence problems caused by non-physical initial conditions in the first frame of the simulation.

[0199] 3. Tire-Brake-Terrain Coupling Calculation Module

[0200] This module handles the most intensive nonlinear calculations. Within the tire calculation submodule, the system calculates the slip ratio and sideslip angle using the vehicle's current wheel center velocity and wheel speed state, and calculates the target force using a linear tire model and adhesion circle constraints. To improve stability, this module also predefines first-order hysteresis differential equations to simulate tire transient hysteresis effects and smooth tire force output.

[0201] In the braking submodule, a braking delay simulator is designed. It calculates the braking command received by the system through a first-order inertial element, which truly reflects the response lag of the air or hydraulic braking system. Finally, it calculates the actual braking torque through physical parameters and feeds it back to the wheel rotation equation.

[0202] In the terrain determination submodule, the system implements a multi-layer height indexing function. For two-layer terrain scenes, the module does not query a single elevation map, but instead opens two independent texture sampling channels to read the height values ​​of the ground layer and the bridge deck layer respectively. The module defines a height distance minimization algorithm, which compares the real-time vertical position of the tire with the elevation values ​​of the two layers, and selects the layer with the smallest absolute difference as the physical contact layer, thereby avoiding the clipping error that occurs with traditional single-layer elevation maps in overpass scenes.

[0203] 4. Dynamics Integral Update Module

[0204] This module acts as the "time-domain stepper" for the entire physical simulation system. In actual operation, this module divides the external main simulation step into multiple tiny sub-steps based on the user-defined model stiffness and accuracy requirements. In each sub-step loop, the module sequentially aggregates the resultant forces and torques from the tires, suspension, braking, and gravity, updates the linear velocity and angular velocity based on the rigid body momentum theorem, and then updates the center of mass position and attitude.

[0205] Specifically, the attitude update employs quaternion integration: this module constructs a quaternion update matrix using the current angular velocity, advances the current quaternion by one substep, and strictly performs quaternion normalization after each substep. This design effectively avoids gimbal lock-up and numerical drift that may occur with traditional Euler angle integration at large pitch or roll angles, ensuring the robustness of the vehicle's three-dimensional attitude evolution.

[0206] 5. High-parallelism simulation execution module

[0207] This module is the core scheduler for achieving high-performance parallelism throughout the system. In actual operation, this module does not directly execute the loop of the above steps via the CPU. Instead, it constructs a computation graph containing loop control nodes from all the computational nodes within the integral update module (including tire force integral, suspension force calculation, quaternion update, and terrain index). Through a built-in deep learning-style compiler, this module optimizes the computation graph by fusion of operators and loops, merging multiple discrete mathematical operations into a single-cycle computation kernel function deeply optimized for the GPU instruction set, and pre-compiling and exporting it as a low-level executable library for the platform.

[0208] When faced with a batch of vehicle inputs, this module concatenates the received control commands, states, and vehicle parameters of vehicle B into a three-dimensional tensor structure, and simultaneously constructs a Boolean mask tensor that matches the tensor's dimension. During the execution of specific acceleration kernel functions, the module performs conditional flow control based on the mask tensor. For vehicles with a mask bit of "1" active, it performs complete physics calculations; for vehicles with a mask bit of "0", it skips inefficient physics calculation steps and directly assigns the data from the previous frame to the output tensor. Finally, the module outputs the latest position, pose, velocity, and tire force data of the batch of vehicles in tensor form, allowing downstream autonomous driving perception algorithms or reinforcement learning environments to interact instantaneously.

[0209] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0210] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0211] It should be particularly noted that, through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, or of course, by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A high-parallelism reconfigurable multi-axle vehicle dynamics simulation method, characterized in that, Includes the following steps: Step S1: Read the vehicle configuration parameters, parse the number of axles, tire mounting positions, vehicle mass, inertia tensor, suspension parameters, tire parameters and braking parameters, and construct a reconfigurable parameter model corresponding to the vehicle configuration. Step S2: Based on the tire mounting position and vehicle center of gravity parameters in the reconfigurable parameter model, construct the vehicle's lateral and longitudinal dynamics model, vertical dynamics model, and rigid body attitude dynamics model to form a six-degree-of-freedom coupled dynamics model for a multi-axle vehicle. Step S3: Based on the static balance relationship of the vehicle's weight, tire support force, and suspension support force, as well as the suspension-tire coupling relationship, solve for the steady-state initial state vector of the vehicle under the initial terrain conditions. Step S4: Based on the tire longitudinal slip ratio, sideslip angle and normal load, calculate the basic longitudinal force and basic lateral force of each wheel, introduce the attachment circle constraint to uniformly scale the longitudinal and lateral forces, and use the tire transient hysteresis state to dynamically correct the tire force to obtain the actual tire force participating in the vehicle dynamics solution. Step S5: Based on the friction dynamics model of the braking system, calculate the actual braking torque of each wheel edge, and input the braking torque, driving torque, and tire longitudinal force into the wheel rotation equation to update the wheel speed state of each wheel. Step S6: Based on the elevation map and terrain coordinate mapping relationship, calculate the road surface height information of each wheel contact point. For a two-layer terrain scene that includes bridges or overpasses, read the elevation information of the ground layer and the bridge deck layer at the same time. Based on the distance relationship between the tire reference height and the height of the two terrain layers, determine the current contact layer and extract the road surface height of the corresponding layer. Step S7: Couple the tire forces, suspension forces, gravity projections and rigid body moments of each wheel, and update the vehicle position, attitude, velocity and tire state using a multi-substep integration method to complete the vehicle single-step dynamics evolution. Step S8: Convert the single-step dynamic evolution process into a batch state transition mapping containing multi-sub-step loops, construct the corresponding simulation calculation graph, perform loop fusion optimization and compilation export on the simulation calculation graph, and generate a compiled simulation execution model that supports variable batch size input tensors at runtime. Step S9: Receive driving, braking, and steering commands from a batch of vehicles, construct a mask matrix containing the vehicle activation states, drive the compiled simulation execution model to perform parallel stepping only on the currently active valid vehicles based on the mask matrix, keep the previous state unchanged for inactive vehicles, and output the pose state, speed state, tire state, and tire force state of each vehicle.

2. The high parallelism reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, Step S1 specifically includes: Based on the configuration relationship that each axle includes left and right wheel sets, and each wheel set includes at least one tire, the total number of tires in the vehicle is determined, and the mounting position vector of each wheel in the vehicle coordinate system is generated. According to the wheel position sequence, the vehicle mass, sprung mass, unsprung mass of each wheel, vehicle body inertia tensor, suspension stiffness matrix, suspension damping matrix, tire vertical stiffness matrix, and tire vertical damping matrix are organized into a unified parameter tensor to adapt to the unified parameter loading and state initialization of three-axle, four-axle, five-axle, and six-axle vehicles.

3. The high parallelism reconfigurable multi-axle vehicle dynamics simulation method of claim 1, wherein, In step S2, the six-degree-of-freedom coupled dynamics model includes a vehicle translational dynamics model, a vehicle rotational dynamics model, a posture kinematics model, and a suspension vertical coupling relationship. The vehicle translational dynamics model incorporates longitudinal velocity, lateral velocity, and vertical velocity into the force balance relationship of the vehicle's center of mass; the vehicle rotational dynamics model establishes the equations for roll, pitch, and yaw motion based on the torque balance relationship of the vehicle around its center of mass. The vertical coupling relationship of the suspension incorporates the vertical displacement of the sprung mass, the roll angle, and the pitch angle into the calculation formulas for the suspension deflection and suspension force of each wheel.

4. The high parallelism reconfigurable multi-axle vehicle dynamics simulation method of claim 1, wherein, Step S3 specifically includes: Based on the lateral and longitudinal lever arms of the normal loads of each wheel relative to the center of mass, the moment balance relationships in the roll and pitch directions are established respectively. Combined with tire compression, suspension deflection and terrain elevation, the vertical force balance equations, roll moment balance equations, pitch moment balance equations and tire-suspension coupling relationships of each wheel are combined into a system of linear equations, and the steady-state initial state vector is obtained by solving the system.

5. The high parallelism reconfigurable multi-axle vehicle dynamics simulation method of claim 1, wherein, Step S4 specifically includes: Based on the velocity decomposition results at the wheel center, the longitudinal slip ratio and sideslip angle of each wheel are obtained. The basic longitudinal force and basic lateral force of each wheel are calculated based on the tire stiffness parameters. The tire-road adhesion constraint is introduced to uniformly scale the longitudinal and lateral tire forces. Furthermore, the tire transient hysteresis state is introduced to obtain the actual tire forces participating in the solution of the vehicle dynamics.

6. The highly parallel reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, Step S5 specifically includes: The frictional torque at each wheel's brake disc is calculated based on the braking input and brake structure parameters. The braking friction coefficient is then introduced into the first-order dynamic hysteresis equation to simulate the hydraulic build-up, friction pair response, and transmission chain hysteresis process in the braking system. The braking torque, driving torque, and tire longitudinal force are then written into the wheel rotation equations for each wheel, and the angular acceleration of the wheel's equivalent moment of inertia is integrated.

7. The highly parallel reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, In step S6, the contact layer in the dual-terrain scene is determined by: using the plane coordinates of each wheel's contact point in the ground coordinate system, and simultaneously querying the ground elevation map and the bridge elevation map, the difference between the tire reference height and the ground layer and the difference between the tire reference height and the bridge layer are obtained respectively. Select the layer with the smaller absolute value of the difference as the current contact layer.

8. The highly parallel reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, In step S7, the simulation step size is divided into multiple sub-step sizes, and the vehicle state is iteratively integrated based on each sub-step size. For the vehicle attitude state, the quaternion update method is adopted, which multiplies the quaternion update matrix constructed by the current angular velocity with the quaternion of the previous moment to complete the attitude cross-sub-step update. After the integration of each sub-step, the updated quaternion is normalized to avoid numerical drift.

9. The highly parallel reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, In step S8, the cyclic fusion optimization specifically includes: The physical model computation nodes in steps S1 to S7, including tire force calculation nodes, suspension force calculation nodes, multi-substep integration nodes, and terrain lookup nodes, are abstracted into computation graph nodes through data dependencies. The same computation nodes in the multi-substep loop are merged into a single compiled computation kernel function by operator fusion. Vehicle parameters, control inputs, and states are tensorized, allowing a single kernel function call to complete the single-step evolution of all activated vehicles.

10. The highly parallel reconfigurable multi-axle vehicle dynamics simulation method according to claim 1, characterized in that, In step S9, the dimension of the mask matrix is ​​consistent with the dimension of the batch of vehicles. When performing parallel stepping, the Boolean value at the corresponding position in the mask matrix is ​​used to determine whether the vehicle is activated. For activated vehicles, the corresponding control commands and state data are passed into the compiled simulation execution model. For inactive vehicles, the computation graph branch of the vehicle is skipped, and the state data of the previous moment is directly assigned to the output result of the current moment.

Images