An adaptive model predictive control method and system with uncertainty perception capability

Abstract

The present application relates to the intelligent network connected automobile and the automatic driving technical field, specifically relates to a kind of self-adapting model predictive control method and system with uncertainty perception ability, method includes: through deep integration network quantification cognitive uncertainty;Uncertainty and environmental state are fused to form risk perception state;Weight factor and intention action are output by reinforcement learning;MPC reconfigures cost function online;Calculate physical feasible region;Intention action is corrected by safety projection.The system includes uncertainty measurement module, risk perception decision module, model predictive control and safety projection module.The present application can perceive scene uncertainty in real time, dynamically adjust control strategy, ensure that control instruction is safe and feasible, significantly improve the safety and robustness of automatic driving in complex unknown scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicles and autonomous driving technology, specifically to an adaptive model predictive control method and system with uncertainty perception capability. Background Technology

[0002] As autonomous driving technology evolves to higher levels, decision-making and planning systems face two core challenges: first, how to make safe and efficient driving decisions in complex and dynamic environments; and second, how to ensure the feasibility of decision results under the physical constraints of the vehicle.

[0003] Currently, there are two main technical routes in the industry: Route 1: End-to-end decision-making based on deep reinforcement learning (DRL). DRL learns the direct mapping from perception to control through trial and error with the environment, and has strong exploration capabilities and scenario generalization potential. However, the pure DRL method has the following defects: (1) the output action may violate vehicle dynamics constraints and produce physically unexecutable instructions; (2) when facing scenarios outside the training set, the model is prone to overconfident erroneous predictions, and its "black box" characteristics make the decision-making process difficult to explain; (3) it lacks a strict safety guarantee mechanism. Route 2: Optimal control based on model predictive control (MPC). MPC can generate the optimal control sequence under the premise of satisfying vehicle dynamics constraints and collision avoidance constraints by solving constrained finite-domain optimization problems online. However, the cost function weight parameters of traditional MPC are usually preset and kept fixed by engineers based on experience, which is difficult to adapt to the rapidly changing traffic flow scenarios in urban roads (such as the different requirements for weight configuration between aggressive following during congestion and smooth and energy-saving driving during high-speed cruising).

[0004] However, the "exploration capability" of DRL and the "safety constraints" of MPC cannot be organically integrated. Existing solutions are mostly simple series or switch-on switching, without establishing a deep mutual empowerment mechanism between the two. Existing systems cannot explicitly perceive their own cognitive uncertainty about the current scenario, and therefore cannot adaptively switch driving styles between "known scenarios" and "unknown scenarios". The intention commands output by DRL lack a strict safety projection mechanism based on the physical feasible domain, making it difficult to pass automotive-grade functional safety certification.

[0005] Therefore, there is an urgent need for an adaptive decision-making and control method that can quantify its own cognitive uncertainty, dynamically adjust the control strategy based on the uncertainty, and ensure that the output command always lies within the physically feasible domain. Summary of the Invention

[0006] To overcome the aforementioned technical problems in the prior art, this invention provides an adaptive model predictive control method and system with uncertainty perception capability, in order to solve the technical problems in the prior art where the exploration capability of deep reinforcement learning (DRL) and the safety constraints of model predictive control (MPC) cannot be deeply integrated, the system lacks cognitive uncertainty self-perception capability, and the decision instructions lack physical safety fallback mechanisms.

[0007] To achieve the above objectives, this invention provides an adaptive model predictive control method with uncertainty perception capability, comprising the following steps: Step S1: Cognitive uncertainty quantification stage, using a deep integrated scene understanding network composed of M models to process current driving environment observation data in parallel, calculating the statistical divergence between the output results of the M models, and generating a quantitative index value representing the cognitive uncertainty of the current scene; the magnitude of the quantitative index value is used to measure the familiarity and judgment of the current scene, with a larger value indicating a lower level of familiarity; Step S2: Risk perception state construction stage, using the quantitative index value as an explicit risk feature dimension, and concatenating it with the environmental state vector obtained after encoding the driving environment observation data to form a risk perception state vector that integrates scene semantic information and system self-awareness confidence; Step S3: Adaptive strategy generation stage, inputting the risk perception state vector into a deep reinforcement learning strategy network, which outputs the following two sets of signals in parallel: Weight adjustment factor: used to dynamically adjust the weights of each penalty term in the cost function of the model predictive controller. The coefficient vector; Intended action command: a preliminary control quantity representing the optimal operation intention in the current driving scenario; Step S4: Controller parameter dynamic reconstruction stage, the model predictive controller receives the weight adjustment factor and reconstructs its own cost function online according to the weight adjustment factor, so that the weight of the safety-related penalty term in the cost function monotonically increases with the increase of the quantification index value; Step S5: Physical feasible region calculation stage, the reconstructed model predictive controller, based on the current vehicle state, road surface adhesion conditions, and actuator saturation constraints, solves online the set of physically feasible actions that satisfy vehicle dynamic constraints and collision avoidance constraints in the future prediction time domain, denoted as the feasible region; Step S6: Safety projection and command correction stage, it is determined whether the intended action command generated in step S3 is located within the feasible region: if it is located within the feasible region, the intended action command is directly output as the final control command; if it is not located within the feasible region, a safe action command that is geometrically closest to the intended action command is solved on the boundary of the feasible region, and the safe action command is output as the final control command to the vehicle actuator.

[0008] Preferably, the formula for calculating the quantitative index value in step S1 is: in, This represents the value of the quantitative indicator. This indicates that the m-th model is based on the input observation data. The output scene feature encoding vector, This represents the mean of the output vectors of the M models; the larger the value of the quantification index, the greater the discrepancy in the understanding of the current scene among the sub-models in the deep integrated scene understanding network, and the higher the uncertainty of the system's cognition of the current scene.

[0009] Preferably, the risk perception state vector in step S2 is constructed by concatenating the environmental state vector obtained after feature extraction from the environmental observation data with the quantified index value to form a risk perception state vector with expanded dimensions. The risk perception state vector enables the deep reinforcement learning policy network to explicitly perceive the system's confidence level in the current scene.

[0010] Preferably, the deep reinforcement learning policy network adopts a Soft Actor-Critic architecture, and the reward function used in the offline training phase is... It has the following forms: in, Positive reward items that characterize traffic efficiency; To characterize the negative penalty for dangerous events, The reward function is designed to adjust the reward coefficient so that the deep reinforcement learning policy network learns the following behavioral strategy during training: when the quantification index value increases, it actively reduces the pursuit of efficiency and prioritizes avoiding safety penalties.

[0011] Preferably, the cost function after the model prediction controller is reconstructed in step S4. It has the following forms: in, Let be the state error vector at the k-th prediction step. To control the increment vector, To predict the length of the time domain, The state weight matrix is... The control weight matrix; the state weight matrix With control weight matrix The weight adjustment factor generated in step S3 is dynamically determined.

[0012] Preferably, the state weight matrix The weighting coefficients corresponding to the following distance error and the weighting coefficients corresponding to the lateral position deviation are both positively correlated with the quantified index value; this enables the model prediction controller to automatically execute a more conservative driving strategy when the system's perception of the current scene becomes more uncertain. The driving strategy is manifested in at least one of the following behaviors: increasing the safe following distance, reducing the steering response speed, and narrowing the feasible speed range.

[0013] Preferably, the step S6 of finding a safe action command that is geometrically closest to the intended action command adopts the following quadratic programming problem: in, The intention action command generated in step S3, The set of physical actions generated in step S5. This represents the square of the Euclidean distance between the action vector and the intention action vector.

[0014] Preferably, the M models in the deep integrated scene understanding network are homogeneous heterogeneous neural networks with the same network structure but different initialization parameters, or heterogeneous neural networks with different network structures.

[0015] Accordingly, the present invention also provides an adaptive model predictive control system with uncertainty perception capability, comprising: The uncertainty measurement module is configured to perform the function of step S1 as described in claim 1, and is equipped with a deep integrated scene understanding network consisting of M models, for receiving driving environment observation data and outputting a quantitative index value of cognitive uncertainty. The risk perception and decision-making module is configured to perform the functions of steps S2 to S3 as described in claim 1, and is equipped with a deep reinforcement learning policy network for fusing the quantitative index value with the environmental state vector and outputting a weight adjustment factor and intention action command. The model prediction control and safety projection module is configured to perform the functions of steps S4 to S6 as described in claim 1. It is equipped with a model prediction controller that can reconstruct the cost function online. The controller is used to reconstruct the control strategy according to the weight adjustment factor, calculate the physical action domain online, perform safety verification and projection correction on the intention action command, and output the final safety control command.

[0016] Preferably, the model prediction control and safety projection module is further configured to: when the intended action command does not fall within the physical action domain, solve a quadratic programming problem on the boundary of the physical action domain to determine the safe action command with the closest Euclidean distance to the intended action command.

[0017] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the adaptive model predictive control method with uncertainty perception capability provided by the present invention.

[0018] The present invention has at least the following technical effects through the technical solution provided by the present invention: This invention, through variance calculation of deeply integrated networks, introduces "cognitive uncertainty" as an explicit differentiable signal into the decision-making link for the first time, enabling the system to quantify its "degree of control" over the current scenario and giving the system "self-awareness." Through the training mechanism of the reward function and the monotonically increasing mapping relationship between the MPC weight matrix and the cognitive uncertainty quantification index, the system automatically becomes conservative (increasing following distance and reducing speed) when encountering long-tailed unknown scenarios, and resumes efficient passage in familiar scenarios, achieving a smooth and stepless switching of driving style. By "safely projecting" the DRL intention command through the physical feasible domain solved online by MPC, it ensures that the final output control command always lies within the intersection of the vehicle dynamics envelope and collision constraints. This mechanism strictly constrains the "imagination" of DRL within the "possibilities" boundary of the physical world, solving the deployment problem of end-to-end learning models in safety-critical systems; the weight adjustment factor output by DRL provides scene-adaptive optimization parameters for MPC, avoiding the lag of manual parameter tuning; at the same time, the intent command output by DRL can be used as the initial value for hot start of MPC solution, significantly reducing the number of optimization iterations and meeting the real-time requirements of millisecond-level control cycle of autonomous driving. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is the overall flowchart of the adaptive model predictive control method with uncertainty perception capability provided by the present invention; Figure 2 This is a schematic diagram of the structure of the deep integrated scene understanding network in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the deep reinforcement learning policy network (SAC architecture) in an embodiment of the present invention; Figure 4 This is a module architecture diagram of the adaptive model predictive control system with uncertainty perception capability provided by the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] In this invention, the terms "system" and "network" are used interchangeably. "Multiple" refers to two or more; therefore, in this invention, "multiple" can also be understood as "at least two." "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / ", unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, it should be understood that in the description of this invention, terms such as "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or order.

[0022] Please see Figure 1 This invention provides an adaptive model predictive control method with uncertainty perception capability. This method runs in the decision planning module of a vehicle's autonomous driving domain controller or central computing platform, and cyclically executes the following steps S1 to S6 at a fixed control cycle (e.g., 50 milliseconds, i.e., 20Hz), specifically including: Step S1: Using a deep integrated scene understanding network composed of M models, the current driving environment observation data is processed in parallel, the statistical divergence between the output results of the M models is calculated, and a quantitative index value representing the uncertainty of the current scene cognition is generated; the magnitude of the quantitative index value is used to measure the familiarity and judgment of the current scene, and the larger the value, the lower the familiarity. Step S2: The quantitative index value is used as an explicit risk feature dimension and is concatenated with the environmental state vector obtained after encoding the driving environment observation data to form a risk perception state vector that integrates scene semantic information and system self-awareness confidence. Step S3: Input the risk-aware state vector into the deep reinforcement learning policy network, which outputs the following two sets of signals in parallel: Weight adjustment factor: A coefficient vector used to dynamically adjust the weights of each penalty term in the cost function of the model predictor controller; Intended action command: The initial control quantity representing the optimal operating intention in the current driving scenario; Step S4: The model prediction controller receives the weight adjustment factor and reconstructs its own cost function online according to the weight adjustment factor, so that the weight of the safety-related penalty term in the cost function increases monotonically as the value of the quantification index increases. Step S5: The reconstructed model predictive controller, based on the current vehicle state, road surface adhesion conditions, and actuator saturation constraints, solves online the set of physically feasible actions that satisfy vehicle dynamics constraints and collision avoidance constraints in the future prediction time domain, denoted as the feasible region. Step S6: Determine whether the intentional action command generated in step S3 is within the feasible domain: If it is within the feasible domain, the intention action instruction will be directly output as the final control instruction; If it is not located within the feasible region, then on the boundary of the feasible region, a safe action command that is geometrically closest to the intended action command is solved, and the safe action command is output as the final control command to the vehicle actuator.

[0023] In this embodiment of the invention, step S1 is the cognitive uncertainty quantification stage. The core purpose of this step is to enable the system to answer the question, "How confident am I in the current scenario?" Specifically, it includes: S1.1: Primarily responsible for acquiring environmental observation data, obtaining current driving environment observation data from the vehicle perception module, including but not limited to: acquiring a bird's-eye view feature map with a size of 256×256×7 (the 7 channels represent: drivable area mask, lane line probability map, road edge probability map, static obstacle occupancy grid, dynamic obstacle (vehicle, pedestrian, cyclist) occupancy grid and its velocity component); the bird's-eye view feature map is generated in real time by the upstream multi-sensor fusion module (integrating data from the forward-looking camera, surround-view camera, LiDAR, and millimeter-wave radar); acquiring target-level state vectors, containing structured information of up to 32 traffic participants around the vehicle, each target represented by a 12-dimensional vector: [target ID, type (0-pedestrian, 1-cyclist, 2-car, 3-truck / bus), relative position x, relative position y, relative speed v_x, relative speed v_y, relative acceleration a_x, relative acceleration a_y, heading angle, length, width, existence probability].

[0024] S1.2: Primarily performs forward inference for the deep integrated scene understanding network. In this embodiment, such as... Figure 2As shown, M=3 hierarchical visual Transformers (Swin Transformer networks) based on shift windows with identical structures but different randomly initialized seeds are used as sub-models of the deep ensemble scene understanding network. The Swin Transformer is particularly suitable for bird's-eye view feature processing in autonomous driving due to its excellent global context modeling ability and linear computational complexity in computer vision tasks. Specifically, the bird's-eye view feature map (256×256×7) obtained by S1.1 is input into three encoders with the SwinTransformer Tiny architecture. The specific structural parameters of each encoder are as follows: Patch Partition: Divides the input image into 4×4 patches, each patch being flattened into a 16×7=112 dimensional vector.

[0025] Stage 1: The Linear Embedding layer maps 112 dimensions to 96 dimensions, followed by two SwingTransformerBlocks with a window size of 7×7.

[0026] Stage 2: The Patch Merging layer performs a 2x downsampling, increasing the feature dimension to 192 dimensions, followed by two SwingTransformer Blocks.

[0027] Stage 3: The Patch Merging layer performs a 2x downsampling, increasing the feature dimension to 384 dimensions, followed by 6 SwingTransformer Blocks.

[0028] Stage 4: The Patch Merging layer performs a 2x downsampling, increasing the feature dimension to 768 dimensions, followed by two SwingTransformer Blocks.

[0029] After Stage 4 output, a 768-dimensional scene feature encoding vector is obtained through global average pooling. The three sub-models then independently perform forward inference, each outputting three 768-dimensional vectors, denoted as... , , .

[0030] S1.3: Quantitative indicators of cognitive uncertainty. First, calculate the mean of the three output vectors using the following formula: Then, the quantitative indicator value is calculated using the following formula. : Among them, and and Both are vectors. This represents the element-wise sum of squares of the vector differences, i.e., the square of the Euclidean distance.

[0031] This quantitative indicator value The magnitude of the quantization index reflects the consistency of the three sub-models' understanding of the same input scene. In typical training set coverage scenarios (such as highways on sunny days), the quantization index value is usually in the range [0, 0.5]. In long-tail scenarios (i.e., complex edge scenarios with low probability of occurrence, few samples but many types; such as irregular obstacles on the road after rain, or irregular cone combinations in construction areas), the quantization index value can jump to [2.0, 5.0] or even higher.

[0032] In this embodiment of the invention, step S2 is the risk perception state construction stage, specifically as follows: S2.1: Obtain the environmental state vector. Flatten the target-level state vector obtained in step S1.1. 32 targets × 12 dimensions / target = 384-dimensional vector. If the actual number of targets is less than 32, fill the remaining positions with zeros. This 384-dimensional vector is denoted as... Simultaneously, the vehicle's status information (including vehicle speed, acceleration, steering wheel angle, yaw rate, and remaining battery SOC, totaling 5 dimensions) is stitched together. In this context, a 389-dimensional basic environmental state vector is formed. .

[0033] S2.2: Feature fusion and splicing, combining the quantified index values ​​calculated in step S1.3. (1-dimensional scalar) concatenated to the basic environment state vector The (389-dimensional) endpoints form a risk perception state vector with a dimension of 390: .

[0034] Furthermore, step S3 is the adaptive policy generation stage. This step utilizes a pre-trained deep reinforcement learning policy network, such as... Figure 3 As shown, the risk perception state vector is mapped to a control strategy. Specifically: S3.1: Policy Network Architecture. In this embodiment, the deep reinforcement learning policy network adopts the Soft Actor-Critic (SAC) architecture because it has good exploration capabilities and training stability in continuous action space control tasks. Specific parameters are as follows: Actor Network (Policy Network): Input layer: 390 dimensions (corresponding to) ) Fully connected layer 1: 1024 neurons, activation function is ReLU Fully connected layer 2: 1024 neurons, activation function is ReLU Fully connected layer 3: 512 neurons, activation function is ReLU Output layer: consists of two parallel fully connected headers—mean header: outputs a 15-dimensional mean vector. Standard deviation head: Outputs a 15-dimensional standard deviation vector. The positive value is ensured by the Softplus function; The actual action is calculated using the following formula: in, Represents element-wise product. This is a 15-dimensional random noise vector sampled from a standard normal distribution. It is a 15-dimensional identity matrix. For the original sampling action, Let be the final output action vector compressed to the interval [-1, 1] by the hyperbolic tangent function.

[0035] Critic Network (Value Network): A dual Q-network structure to alleviate the overestimation problem; each Q-network: input 390-dimensional state + 15-dimensional action (405-dimensional after concatenation), outputs a single Q-value after passing through two 1024-dimensional fully connected layers.

[0036] S3.2: Output Signal Analysis. After analysis, the 15-dimensional vector output by the Actor network corresponds to the following four types of signals with different physical meanings: Dimensions 1 to 6 together constitute a six-dimensional weight adjustment factor vector. After linear mapping, the value range is [0.1, 5.0] (where the first dimension...). Represents the longitudinal velocity tracking error weight scaling factor; 2nd dimension The third dimension represents the lateral positional deviation weight scaling factor. This represents the following distance error weight scaling factor; the 4th dimension. This represents the steering angular rate weight scaling factor, which is related to comfort constraints; the 5th dimension. This represents the acceleration rate of change weight scaling factor, which is related to comfort constraints; the 6th dimension. The 7th dimension represents the target vehicle speed correction factor, which is related to the collision potential field weight scaling factor and is associated with safety constraints. The unit is meters per second, and after linear mapping, its value ranges from [-5.0, 5.0]; the 8th dimension represents the expected following distance correction. The unit is seconds, and after linear mapping, the value ranges from [-5.0, 1.0]. Dimensions 9 to 15 together constitute the lateral offset sequence (in meters) of the reference trajectory points for the next 7 control cycles. Each dimension corresponds to the expected lateral offset of the trajectory at the future times of 0.5 seconds, 1.0 seconds, 1.5 seconds, 2.0 seconds, 2.5 seconds, 3.0 seconds, and 3.5 seconds, respectively, and after linear mapping, the value ranges from [-3.0, 3.0]. This offset is defined in the Frenet coordinate system and represents the lateral distance relative to the reference centerline; positive values ​​indicate a leftward offset, and negative values ​​indicate a rightward offset.

[0037] From the Actor network output, the mapped reference trajectory point sequence is used to construct the initial intention action command. In this embodiment, Defined as the expected trajectory sequence for the next N steps (aligned with the MPC prediction time domain), it is expressed as: .

[0038] S3.3: Offline training of the policy network. The training of the policy network is completed offline in the simulation environment, using the following specially designed reward function: in, Positive rewards for improving traffic efficiency encourage vehicles to travel at near-target speeds and complete tasks. This is a negative penalty for dangerous events; The adjustment coefficient; the key training mechanism: in a simulation environment, by artificially injecting perceptual noise and generating extreme adversarial scenarios, the system undergoes different amplitudes of [various effects] during training. The coefficient of the security penalty term in the reward function dynamically amplifies the cost of a security event. After millions of iterations, the policy network will learn: when When the impact is significant, proactively adopt conservative actions to avoid amplified safety penalties; when When they are younger, they actively pursue efficiency rewards.

[0039] In this embodiment of the invention, step S4 is the dynamic reconfiguration stage of the controller parameters, where the model prediction controller receives the weight adjustment factor vector output in step S3. The cost function is reconstructed online. In this embodiment, MPC uses a vehicle monorail dynamics model as the prediction model, and the state vector... , representing lateral deviation, heading angle deviation, speed error, and following distance error, respectively. Control vector , representing the front wheel steering angle command and the longitudinal acceleration command, respectively. Then the cost function of MPC... The specific form is as follows: Among them, the prediction time domain The value is 20, the control cycle is 0.1s, and the total prediction time is 2.0 seconds. The state weight matrix is... To control the weight matrix; all weight adjustment factor vectors are generated in step S3. Determined dynamically.

[0040] Furthermore, the reconstruction rule for the state weight matrix is ​​as follows: the weight matrix includes the lateral deviation weights (dynamic reconstruction formula: base value × ... The base value is 10.0, and the speed error weight (dynamic reconstruction formula: base value × ...) (Base value 5.0), following distance error weight (dynamic reconstruction formula: base value × ... (Base value is 20.0), steering angle weight (dynamic reconfiguration formula: base value × ... (Base value 1.0) and acceleration weight (dynamic reconstruction formula: base value × (The base value is 0.5). Due to the design of the reward function during the training phase, when step S1 outputs... When the value increases, the output of the policy network... (Safety-related) will tend to increase. Therefore, in the MPC cost function reconstructed in step S4, the weight of the safety-related penalty term will increase accordingly. The monotonically increasing value causes the controller to automatically exhibit a conservative trend.

[0041] In this embodiment of the invention, step S5 is the physical feasible region calculation stage, in which the reconstructed model predictive controller solves online for the set of physical feasible actions, i.e., the feasible region. The feasible region is defined as the set of predicted trajectories corresponding to all control sequences that simultaneously satisfy vehicle dynamics constraints, actuator physical saturation constraints, collision avoidance constraints, and road boundary constraints within the future prediction time domain. Specifically, the feasible region is defined by the following four types of constraints: Constraint 1: Vehicle Dynamics Constraint. This embodiment employs a three-degree-of-freedom single-track dynamics model, which describes the vehicle's motion response in the longitudinal, lateral, and yaw directions, and considers the nonlinear lateral force characteristics between the tire and the road surface. The tire lateral force is calculated based on the Pacejka magic formula tire model, where the road adhesion coefficient is provided in real time by the upstream road state estimation module; this constraint ensures that the trajectory generated by the controller is physically a motion that the vehicle can actually perform.

[0042] Constraint 2: Actuator Saturation Constraint. This constraint limits the control output and its rate of change of the controller to the physical limits of the vehicle's actuators. Specifically, it includes: maximum amplitude and maximum speed limits for the front wheel steering angle, maximum amplitude limits for longitudinal acceleration, and limits on the rate of change of longitudinal acceleration (i.e., jerk constraints). These constraints ensure that control commands do not exceed the actual response capabilities of the steering, braking, and drive systems.

[0043] Constraint 3: Collision Avoidance Constraint. This constraint requires that at every prediction step within the future prediction time domain, the vehicle must maintain a distance of at least a preset safe distance from all detected traffic participants (including vehicles, pedestrians, cyclists, and static obstacles). The safe distance is set by comprehensively considering the static minimum clearance and a dynamic allowance proportional to relative speed to ensure sufficient safety margin under extreme conditions such as emergency braking.

[0044] Constraint 4: Road Boundary Constraint. This constraint requires that the predicted trajectory of the vehicle always remain within the drivable area of ​​the current road. The drivable area is defined by road boundary elements such as lane lines, shoulders, and guardrails, ensuring that the vehicle does not leave the road surface or enter impassable areas such as oncoming lanes.

[0045] The model predictive control solver searches for control sequences that minimize the reconstructed cost function in step S4 within the feasible space that satisfies all the above constraints. All candidate trajectories evaluated and verified to satisfy all constraints during the solution process constitute the set of physically feasible actions. In practical deployments, to reduce the online computing burden while meeting real-time requirements, the feasible domain... It can be approximated as a confidence ellipsoidal region centered on the current optimal solution. The shape of the ellipsoid is determined by the curvature characteristics of the cost function at the optimal solution, thereby significantly reducing computation time while ensuring solution accuracy.

[0046] In this embodiment of the invention, step S6 is the security projection and instruction correction stage, in which the intentional action instruction generated in step S3 is determined. Whether it is safe and executable is determined. If it is not safe, it is corrected based on the nearest neighbor projection method of the physical feasible region boundary, thereby constructing a physical safety barrier for the final output of the system. Specifically: S6.1: Perform attribution determination and check the intended action instruction. Does it satisfy all the constraints defined in step S5? The basis for the judgment is: The predicted trajectory is then substituted into the vehicle dynamics model and environmental constraint model to verify whether it violates actuator saturation limits, collision safety distance requirements, road boundary limits, and vehicle physical motion limits. If all the above constraints are satisfied, and the residuals calculated by the dynamics model are within a preset small tolerance range, then the intended action command is determined. Located in the set of physical actionable actions If any constraint is not met, the intended action instruction is determined to be outside the feasible domain; conversely, if any constraint is not met, the intended action instruction is determined to be outside the feasible domain.

[0047] S6.2: Direct output scenario: If the judgment result of step S6.1 is "yes", that is, the intended action instruction is executed. If it has been verified to be completely safe and physically executable, the system will not modify it further and will directly execute it. This serves as the final control command output. In specific operation, the front wheel steering angle command and longitudinal acceleration command corresponding to the first step are extracted from the intended action command sequence, encoded according to the vehicle communication protocol, and sent to the electric power steering system and the power and braking system for execution, respectively.

[0048] S6.3: Safety projection correction scenario, if the judgment result of step S6.1 is "no", that is, the intended action command... Determined to be beyond the set of physically possible actions If the boundary of the road contains, for example, an excessive steering request that would cause the vehicle to go beyond the road boundary, or an excessive acceleration request that would pose a collision risk, the system will initiate a safety projection correction procedure. In this case, the system does not simply discard the intent instruction, but rather, based on the nearest neighbor safety projection principle, within the feasible region... Find an optimal alternative instruction. The mathematical essence of this search process is to solve the following quadratic programming problem, the geometric meaning of which is to determine, among all safe alternative actions, one that matches the original intention. The closest action: in, This indicates the revised safety action instructions. This represents the original intention action command generated in step S3. This represents the set of physical possible actions generated in step S5. It represents the square of the Euclidean distance between two vectors.

[0049] To efficiently solve this quadratic programming problem on an in-vehicle embedded computing platform, this embodiment employs numerical optimization techniques such as projective gradient descent. The solution process mainly includes two sub-steps: iteratively executing constrained projection and gradient descent along the tangent space, until the calculation result converges to the nearest point on the feasible region boundary. Finally, the safety action command corrected by the projection is generated. As the final control command output, it respects the decision-making intent of deep reinforcement learning to the greatest extent while strictly ensuring the physical feasibility of the output signal and driving safety.

[0050] S6.4: Control command issuance. The final safety control command is encoded into a CAN message and sent to the vehicle chassis domain controller through the gateway. This drives the electric power steering system (EPS), electronic stability control system (ESC), and motor controller to perform corresponding steering, braking, and driving operations. After completing the current control cycle, when the next cycle arrives, return to step S1 and repeat the process.

[0051] Please see Figure 4 Based on the same inventive concept, embodiments of the present invention provide an adaptive model predictive control system with uncertainty perception capability. This system is deployed on a vehicle's high-performance computing platform or autonomous driving domain controller and comprises three core software modules: 1. Uncertainty Measurement Module: This module runs on the graphics processing unit or neural network processing unit of the domain controller; it internally deploys a deep integrated scene understanding network consisting of M Swin Transformer Tiny networks. This module subscribes to bird's-eye view feature map data from the perception fusion module via shared memory, performs the inference calculation in step S1, and outputs a quantitative index value of cognitive uncertainty. .

[0052] 2. Risk Awareness and Decision-Making Module: This module runs on the GPU of the domain controller or a high-performance CPU core, and internally deploys a deep reinforcement learning policy network (Actor network) with a Soft Actor-Critic architecture. This module subscribes to data from the Uncertainty Measurement module. The values ​​and the target-level state vector from the perception module are used to perform feature concatenation in step S2 and policy network forward inference in step S3, outputting a weight adjustment factor vector. and intentional action instructions .

[0053] 3. Model Predictive Control and Secure Projection Module: This module runs on the real-time CPU core of the domain controller. It contains a nonlinear model predictive controller capable of online cost function reconstruction and a quadratic programming solver. This module performs cost function reconstruction in step S4, feasible region calculation in step S5, and secure projection verification in step S6.

[0054] Inter-module communication mechanism: The three modules exchange data with low latency through the DDS or SOME / IP middleware inside the domain controller.

[0055] Furthermore, embodiments of the present invention also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the present invention.

[0056] This invention, through variance calculation of deeply integrated networks, introduces "cognitive uncertainty" as an explicit differentiable signal into the decision-making link for the first time, enabling the system to quantify its "degree of control" over the current scenario and giving the system "self-awareness." Through the training mechanism of the reward function and the monotonically increasing mapping relationship between the MPC weight matrix and the cognitive uncertainty quantification index, the system automatically becomes conservative (increasing following distance and reducing speed) when encountering long-tailed unknown scenarios, and resumes efficient passage in familiar scenarios, achieving a smooth and stepless switching of driving style. By "safely projecting" the DRL intention command through the physical feasible domain solved online by MPC, it ensures that the final output control command always lies within the intersection of the vehicle dynamics envelope and collision constraints. This mechanism strictly constrains the "imagination" of DRL within the "possibilities" boundary of the physical world, solving the deployment problem of end-to-end learning models in safety-critical systems; the weight adjustment factor output by DRL provides scene-adaptive optimization parameters for MPC, avoiding the lag of manual parameter tuning; at the same time, the intent command output by DRL can be used as the initial value for hot start of MPC solution, significantly reducing the number of optimization iterations and meeting the real-time requirements of millisecond-level control cycle of autonomous driving.

[0057] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention.

[0058] It should also be noted that the various specific technical features described in the above embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not describe the various possible combinations separately.

[0059] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0060] Furthermore, various different implementations of the present invention can be combined arbitrarily, as long as they do not violate the spirit of the present invention, they should also be regarded as the content disclosed in the present invention.

Claims

1. An adaptive model predictive control method with uncertainty perception capability, characterized in that, Includes the following steps: Step S1: Using a deep integrated scene understanding network composed of M models, the current driving environment observation data is processed in parallel, the statistical divergence between the output results of the M models is calculated, and a quantitative index value representing the uncertainty of the current scene cognition is generated; the magnitude of the quantitative index value is used to measure the familiarity and judgment of the current scene, and the larger the value, the lower the familiarity. Step S2: The quantitative index value is used as an explicit risk feature dimension and is concatenated with the environmental state vector obtained after encoding the driving environment observation data to form a risk perception state vector that integrates scene semantic information and system self-awareness confidence. Step S3: Input the risk-aware state vector into the deep reinforcement learning policy network, which outputs the following two sets of signals in parallel: Weight adjustment factor: A coefficient vector used to dynamically adjust the weights of each penalty term in the cost function of the model's predictive controller; Intended action command: The initial control quantity that represents the optimal operating intention in the current driving scenario; Step S4: The model prediction controller receives the weight adjustment factor and reconstructs its own cost function online according to the weight adjustment factor, so that the weight of the safety-related penalty term in the cost function increases monotonically as the value of the quantification index increases. Step S5: The reconstructed model predictive controller, based on the current vehicle state, road surface adhesion conditions, and actuator saturation constraints, solves online the set of physically feasible actions that satisfy vehicle dynamics constraints and collision avoidance constraints in the future prediction time domain, denoted as the feasible region. Step S6: Determine whether the intentional action command generated in step S3 is within the feasible domain: If it is within the feasible domain, the intention action instruction will be directly output as the final control instruction; If it is not located within the feasible region, then on the boundary of the feasible region, a safe action command that is geometrically closest to the intended action command is solved, and the safe action command is output as the final control command to the vehicle actuator.

2. The method according to claim 1, characterized in that, The formula for calculating the quantitative indicator value in step S1 is as follows: in, This represents the value of the quantitative indicator. This indicates that the m-th model is based on the input observation data. The output scene feature encoding vector, This represents the mean of the output vectors of the M models; the larger the value of the quantification index, the greater the discrepancy in the understanding of the current scene among the sub-models in the deep integrated scene understanding network, and the higher the uncertainty of the system's cognition of the current scene.

3. The method according to claim 2, characterized in that, The risk perception state vector described in step S2 is constructed as follows: The environmental state vector obtained after feature extraction from the environmental observation data is concatenated with the quantitative index value to form a risk perception state vector with expanded dimensions. The risk-aware state vector enables the deep reinforcement learning policy network to explicitly perceive the system's confidence level in the current scenario.

4. The method according to claim 3, characterized in that, The deep reinforcement learning policy network adopts the SoftActor-Critic architecture, and the reward function used in the offline training phase... It has the following forms: in, Positive reward items that characterize traffic efficiency; To characterize the negative penalty for dangerous events, The reward function is designed to adjust the reward coefficient so that the deep reinforcement learning policy network learns the following behavioral strategy during training: when the quantification index value increases, it actively reduces the pursuit of efficiency and prioritizes avoiding safety penalties.

5. The method according to claim 1, characterized in that, The cost function after model prediction controller reconstruction in step S4 It has the following forms: in, Let be the state error vector at the k-th prediction step. To control the increment vector, To predict the length of the time domain, The state weight matrix is... The control weight matrix; the state weight matrix With control weight matrix The weight adjustment factor generated in step S3 is dynamically determined.

6. The method according to claim 5, characterized in that, The state weight matrix The weighting coefficients corresponding to the following distance error and the weighting coefficients corresponding to the lateral position deviation are both positively correlated with the quantified index value; this enables the model prediction controller to automatically execute a more conservative driving strategy when the system's perception of the current scene becomes more uncertain. The driving strategy is manifested in at least one of the following behaviors: increasing the safe following distance, reducing the steering response speed, and narrowing the feasible speed range.

7. The method according to claim 1, characterized in that, Step S6, which involves finding a safe action command that is geometrically closest to the intended action command, employs the following quadratic programming problem: in, The intention action command generated in step S3, The set of physical actions generated in step S5. This represents the square of the Euclidean distance between the action vector and the intention action vector.

8. The method according to claim 1, characterized in that, The M models in the deep integrated scene understanding network are either homogeneous heterogeneous neural networks with the same network structure but different initialization parameters, or heterogeneous neural networks with different network structures.

9. An adaptive model predictive control system with uncertainty perception capability, characterized in that, include: The uncertainty measurement module is configured to perform the function of step S1 as described in claim 1, and is equipped with a deep integrated scene understanding network consisting of M models, for receiving driving environment observation data and outputting a quantitative index value of cognitive uncertainty. The risk perception and decision-making module is configured to perform the functions of steps S2 to S3 as described in claim 1, and is equipped with a deep reinforcement learning policy network for fusing the quantitative index value with the environmental state vector and outputting a weight adjustment factor and intention action command. The model prediction control and safety projection module is configured to perform the functions of steps S4 to S6 as described in claim 1. It is equipped with a model prediction controller that can reconstruct the cost function online. The controller is used to reconstruct the control strategy according to the weight adjustment factor, calculate the physical action domain online, perform safety verification and projection correction on the intention action command, and output the final safety control command.

10. The system according to claim 9, characterized in that, The model prediction control and safe projection module is further configured as follows: When the intended action command does not fall within the physical actionable domain, a quadratic programming problem is solved on the boundary of the physical actionable domain to determine the safe action command with the closest Euclidean distance to the intended action command.

Images