Unmanned global path planning method and system based on quantum annealing optimization

Abstract

The application discloses an unmanned global path planning method and system based on quantum annealing optimization, and the method steps comprise the following steps: in step S01, real-time surrounding environment information of a controlled vehicle is acquired; in step S02, an improved Hamiltonian is used in a quantum modeling process and is converted into a QUBO model, and parameters in the improved Hamiltonian change according to real-time environment information; in step S03, an initial path is generated by using a classical preprocessing algorithm, and priority sorting is performed; parameters of the improved Hamiltonian and the QUBO model are updated, and are input into a quantum annealer for solving; a classical algorithm result is used to guide quantum solving, and parameters of the classical algorithm are optimized according to a solving result; and in step S04, a global optimal path is converted into a control instruction to control the controlled vehicle to travel. The application can take into account hardware limitation conditions, operation efficiency requirements and environment adaptability of the vehicle, and improve calculation efficiency, global optimization capability and environment adaptability of planning.

Description

Technical Field

[0001] This application relates to the field of autonomous driving path planning technology, and in particular to an autonomous driving global path planning method and system based on quantum annealing optimization. Background Technology

[0002] Global path planning is one of the key tasks of autonomous vehicles. It involves selecting an efficient and reasonable route based on the vehicle's real-time driving information and current road conditions. Current technologies primarily employ the following methods for autonomous driving path planning:

[0003] 1. Path planning method based on graph retrieval algorithm

[0004] This type of method implements path planning using graph retrieval algorithms such as A* and Dijkstra. However, under large-scale road network conditions, graph search-based algorithms need to traverse a large number of nodes and paths, which consumes a lot of time, has low computational efficiency, and cannot provide the path required by autonomous vehicles in a timely manner. For example, during the morning and evening rush hours in cities, there are tens of thousands of nodes on the road. If the A* algorithm is used, it is necessary to continuously expand the corresponding nodes, and each time a new cost needs to be calculated and sorted. This will cause the computational load of each step to increase exponentially. The larger the number of given nodes, the longer the algorithm takes, making it difficult to meet the real-time requirements. On the other hand, if heuristic search path planning algorithms such as Dijkstra are used, they may get stuck in local optima (such as being unable to escape suboptimal paths in complex obstacle environments). As the number of path nodes increases, the time complexity of path planning will also increase exponentially. In addition, traditional graph retrieval-based path planning methods are poorly adaptable to dynamically changing environments. When the environment changes (such as traffic accidents, road construction, etc.), it is necessary to re-traverse all nodes, which consumes a lot of time.

[0005] 2. Path planning methods based on optimization algorithms

[0006] This type of method transforms the control problem into a constrained optimization problem and uses optimization algorithms such as genetic algorithms for path planning. However, the results obtained from optimization algorithms generally only yield approximate optimal solutions. When dealing with multi-objective optimization problems with complex constraints or multiple objective functions, it is difficult to guarantee a global optimal solution. Furthermore, in urban road networks, if multiple detour paths are considered simultaneously due to multiple obstacles, the genetic algorithm may prematurely fall into local optima due to unreasonable crossover and mutation operators, thus failing to discover better detour paths.

[0007] Quantum annealing technology utilizes the quantum tunneling effect to escape local extrema, offering significant advantages in solving combinatorial optimization problems. Applying quantum annealing to UAV path planning can effectively improve global optimization capabilities, computational efficiency, and adaptability to dynamic environments. However, existing quantum annealing machines are bulky and energy-intensive, making them unsuitable for vehicle-mounted scenarios. A hybrid architecture combining quantum annealing and classical algorithms can alleviate this problem, but this type of quantum-classical hybrid architecture suffers from poor synergy, typically employing a simple "quantum solution + classical simple post-processing" approach. This means the quantum annealing algorithm and classical simple post-processing are used independently, with the classical part only performing preliminary data preparation. This can lead to wasted quantum resources (i.e., searching for invalid solutions) or the classical algorithm's accuracy failing to iterate. For example, even in a classical hybrid approach, it still requires complete enumeration of all node information, failing to improve computational efficiency and still exhibiting low computational efficiency. Furthermore, current dynamic traffic conditions require path planning systems to have millisecond-level response capabilities, which the aforementioned path planning methods cannot meet in terms of real-time performance. In addition, the Hamiltonian parameters used in the existing quantum annealing system (such as transverse field strength, problem coupling strength, and constraint penalty coefficient) are all constant values, which cannot be adjusted in real time according to the dynamic traffic environment, resulting in a reduction in the solution quality and computational efficiency of quantum solutions. Summary of the Invention

[0008] The technical problem to be solved by this application is: In view of the above-mentioned problems existing in the prior art, this application provides a quantum annealing-based global path planning method and system for autonomous driving that is simple to implement, has high computational efficiency, strong global optimization capability and environmental adaptability, and can take into account the vehicle's hardware constraints, operating efficiency requirements and environmental adaptability, ensuring that autonomous vehicles can drive safely and stably in complex environments, and solving the problems of traditional autonomous vehicle path planning methods being unable to achieve global optimization and lacking flexibility.

[0009] To solve the above-mentioned technical problems, the technical solution proposed in this application is as follows:

[0010] A global path planning method for autonomous driving based on quantum annealing optimization includes the following steps:

[0011] Step S01. Environmental Information Acquisition: Acquire real-time environmental information surrounding the controlled vehicle;

[0012] Step S02. Quantum Modeling: The path planning problem is modeled as a quantum annealing model. The quantum annealing model uses an improved Hamiltonian and is transformed into a QUBO model. The globally optimal path corresponds to the minimum energy value of the improved Hamiltonian. The globally optimal path is obtained by solving the QUBO model. The improved Hamiltonian includes a transverse field term, a problem coupling term, and a constraint penalty term to represent the degree of path violation. The transverse field term is used to drive the transition of qubits between the ground state and the excited state. The problem coupling term represents the quality of the path by its energy level. The parameters in the improved Hamiltonian change according to real-time environmental information. The parameters include the transverse field strength in the transverse field term, the problem coupling strength in the problem coupling term, and the constraint penalty coefficient in the constraint penalty term.

[0013] Step S03. Hybrid Quantum Annealing-Classical Computation: First, candidate road network nodes are selected using a classical preprocessing algorithm to generate an initial path and prioritize them. Then, the parameters of the improved Hamiltonian and the QUBO model are updated based on the environmental information surrounding the controlled vehicle. The updated QUBO model is then input into the quantum annealer for solving. The initial path and priority ranking generated by the classical preprocessing algorithm guide the solution of the QUBO model, and the parameters of the classical preprocessing algorithm are optimized based on the solution results. Finally, the quantum annealer outputs the optimal solution to obtain the globally optimal path.

[0014] Step S04. Control Execution: Convert the currently generated global optimal path into control commands to control the driving of the controlled vehicle.

[0015] Furthermore, the calculation expression for the improved Hamiltonian is as follows:

[0016]

[0017] in, For the transverse field term, Let be the transverse field intensity at time t. It is a transverse field; For the problem coupling term, Let be the problem coupling strength at time t. The problem energy is calculated using the following expression: , , and It is a binary variable. Indicates passing through nodes , Indicates not passing through nodes , Indicates passing through nodes , Indicates not passing through nodes , For nodes With nodes Euclidean distance between them Represents a node With nodes Real-time passage time between locations For nodes With nodes Energy consumption coefficient between These are the weighting coefficients; To constrain penalties, Let be the coefficient of the constraint penalty term at time t. To restrain punishment;

[0018] Transverse field strength Using the exponential decay model, the calculation expression is as follows: ,in, t Indicates time, The initial transverse field intensity, The decay time constant; the problem coupling strength Using the Sigmoid growth model, the calculation expression is: ,in, For saturated coupling strength, The inflection point time, is the growth rate constant.

[0019] Furthermore, the constraint penalty term includes hard constraints and soft constraints. The hard constraints include assigning a specified energy penalty to paths that violate the prohibition of reversing or obstacle avoidance. The soft constraints include applying gradient energy penalties when turning radius constraints or safety distance constraints are violated, so as to guide quantum annealing to select a path that meets the constraints. The gradient energy penalty is a penalty amount generated based on gradient according to the degree of violation of turning radius constraints or safety distance constraints.

[0020] Further, step S03 includes:

[0021] Step S301. Classical preprocessing of road network nodes: A suboptimal feasible path is obtained using a classic preprocessing algorithm. The range of candidate road network nodes is determined with the path as the center. Starting from the current vehicle location, the cost of all adjacent nodes is calculated. The node with the lowest cost is selected as the next target node. The node with the lowest total cost is expanded first to obtain all candidate road network nodes. The candidate road network nodes are prioritized according to their state information. The state information of the candidate road network nodes includes any one or more of path criticality, congestion sensitivity, and dynamic correlation. The dynamic correlation is determined based on the distance between the node and dynamic obstacles.

[0022] Step S302. Quantum Solving: The quantum solver determines the solution space to be solved based on the range of candidate road network nodes, updates the parameters of the improved Hamiltonian based on the environmental information around the controlled vehicle, and dynamically updates the Q matrix parameters of the QUBO model to obtain the currently updated QUBO (Quadratic Unconstrained Binary Optimization) model. The currently updated QUBO model is input into the quantum annealer for solving. During the solving process, each candidate path node is mapped to a quantum bit according to the candidate path nodes and priorities obtained in step S301, and finally the binary optimal solution is obtained.

[0023] Step S303. Multi-round annealing optimization: Use multi-round annealing to obtain several path solutions, and select the final optimal solution based on the overlap rate of path nodes.

[0024] Furthermore, step S302 also includes sparsifying the Q matrix based on node adjacency relationships:

[0025] Based on the road network topology of the environment where the controlled vehicle is located, a node adjacency matrix is ​​constructed. If there is a direct passage between two nodes, they are determined to be adjacent nodes; otherwise, they are determined to be non-adjacent nodes.

[0026] Retain the Q matrix elements corresponding to adjacent nodes, and assign a value of 0 to non-adjacent nodes to form a sparse matrix;

[0027] Store the sparse matrix and record the triples formed by the row index, column index, and element value of the non-zero elements.

[0028] Furthermore, in step S302, dynamically updating the Q matrix parameters includes:

[0029] Step S321. Dynamic Data Exponential Weighted Smoothing Calculation: An exponential weighted smoothing strategy is adopted to smooth the data within the window. The weights of the data points are: , The attenuation coefficient is used, and each weight is normalized.

[0030] Step S322. Trend Compensation Correction: Calculate the gradient change of the data within the window, predict the environmental trend based on the gradient change, and add compensation terms to optimize the path. The environmental trend includes congestion trend.

[0031] Step S323. Parameter calibration: Set the initial values ​​of each element in the Q matrix, analyze the degree of change of each element in the Q matrix to the QUBO model to analyze the sensitivity, select the multiple elements with the greatest influence on the QUBO model as the key elements of the Q matrix, and perform iterative optimization of each key element of the Q matrix using the particle swarm optimization algorithm. For each particle composed of a set of key elements of the Q matrix, solve the QUBO model through a quantum annealing simulator, calculate the objective function after obtaining the path and use it as the fitness value of the particle. After multiple iterations, output the key elements of the Q matrix corresponding to the optimal position of the swarm.

[0032] Furthermore, in step S3, before the quantum annealer solves the problem, each candidate node is split into two or more sub-problems using K-means clustering. There are multiple overlapping nodes between the sub-problems. The sub-problems are assigned to the quantum annealer in sequence to solve the quantum annealing sub-problems. When a local environmental change is detected, a local update of the Hamiltonian is triggered, so that during the quantum solution process, only the changed road network nodes are constructed to solve the local sub-problems. The local environmental change includes the movement of obstacles less than a preset threshold or the rate of change of congestion coefficient less than a preset threshold.

[0033] Furthermore, step S03 also includes constrained verification of whether the path satisfies any one or more of the following constraints: dynamic constraints, environmental constraints, and specified traffic rule constraints.

[0034] An autonomous driving global path planning system includes a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the method described above.

[0035] A computer-readable storage medium storing a computer program that, when executed, implements the method described above.

[0036] Compared with the prior art, the advantages of this application are:

[0037] 1. This application achieves global optimal path planning for autonomous driving by integrating quantum annealing and classical preprocessing algorithms. First, the classical preprocessing algorithm is used to select candidate road network nodes, generate initial paths, and sort them by priority. This significantly reduces the number of road network nodes. During the quantum annealing process, the results of the classical preprocessing algorithm guide the solution process. At the same time, the parameters of the classical preprocessing algorithm are optimized based on the solution results, which can greatly improve the computational efficiency and real-time performance of the path planning process.

[0038] 2. This application uses an improved Hamiltonian in the quantum annealing model to break through local optima through the quantum tunneling effect. The transverse field term is used to drive the transition of qubits between the ground state and the excited state. The problem coupling term represents the quality of the path by its energy level, and the constraint penalty term represents the degree of path violation. At the same time, by adjusting the transverse field strength, problem coupling strength and constraint penalty coefficient term in real time according to the dynamic traffic environment, the Hamiltonian can be dynamically adjusted according to the environmental state in real time, which can further improve the solution quality and computational efficiency of quantum solution, thereby improving the efficiency and accuracy of path planning. Attached Figure Description

[0039] Figure 1 This is a schematic diagram illustrating the implementation process of the global path planning method for autonomous driving based on quantum annealing optimization in this embodiment. Detailed Implementation

[0040] The present application will be further described below with reference to the accompanying drawings and specific preferred embodiments, but this does not limit the scope of protection of the present application.

[0041] like Figure 1 As shown, the steps of the quantum annealing-optimized global path planning method for autonomous driving in this embodiment include:

[0042] Step S01. Environmental Information Acquisition: Acquire real-time environmental information surrounding the controlled vehicle;

[0043] Step S02. Quantum Modeling: The path planning problem is modeled as a quantum annealing model. The quantum annealing model uses an improved Hamiltonian and is transformed into a QUBO model. The globally optimal path corresponds to the minimum energy value of the improved Hamiltonian. The globally optimal path corresponding to the minimum energy value of the improved Hamiltonian is obtained by solving the QUBO model. The improved Hamiltonian includes a transverse field term, a problem coupling term, and a constraint penalty term to represent the degree of path violation. The transverse field term is used to drive the transition of the qubit between the ground state and the excited state. The problem coupling term represents the quality of the path by its energy level. The parameters in the improved Hamiltonian change according to the real-time environmental information. The parameters include the transverse field strength in the transverse field term, the problem coupling strength in the problem coupling term, and the constraint penalty coefficient in the constraint penalty term.

[0044] Step S03. Hybrid Quantum Annealing-Classical Computation: First, the classical preprocessing algorithm is used to select candidate road network nodes, generate initial paths, and sort them by priority. Then, the parameters of the improved Hamiltonian and the QUBO model are updated based on the environmental information around the controlled vehicle. The updated QUBO model is then input into the quantum annealer for solving. The initial paths and priority sorting generated by the classical preprocessing algorithm guide the solution of the QUBO model, and the parameters of the classical preprocessing algorithm are optimized based on the solution results. Finally, the quantum annealer outputs the optimal solution to obtain the globally optimal path.

[0045] Step S04. Control Execution: Convert the currently generated global optimal path into control commands to control the driving of the controlled vehicle.

[0046] This embodiment achieves globally optimal path planning for autonomous driving by integrating quantum annealing and classical preprocessing algorithms. First, classical preprocessing algorithms are used to select candidate road network nodes, generate initial paths, and prioritize them, significantly reducing the number of road network nodes. During the quantum annealing process, the results of the classical preprocessing algorithm guide the solution process, and the parameters of the classical preprocessing algorithm are optimized based on the solution results, greatly improving the computational efficiency and real-time performance of the path planning process. Furthermore, an improved Hamiltonian is used in the quantum annealing model to break through local optima through quantum tunneling. The transverse field term drives the transitions of qubits between the ground and excited states, the problem coupling term represents the path quality through energy levels, and the constraint penalty term represents the degree of path violation. By adjusting the parameters of the transverse field term, problem coupling term, and constraint penalty term in the improved Hamiltonian in real time according to the dynamic traffic environment, the solution quality and computational efficiency of the quantum solution can be further improved, thus meeting the current real-time requirements of dynamic traffic.

[0047] In this embodiment, the environmental information surrounding the controlled vehicle in step S01 includes surrounding 3D point cloud data, traffic signs detected by the camera, object recognition and classification information such as pedestrians and vehicles, and vehicle position and speed information. Specifically, the environmental perception module can use onboard LiDAR to obtain 360-degree 3D point cloud data around the vehicle, use the camera to detect traffic signs, pedestrians, and vehicles, and obtain vehicle position and speed information via GPS. Based on the real-time data from LiDAR, camera, and GPS, key features such as congestion coefficient, obstacle movement speed, and traffic light status are then extracted.

[0048] Optionally, after collecting data from each sensor in step S01, the process also includes preprocessing operations such as denoising, filtering, and spatial alignment of the collected raw data to obtain data usable for path planning, such as road boundaries, obstacle categories and motion information, and traffic light status. Then, the information collected by multiple sensors is fused to provide accurate environmental information for the path planning process.

[0049] The essence of quantum annealing is to simulate the "cooling optimization" process of classical annealing, and it can break through local optima through the quantum tunneling effect. In this embodiment, during the quantum modeling process in step S02, an improved Hamiltonian with a three-part composite structure (transverse field term, problem coupling term, and constraint penalty term) is adopted. The transverse field term drives the qubit to transition between the ground state and the excited state to ensure global exploration without affecting the final energy minimization. The problem coupling term uses energy levels to represent the quality of the path, corresponding to the objective function of path planning. The better the path, the lower the corresponding energy, and the worse the path, the higher the corresponding energy. The constraint penalty term represents the degree of path violation. Based on the three-part composite structure, the quantum tunneling effect and engineering constraints can be organically integrated. Its goal is to make the minimum energy value of the Hamiltonian corresponding to the globally optimal path the ultimate value. This minimum value is finally found through the evolution process of quantum annealing, that is, the globally optimal path is solved. Since the horizontal field term is not directly related to the "quality" or "compliance" of the path, it only affects the exploration process. When the system finally converges, the energy of the horizontal field term approaches 0. The minimum value of the Hamiltonian is determined by the "problem coupling term + constraint penalty term", which corresponds to the "compliant optimal path".

[0050] Specifically, in this embodiment, the calculation expression for the improved Hamiltonian can be expressed as:

[0051] (1)

[0052] in, For the transverse field term, Let be the transverse field intensity at time t. It is a transverse field; For the problem coupling term, Let be the problem coupling strength at time t. The problem energy is calculated using the following expression: , , and It is a binary variable. Indicates passing through nodes , Indicates not passing through nodes , Indicates passing through nodes , Indicates not passing through nodes , For nodes With nodes Euclidean distance between them Represents a node With nodes Real-time passage time between locations For nodes With nodes Energy consumption coefficient between These are the weighting coefficients; To constrain penalties, Let be the coefficient of the constraint penalty term at time t. To constrain and punish.

[0053] In this embodiment, the transverse field strength An exponential decay model can be used, which means the decay occurs exponentially over time. For example, the calculation expression is: ,in, t Indicates time, The initial transverse field intensity, This is the decay time constant. Preferably, the initial time (t=0) can be configured... Maximum (e.g.) ), to ensure the system can explore the complete solution space, later It gradually decays to 0, no longer interfering with the energy dominance of the objective function and constraints.

[0054] In this embodiment, the problem coupling strength We can use the Sigmoid growth model, that is, grow according to the Sigmoid model. For example, the calculation expression is: ,in, For saturated coupling strength, The inflection point time, is the growth rate constant.

[0055] In this embodiment, the constraint penalty term Including hard constraints (Prohibition of driving against traffic, obstacle avoidance, etc.) and soft constraints (Turning radius, safety distance, etc.), the specific expression is as follows: Hard constraints include assigning a specified energy penalty to paths that violate the prohibition against reversing or obstacle avoidance rules. Soft constraints include applying gradient energy penalties when turning radius or safety distance constraints are violated; that is, energy penalties are triggered when constraints are violated, and the amount of energy penalty varies in a gradient. By employing the above constraint penalty terms, energy values ​​can be directly added to road engineering constraints (no reversing, turning radius, safety distance), i.e., when a path violates these constraints, until it becomes an "infeasible high-energy solution," which is then automatically excluded during quantum annealing calculations.

[0056] Specifically, hard constraints (prohibition of reverse driving, obstacle avoidance) adopt an "energy barrier term" design, if the node For vehicles traveling in the wrong direction or located in an obstacle area, a penalty coefficient can be set corresponding to the hard constraint. To make it much larger than the soft constraint penalty coefficient, for example, by setting... The value 100 forces the energy of the path to approach infinity, ensuring that quantum annealing does not converge to a violation solution. Soft constraints (turning radius, safety distance) are designed using a "gradient penalty term" to guide quantum annealing in selecting a path that meets the constraints. This gradient energy penalty is generated based on the gradient, according to the degree of violation of the turning radius or safety distance constraint. For example, the energy expression for the turning radius constraint is... , The function is a Heaviside step function; it triggers an energy penalty when the constraint is violated, and there is no additional penalty when it is compliant. The gradient coefficient is positively correlated with the angle difference. For nodes arrive The steering angle, The maximum permissible steering angle, if possible. The larger the angle difference, the faster the penalty energy increases, thus guiding quantum annealing to choose a path that conforms to the dynamic constraints; the expression for the safe distance constraint is... , Real-time vehicle distance can be obtained through LiDAR measurements. This is a dynamic safety distance.

[0057] This embodiment classifies constraints into hard and soft constraints based on their "degree of insurmountability." Hard constraints (prohibition of reverse driving, obstacle avoidance) are transformed into "energy barrier terms," ​​causing the energy of illegal paths to approach infinity, which is automatically eliminated by quantum annealing. Soft constraints (turning radius, safety distance) are transformed into "gradient penalty terms," ​​with higher penalty energy for larger deviations. This guides the quantum solution to select paths that conform to dynamic constraints. "Illegal paths" are assigned "high-energy penalties," allowing engineering constraints to be encoded and coupled to the energy function according to their "degree of insurmountability." This establishes a direct correlation between constraint classification and energy functions, thereby constructing a direct coupling system between multiple constraints ("vehicle dynamics-environment-traffic rules") and the QUBO energy function. This ensures that the quantum solution results possess both mathematical optimality and engineering feasibility, solving the problem that "mathematically optimal solutions cannot be implemented in engineering" in traditional quantum annealers.

[0058] In this embodiment, the parameter in the improved Hamiltonian (transverse field strength) Problem Coupling Strength and constraint penalty coefficient The improved Hamiltonian is based on real-time environmental information changes and can adapt to environmental changes in a timely manner, solving the problem of "response lag" of traditional fixed values, thereby improving computational efficiency and solution quality. After constructing the improved Hamiltonian by performing quantum modeling according to step S02, before starting quantum annealing in step S03, the three types of parameters in the improved Hamiltonian, namely the transverse field strength, problem coupling strength, and constraint penalty coefficient, are adjusted in real time according to real-time traffic environment data (such as traffic jams, obstacle movement speed, etc.). The parameters of the Hamiltonian are changed in real time, and the parameters of the improved Hamiltonian are updated according to the new environmental information when there are changes, so that it can adapt to the dynamic traffic environment and quantum annealing solution logic.

[0059] Specifically, the Hamiltonian parameters are updated every 10ms (configurable). When a sudden change is detected, the constraint penalty term and the problem coupling term parameters are automatically associated, and the constraint penalty coefficient is adjusted. For example, if a sudden congestion (travel time fluctuation greater than 30%) is detected by the LiDAR measurement within 50m, the penalty coefficient of the automatically generated safe distance constraint is increased from 5 to 8. At the same time, the weight of the real-time travel time in the problem coupling term is increased from 0.4 to 0.6, so that quantum annealing will prioritize the congestion avoidance problem. If the camera captures that the traffic light at a certain intersection has turned red, the hard constraint penalty coefficient of the corresponding road network node will be automatically generated and increased to 150, so that the optimized path that avoids passing through the intersection cannot meet the traffic light rules of all intersections, thereby achieving a two-way balance between global optimal exploration and dynamic constraint adaptation.

[0060] The goal of the quantum annealing mathematical model is to minimize the energy value of the Hamiltonian corresponding to the "globally optimal and engineering-compliant path". In this embodiment, during the quantum modeling process in step S02, a QUBO (Quadratic Unconstrained Binary Optimization) model (energy function) is constructed. The optimal binary solution (corresponding to path node selection) is found by minimizing this QUBO model. The Hamiltonian and the QUBO model are directly equivalent, with the problem coupling term directly corresponding to the QUBO model. Therefore, by controlling the strength of the transverse field term and the problem coupling term (the transverse field decays over time, while the problem coupling strength increases over time), the quantum system can evolve towards its energy ground state. The binary variable vector corresponding to the final ground state is the optimal solution for path planning. The standard form of the QUBO model is... ,in, It is a binary variable vector. It is the quadratic coefficient matrix (QUBO matrix). The core of the QUBO model is to transform the path planning problem into a quantum-solvable mathematical model, with the objective of minimizing this QUBO model, i.e., minimizing... Take the corresponding time Vector (the selection of corresponding path nodes). This embodiment transforms the "problem coupling term + constraint penalty term" into a QUBO matrix using a quantum modeling method. Q The path planning problem can be transformed into finding the optimal solution of the QUBO model.

[0061] In this embodiment, step S03 employs a hybrid quantum annealing-classical computation approach combining classical pre-solution and quantum optimization. First, a classical preprocessing algorithm (e.g., a greedy algorithm) is used to quickly generate feasible solutions, creating a "suboptimal feasible path." A "candidate node cluster" is then defined around this path, reducing the number of nodes considered. Simultaneously, nodes are prioritized based on path criticality, congestion sensitivity, and dynamic correlation. Then, a quantum annealing algorithm is used for optimization. During the quantum solution process, the results of the classical preprocessing algorithm guide the direction of the quantum solution, focusing it on the "local space where the optimal solution resides." This avoids the problem of "undirected exploration" in the quantum system caused by directly inputting compressed nodes into quantum annealing, thus significantly improving the speed and accuracy of the solution. Using this hybrid quantum annealing-classical computation approach, a bidirectional collaborative system can be constructed: "classical guidance for quantum focusing on the effective solution space, and quantum feedback optimization of classical model parameters." Classical preprocessing computation compensates for the limitations of quantum hardware, while quantum computation provides feedback to classical preprocessing, achieving the transformation from "quantum capability to engineering usability."

[0062] In this embodiment, the specific steps of step S03, hybrid quantum annealing-classical computation, include:

[0063] Step S301. Classical preprocessing of road network nodes: Use the classic preprocessing algorithm to obtain a suboptimal feasible path, determine the range of candidate road network nodes with the path as the center, calculate the cost of all adjacent nodes starting from the current vehicle location, select the node with the smallest cost as the next target node, prioritize the expansion of the node with the smallest total cost, obtain all candidate road network nodes, and prioritize each candidate road network node according to the status information of the candidate road network nodes. The status information of the candidate road network nodes includes any one or more of path criticality, congestion sensitivity, and dynamic correlation. The dynamic correlation is determined based on the distance between the node and dynamic obstacles.

[0064] Step S302. Quantum Solving: The quantum solver determines the solution space to be solved based on the range of candidate road network nodes, updates the parameters of the improved Hamiltonian based on the environmental information around the controlled vehicle, and dynamically updates the Q matrix parameters of the QUBO model to obtain the currently updated QUBO model. The currently updated QUBO model is input into the quantum annealer for solving. During the solving process, each candidate path node is mapped to a quantum bit according to the candidate path nodes and priorities obtained in step S301. High-priority nodes are mapped to low-noise quantum bits, and finally the binary optimal solution is obtained.

[0065] Step S303. Multi-round annealing optimization: Use multi-round annealing to obtain several path solutions, and select the final optimal solution based on the overlap rate of path nodes.

[0066] This embodiment employs the aforementioned hybrid quantum annealing-classical computation. First, a suboptimal feasible path is generated through classical preprocessing of the road network nodes. Simultaneously, nodes are prioritized based on path criticality, congestion sensitivity, and dynamic correlation to select "candidate path nodes" before quantum computation. This reduces the problem size, matches the number of qubits, and lowers the computational cost, thereby guiding deep optimization through quantum annealing. During quantum computation, high-priority nodes are mapped to low-noise qubits, focusing the quantum computation on the effective solution space and reducing the exploration of invalid solutions. The quantum annealing results are then fed back to iteratively optimize the classical preprocessing model. The "deviation between the optimal path and the initial path" output by the quantum computation is fed back to the classical preprocessing model, iteratively improving the quality of the initial path generated by the classical model. This forms a closed loop of "classical guidance → quantum optimization → classical iteration," effectively improving computational efficiency and solution quality.

[0067] For example, a classical preprocessing algorithm can employ a greedy algorithm. First, a suboptimal feasible path is generated using the greedy algorithm to create an initial path. Based on this initial path, a baseline value for the path cost is determined. During parameter calibration in QUBO modeling, this baseline value can be used as the target energy threshold for particle swarm optimization to further narrow the optimization range. Simultaneously, the "deviation between the optimal path and the initial path" output by the quantum solution (e.g., a 10% reduction in path length) is fed back to the greedy algorithm, dynamically adjusting the "estimated cost weight" of the greedy algorithm (e.g., adjusting the Manhattan distance weight from 0.3 to 0.35). By dynamically adjusting the "cost weight" of the classical preprocessing algorithm through the particle swarm optimization algorithm, the quality of the initial path in the next round is improved.

[0068] In a specific application embodiment, the detailed steps of classic preprocessing of road network nodes when using a greedy algorithm are as follows:

[0069] Step S311. Initial Path Guidance: Use a greedy algorithm or an improved greedy algorithm to obtain a suboptimal feasible path. Determine the range of candidate road network nodes centered on the path, so that the quantum solver only solves the effective solution space. Starting from the current vehicle's location, calculate the "instant cost" of all adjacent nodes (i.e., the sum of the Euclidean distance from the node to the vehicle's location and the real-time travel time from the node to the next node), and select the node with the minimum cost as the next target node. Preferably, each expansion step needs to consider both the cost from the current node to adjacent nodes and the estimated cost from adjacent nodes to the target (which can be estimated using Manhattan distance). The total cost is the selected cost, and the node with the lowest total cost is expanded first.

[0070] Step S312. Node Priority Ranking: The selected candidate road network nodes are prioritized according to their path criticality, congestion sensitivity, dynamic correlation, and other state information. This allows for the selection of corresponding low-noise qubits for "critical nodes" (such as start point, end point, and path turning point) during the subsequent quantum solution process, thereby improving the accuracy of the quantum solution.

[0071] Taking the calculation of the initial path using a greedy algorithm in a specific application embodiment as an example, the steps include:

[0072] Node selection: Taking the second selected node as an example, the starting point (node ​​1) is adjacent to nodes 2, 3, and 4. Calculate the instantaneous cost C=0.6d+0.4t: Node 2 (d=10m, t=5s, C=8.0), Node 3 (d=12m, t=6s, C=9.6), Node 4 (d=15m, t=8s, C=12.2). Select node 2, and the initial path is [1→2].

[0073] Path extension: Node 2 is adjacent to nodes 5, 6, and 7. The total cost is Total = C + 0.3H (H is the Manhattan distance): Node 5 (C=7.0, H=700m, Total=217.0), Node 6 (C=6.0, H=680m, Total=210.0), Node 7 (C=8.0, H=720m, Total=224.0). Select node 6 and update the path [1→2→6].

[0074] Path compression: Generate the initial complete path [1→2→6→9→15→22→30→35→40→…→Destination B]. Remove the intermediate node 9 from the continuous collinear nodes (6→9→15, angle 175°≥170°). The compressed path [1→2→6→15→22→30→35→40→…→Destination B] is the desired initial path.

[0075] The qubits in a quantum annealer have "uniform noise" (some qubits have low noise). If the optimization accuracy of the quantum system for nodes with low qubit noise is insufficient, there may be intersections between the planned paths and obstacles. This embodiment prioritizes the candidate nodes after screening them, ensuring that nodes closer to obstacles have a higher frequency of changes in their "occupied" and "passable" states. For example, a priority score can be calculated for each node based on state information such as path criticality, congestion sensitivity, and dynamic correlation. .in, This indicates the criticality of the path. For example, the starting point / end point can be configured to receive 10 points, nodes on the initial path can receive 8 points, adjacent nodes 5 points, and other nodes 3 points. This indicates congestion sensitivity. For example, nodes with a travel time fluctuation of >20% can be assigned 8 points, requiring key optimization, while nodes with a fluctuation of ≤20% can be assigned 5 points. This indicates the dynamic correlation degree. For example, nodes less than 50m from dynamic obstacles can be configured to score 8 points, while nodes 50m or more can score 5 points. Then press... Sort the nodes from highest to lowest priority, mark the top N (configurable) high-priority nodes as "core nodes", and output the "node-priority" list to the quantum optimization module to guide the qubit mapping, so that the core nodes are preferentially mapped to low-noise qubits.

[0076] In this embodiment, during the process of determining candidate nodes using the classic preprocessing algorithm for road network nodes, the spatial correlation priority rule can be further used to filter nodes. For example, only nodes related to the connection between the current vehicle position and the target position can be retained, and redundant nodes that are too far away can be deleted. Invalid nodes can also be filtered out in advance using constraints, such as nodes in prohibited areas or with permanent obstacles, nodes in "permanent obstacle areas" or "prohibited road sections", and nodes "currently occupied by dynamic obstacles" such as those involved in traffic accidents. This achieves the optimization of the original road network nodes.

[0077] In this embodiment, step S302 further includes sparsifying the Q matrix based on node adjacency relationships to achieve sparse matrix optimization. Specific steps include:

[0078] Based on the road network topology of the environment where the controlled vehicle is located, a node adjacency matrix is ​​constructed. If there is a direct passage between two nodes, they are determined to be adjacent nodes; otherwise, they are determined to be non-adjacent nodes.

[0079] Retain the Q matrix elements corresponding to adjacent nodes, and assign a value of 0 to non-adjacent nodes to form a sparse matrix;

[0080] Store a sparse matrix and record the triples formed by the row index, column index, and element value of the non-zero elements.

[0081] Considering that the large scale of urban road network nodes can lead to data redundancy in the Q matrix, this embodiment solves the problem of Q matrix data redundancy by sparsifying the Q matrix based on node adjacency relationships, retaining only physically non-zero terms and eliminating invalid redundant terms.

[0082] In a specific application embodiment, the detailed steps for sparsifying the Q matrix based on node adjacency relationships are as follows:

[0083] Adjacency determination: Based on the road network topology of the high-precision map, a node adjacency matrix is ​​constructed. If there is a direct road segment between node i and node j, it is determined to be an "adjacent node"; otherwise, it is a "non-adjacent node".

[0084] Non-zero term retention: Only the Q matrix elements corresponding to adjacent nodes are retained. diagonal elements (adjacent non-diagonal elements), non-adjacent nodes A uniform value of 0 is assigned, meaning that non-adjacent nodes do not need to be included in energy calculations because they have no direct access routes.

[0085] Sparse storage format: The sparse matrix is ​​stored using the Coordinate Format (COO), which records only the "row index-column index-element value" triples of non-zero elements. Taking 200 candidate nodes as an example, the average number of adjacent nodes is 3-4 per node, and the total number of non-zero elements is approximately 800, with a sparsity of >95%. Compared to full storage (200×200=40000 elements), this can reduce memory usage by more than 90% and improve the data reading efficiency for subsequent quantum solutions.

[0086] In step S302 of this embodiment, dynamically updating the Q matrix parameters includes:

[0087] Step S321. Dynamic Data Exponential Weighted Smoothing Calculation: An exponential weighted smoothing strategy is adopted to smooth the data within the window. The weights of the data points are: , The attenuation coefficient is used, and each weight is normalized.

[0088] Step S322. Trend Compensation Correction: Calculate the gradient change of the data within the window, predict the environmental trend based on the gradient change, and add compensation terms to optimize the path. The environmental trend includes congestion trend.

[0089] Step S323. Parameter calibration: Set the initial values ​​of each element in the Q matrix, analyze the degree of change of each element in the Q matrix to the QUBO model to analyze the sensitivity, select the multiple elements with the greatest influence on the QUBO model as the key elements of the Q matrix, and perform iterative optimization of each key element of the Q matrix using the particle swarm optimization algorithm. For each particle composed of a set of key elements of the Q matrix, solve the QUBO model through a quantum annealing simulator, calculate the objective function after obtaining the path and use it as the fitness value of the particle. After multiple iterations, output the key elements of the Q matrix corresponding to the optimal position of the swarm.

[0090] Q matrix parameters (such as) The path weights and constraint penalty coefficients in the model rely on dynamic data from multiple sensors (real-time travel time, obstacle distance, vehicle speed). This embodiment uses a fixed time window combined with an exponentially weighted smoothing sliding window mechanism to denoise and compensate for trends in the dynamic data, providing stable basic data for parameter calibration. This solves the problem of parameter fluctuations caused by instantaneous interference (such as lidar obstruction, GPS signal fluctuations) in the original data. Furthermore, through the above steps of sparse matrix optimization, sliding window data preprocessing, and parameter calibration, the QUBO model can also be well adapted to the hardware input requirements of the quantum annealer (such as qubit number limitations and data format requirements).

[0091] In a specific application embodiment, when step S321 adopts an exponentially weighted smoothing strategy, the first... The weights of the data points are: ( (This is the attenuation coefficient), and each weight is normalized to ensure that the latest data has the highest proportion, balancing real-time performance and smoothness. For example, combined with the sensor sampling frequency set to 3 data update cycles and a total window duration of 300ms, this ensures sufficient data points are covered to smooth fluctuations, while avoiding excessively large windows that could cause data lag; for road segments... Real-time passage time Let the current time be the time of the original data for the three cycles within the window. The previous cycle was The first two cycles are The smoothed travel time is , Using the same logic, the distance between the vehicle and the obstacle can be calculated, with these being weighting coefficients. Speed Smoothed distance , This can prevent false detection data from causing abnormal changes in constraint penalty coefficients and frequent path adjustments.

[0092] After the dynamic data index weighted smoothing calculation in step S311 is completed, the trend compensation correction step S322 is executed. By calculating the gradient changes of the data within the calculation window, the environmental trend is predicted and a compensation term is added, which can solve the problem of the dynamic data smoothing lagging behind the actual environmental changes. Specifically, gradient calculation is first performed, based on the real-time travel time. For example, gradient This reflects the rate of change in travel time within 200ms; then, the trend of change is determined. If (i.e., a change in travel time of >10s within 200ms) is judged as a "significant trend". A value greater than 0.05 indicates increased congestion. <-0.05 indicates congestion relief); finally, compensation calculations are performed. When congestion worsens, the congestion trend is amplified, and path weights are adjusted in advance. When congestion eases, travel time is reduced, and paths are optimized in a timely manner.

[0093] In a specific application embodiment, step S323 can utilize a three-stage optimization process—initial assignment, sensitivity analysis, and particle swarm optimization—to complete the calculation of Q matrix elements and achieve parameter calibration. By directly integrating the categorized and coded constraint terms into the QUBO energy function, and through the aforementioned three-stage parameter calibration, the Q matrix elements can accurately match the constraint requirements, allowing the quantum solution process to simultaneously satisfy multiple constraints and improving path executability. Specifically, the detailed steps of step S323 are as follows:

[0094] 1) Initial assignment

[0095] In the initial assignment phase, the elements of the Q matrix are set within a reasonable initial range. Specifically, based on the engineering characteristics of autonomous driving path planning and the energy function logic of quantum annealing, different initial values ​​are determined for each type to prevent subsequent optimization from entering a solution-less space. The QUBO matrix Q is Symmetric matrix (n represents the number of candidate nodes), elements Corresponding binary variable (node Whether or not it was selected) and The coupling relationship (whether node j is selected) can be divided into three categories according to different situations, as shown in Table 1. Constraints and gapless mapping to the QUBO model:

[0096] Table 1: Q matrix element classification and assignment rules.

[0097]

[0098] Preferably, in the classic preprocessing algorithm, the constraint type (hard constraint) and penalty level can be marked according to the nodes removed before the constraint is imposed (such as reverse road segments and obstacle nodes). Then, in the QUBO modeling process, the corresponding Q matrix elements can be directly assigned high-energy penalties to simplify the hard constraint coding process of parameter calibration.

[0099] Furthermore, expert experience can be incorporated to modify the initial assignments. For example, hard constraint modifications can be implemented, such as assigning a high energy penalty to the reverse direction of travel in a one-way road segment. Dynamic scenario modifications can also be performed, such as increasing the road segment energy consumption coefficient (e.g., increasing the weight of the road segment energy consumption coefficient from 0.2 to 0.3) and the road segment travel time coefficient (e.g., increasing the weight of the road segment travel time coefficient from 0.5 to 0.6) in congested road segments. Vehicle characteristic modifications can also be made; for example, if the minimum turning radius of a vehicle is 8m, then for all roads with turning angles greater than or equal to 150°... Increase all of them by a certain percentage to achieve routes with larger turning radii.

[0100] 2) Sensitivity analysis

[0101] In the sensitivity analysis phase, the key parameters that have the greatest impact on the QUBO model (energy function) are selected to reduce the computational burden during subsequent optimization. Matrix elements Sensitivity refers to the energy function Regarding the degree of change, it can be used as... right The partial derivative is expressed as follows. When (Diagonal elements) ( (binary characteristics), sensitivity ( Reflecting nodes The probability of being selected); when (Non-diagonal elements) Sensitivity , Reflecting nodes The probability of being selected at the same time.

[0102] 3) Particle Swarm Optimization

[0103] In the particle swarm optimization phase, several candidate paths are first generated using a quantum annealing simulator (D-Wave LeapSimulator) based on the initialized Q matrix, and a binary vector corresponding to each candidate path is recorded. Then, the Particle Swarm Optimization (PSO) algorithm is used to iteratively optimize the key elements so that the path obtained by solving the QUBO model satisfies both the requirements of global optimum and engineering constraints.

[0104] Specifically, particle swarm optimization multi-objective weighted function .in, Indicates the optimality of the path (the smaller the better). Indicates the degree of satisfaction of dynamic constraints (the smaller the better). This indicates the degree of environmental constraint satisfaction (the smaller the better). This represents the weighting coefficients. For each particle (corresponding to a set of key elements in the Q matrix), the QUBO model is solved using a quantum annealing simulator to obtain the path, and then the objective function is calculated. The current particle's fitness value is used as its fitness value; if the current particle's fitness value is less than its historical best fitness value, then its current position is updated. If the change in the population's best fitness value is less than 10 consecutive iterations... If convergence is achieved, the key element of the Q matrix corresponding to the optimal position of the population is output; otherwise, the iteration continues.

[0105] This embodiment dynamically updates the Q-matrix parameters through the above steps. It uses a time window to exponentially smooth the path weight coefficients (node ​​distance, travel time, energy consumption, etc.), which can eliminate instantaneous noise interference. Through parameter gradient trend prediction, it can compensate for the inherent solution delay of the quantum annealing model and ensure that parameter adjustment is synchronized with environmental changes. At the same time, it uses the particle swarm optimization algorithm to iteratively optimize key elements, so that the path obtained by solving the QUBO model can meet both the requirements of global optimum and engineering constraints. Combined with the real-time update of the improved Hamiltonian, a fully closed-loop update system of "real-time perception - trend prediction - constraint coordination - parameter iteration" can be constructed to achieve accurate matching between the quantum model and the dynamic scene.

[0106] Considering that the number of binary variables in the classical preprocessed QUBO model may approach or slightly exceed the number of qubits that a quantum annealer can process in a single run (e.g., some small-scale quantum annealers can only process 160 qubits at a time), this embodiment employs block quantum computing logic to decompose the original problem into subproblems for adaptation. Specifically, before the quantum annealer solves the problem, each candidate node is split into two or more subproblems using K-means clustering. These subproblems have multiple overlapping nodes. By sequentially assigning each subproblem to the quantum annealer and solving the quantum annealing subproblems, the limitation on the number of qubits can be resolved.

[0107] In a specific application embodiment, the detailed steps for sub-problem decomposition and solution are as follows:

[0108] 1) Sub-problem spatial clustering and splitting: The candidate node set output by the classical preprocessing algorithm is spatially distributed, and the K-means clustering algorithm (K value = number of variables / number of qubits, rounded up) is used to split multiple nodes into two or more sub-problems. For example, 200 nodes are split into 2 sub-problems, that is, after splitting, they become 100 nodes / sub-problems. It is ensured that the nodes on each sub-problem are spatially continuous, that is, the route after splitting will not be broken.

[0109] 2) Subproblem boundary handling: Add overlapping nodes between adjacent subproblems so that each subproblem has multiple identical nodes, and ensure that the Q matrix of the same node has the same parameters in the two adjacent subproblems, so as to ensure the continuity of the path when performing classic splicing in the future.

[0110] 3) Sub-problem solving scheduling: The edge computing unit task scheduling module dispatches sub-problems to the quantum annealer according to spatial order, such as the shortest distance priority principle, to ensure that the simultaneous execution of different sub-problems does not cause conflicts between hardware resources. Assuming that the solution time of each sub-problem is ≤15ms, the total time spent between two sub-problems is ≤30ms (including scheduling delay). Therefore, by combining sub-problem decomposition with the millisecond-level convergence characteristics of quantum annealing, this embodiment can achieve a total planning time of ≤100ms. Compared with traditional graph retrieval methods such as the A* algorithm, this can greatly improve response efficiency and meet the current millisecond-level response requirements of dynamic traffic.

[0111] In a specific application embodiment, during the quantum solution process in step S03, candidate nodes and priority ranking results (such as a "node-priority" list) are first obtained according to the classical preprocessing algorithm. High-priority core nodes are preferentially mapped to low-noise qubits, and then ordinary nodes are mapped to conventional qubits to ensure the solution accuracy of core nodes. A greedy mapping algorithm can be used during the mapping process, prioritizing the allocation of optimal qubits to the highest-priority nodes, and continuously allocating qubits to each node in descending priority order to maximize the utilization of qubit resources and prevent qubits from being idle. Then, a preprocessed small-scale QUBO model is used as the input to the quantum annealer to obtain the binary optimal solution. For example, x=[1,0,1,0,1] represents the node path 1→3→5. The quantum annealer (taking D-Wave Advantage2 as an example) can only accept QUBO model inputs in a specific format (commonly sparse matrix instruction sets), therefore, the preprocessed QUBO energy function ( , A binary variable vector, This involves converting the quadratic coefficient matrix into a hardware-recognizable instruction format. For example, the COO (Coordinate Format) format can be used to compress the data size, saving only the non-zero "row index-column index-value" triples in the Q matrix, thus optimizing the sparse matrix format. Then, the quantum-classical interface can be used to convert the COO format Q matrix into dedicated instructions for the quantum annealer, such as qubit mapping instructions (specifying the correspondence between binary variables and qubits) and coupling strength configuration instructions (setting the interaction weights between qubits to determine the coupling strength of the problem). Hardware instruction encoding is implemented, and data verification is performed to ensure that the transmitted data is correct, avoid errors in Q matrix parameters due to transmission noise, and maintain the accuracy of model input.

[0112] Preferably, considering that the qubit performance of the quantum annealer (such as decoherence time and coupling strength) will slightly change with the operating time (fluctuation in a constant temperature environment of 100mK±5mK), the parameters of the QUBO model can also be calibrated based on the real-time hardware state to match the real-time state of the quantum annealer. The specific steps include:

[0113] First, real-time qubit noise values ​​(such as single-qubit gate error rate) are obtained through the monitoring interface of the quantum annealer (sampling frequency 1kHz). Coupler errors (such as the interaction deviation between two qubits) );

[0114] Then, the Q matrix parameters are adjusted; if the noise value of a certain qubit... If the value exceeds the design threshold, then the corresponding binary variable will be... Matrix coefficients (Diagonal terms) are increased by a specified amount (e.g., 5%-10%) to enhance the solution weight of the variable and offset the effects of noise; if coupler error Then the corresponding off-diagonal terms of the Q matrix Reverse compensation based on error proportion, such as hour, Reduced by 6%.

[0115] Finally, the calibration effect can be verified by judging the path length error rate through the solution results of small-scale test cases, and determining whether to switch to formal solution.

[0116] The quantum annealer solution process evolves from quantum state evolution to output a binary optimal solution. After inputting the adapted QUBO model into the quantum annealer, the solution is completed through sequential qubit initialization, annealing process control, and quantum state measurement. It utilizes the quantum tunneling effect to overcome local optima, ensuring that the output binary solution corresponds to the globally optimal path. In a specific application embodiment, the detailed solution process of the quantum annealer is as follows:

[0117] 1) Quantum bit initialization

[0118] First, achieve a quantum superposition state that covers the entire solution space, then use the corresponding... The QUBO model obtained after the goalkeeper adaptation is applied to the target qubit, and the process is as follows:

[0119] a) Quantum bit selection: Based on the node priority list obtained from classical preprocessing, the binary variables corresponding to high-priority nodes are preferentially mapped to low-noise qubits to ensure the accuracy of solving key variables;

[0120] b) Superposition state construction: Apply an H gate to each selected qubit to change it from the ground state |0|. Transform into superposition state The superposition of 200 qubits can simultaneously cover 2^{200} potential solution spaces, ensuring that no global optimal solution is missed;

[0121] c) Initialization Verification: Measure the initial state of each qubit through the readout interface of the quantum annealer. If the purity of the superposition state ( , If the density matrix is ​​≥0.98, the initialization is successful; otherwise, it is reapplied. Gate initialization time ≤ 1ms.

[0122] 2) Annealing process control

[0123] Energy minimization based on time-varying Hamiltonians enables annealing process control. This involves addressing the energy function... By controlling annealing (Temperature, time) is used to realize the evolution of the quantum state from a superposition state to the ground state (the lowest energy state). The specific control parameters and logic are as follows:

[0124] a) Annealing time and temperature settings: Set the total annealing time while considering both real-time performance and accuracy. initial temperature (Dimensionless, corresponding to the intensity of quantum fluctuations), the temperature decreases linearly with time, ensuring that quantum fluctuations dominate in the early stage (global exploration), and that the coupling terms of the problem dominate in the later stage (converging to the optimal solution).

[0125] b) Dynamic adjustment of Hamiltonian parameters: transverse field strength , (Initial value) (Decay time constant), ensure The quantum fluctuation effect is gradually reduced by exponential decay from 100 GHz to 0; problem coupling strength. ), (Saturation value) (Inflection point time), k=10ms (growth rate), ensuring B(t) grows from 0 to 50GHz, so that the QUBO energy function (corresponding to) Gradually take the lead; constraint penalty coefficient (Benchmark value) The standard deviation of the weight coefficients of the current Q matrix. This ensures that quantum states that violate constraints (such as insufficient safe distance) have significantly increased energies and are automatically excluded.

[0126] c) Annealing process monitoring: Real-time acquisition of quantum state energy values. If the energy value remains stable within a specified duration (e.g., fluctuation ≤ 0.1%), the annealing is considered to have converged, and the annealing process can be terminated early to reduce time consumption.

[0127] 3) Quantum state measurement

[0128] After annealing, quantum measurement is used to collapse the superposition state of the qubit into a classical binary state (0 or 1), thus obtaining the optimal solution of the QUBO model and completing the transformation of the quantum state into the optimal binary solution.

[0129] a) Quantum state measurement operation: Apply a measurement gate to all qubits and record the result (i.e., 0 or 1) to measure a multidimensional binary vector x (e.g., x=[1,0,1,…,0,1]) using multiple qubits.

[0130] b) Repeated measurements and optimal solution selection: To counteract quantum measurement noise, repeated measurements are performed multiple times, and the binary vector that appears most frequently in the multiple measurements is used as the final binary optimal solution;

[0131] c) Solution Energy Verification: Calculate the QUBO energy corresponding to the optimal solution. If the energy value is lower than the preset threshold (e.g., 750), and meets the requirements... ( If the maximum energy of all measured solutions is found, then it is considered a valid optimal solution; otherwise, the annealing process is repeated.

[0132] Preferably, to improve the stability of the obtained solutions, a multi-round annealing combined with result voting method can be adopted. The detailed steps are as follows: after multiple consecutive annealings, several path solutions are obtained. The overlap rate of path nodes is calculated and analyzed. For example, a ratio of ≥80% of nodes on the same path is considered high overlap. If a solution with high overlap exists, it is selected as the final result; otherwise, the number of annealing rounds is increased, and solutions whose overlap rate is lower than a preset threshold compared to other solutions are excluded. The average of the remaining solutions is taken, and then it is determined whether a solution can be determined. If it is still not possible to directly discard a solution, the average path of several solutions with the highest scores is retained as the final result. By adopting the above-mentioned multi-round annealing combined with result voting method, the problem of traditional optimization methods such as genetic algorithms getting stuck in suboptimal paths can be avoided.

[0133] The optimal solution (i.e., the state of the qubit) is obtained from the quantum annealer. The states of the qubits 0 and 1 represent whether nodes in the path are selected or not, respectively, thus obtaining a complete path. The binary solution of the quantum output is then converted into path coordinates. When the quantum annealer outputs the final globally optimal path, classical algorithms can be used to verify the quantum solution, making the quantum solution engineering feasible and avoiding the situation where the mathematically optimal solution is not engineering feasible.

[0134] In this embodiment, step S03 updates the parameters of the improved Hamiltonian and the QUBO model based on the current environmental information, generating the latest globally optimal path. During vehicle operation, the environmental perception module continuously receives the latest environmental information and updates the model and Hamiltonian corresponding to the path planning problem accordingly. When encountering construction conditions on the preceding road segment, it updates the road topology information and obstacle location information, and also updates the generated optimal path. The real-time calculated globally optimal path solution is then sent to the path optimization module, where the route is further optimized while ensuring vehicle dynamics constraints and driving comfort.

[0135] In step S03 of this embodiment, after obtaining the globally optimal path, the process further includes checking whether the path satisfies one or more of the following constraints: dynamic constraints (e.g., turning radius ≥ minimum turning radius of the autonomous vehicle), environmental constraints (e.g., distance to obstacles ≥ safety threshold), and specified traffic rule constraints (e.g., no driving against traffic, obeying traffic lights). If the check fails, the QUBO model parameters are adjusted (e.g., the penalty coefficient is adjusted), and step S03 is re-executed to solve the problem again. Traditional quantum mechanics has poor model adaptability, and the lowest energy solution obtained by quantum annealing may only be the mathematically optimal solution. This embodiment, by considering multiple constraints such as the autonomous vehicle's dynamic constraints (turning radius, maximum acceleration, etc.), environmental constraints (safe distance to various types of obstacles), and traffic rules (prohibition of driving against traffic, violation of traffic lights), can effectively avoid the problem of the lowest energy solution not being able to be safely executed by using multi-constraint check.

[0136] In step S03 of this embodiment, when a local environmental change is detected, a local update of the Hamiltonian is triggered. This allows the quantum solution process to construct a local QUBO sub-model only for the changed road network nodes to solve the local sub-problems. In other words, the quantum part only solves the local sub-problems, achieving incremental solution. Local environmental changes include obstacle movement less than a preset threshold or congestion coefficient change rate less than a preset threshold. For example, a small obstacle movement (e.g., 1m) or a congestion coefficient change rate... c <0.2 / sBy using the incremental solution method based on local environmental changes described above, the time delay caused by the full solution can be reduced.

[0137] Specifically, when a local environmental change is detected, the affected local nodes are identified. By updating the range of the local nodes and their adjacent nodes, a local QUBO sub-model is constructed. The quantum annealer solves the local sub-model in a single round, outputs the local optimal path, and then replaces the original path with the local sub-path, retaining the path in the unchanging region. B-spline curves are used to smooth the nodes at the splicing points, making the path continuous without abrupt changes. The time consumed by this type of incremental update is much lower than that of traditional path replanning, enabling vehicles to smoothly avoid vehicles involved in traffic accidents.

[0138] Preferably, the system can also predict the trajectory of obstacles ahead (e.g., the obstacle moves 0.5 m / s northeast) based on real-time data acquired by the sensing system, and initiate quantum solution in advance to achieve predictive scheduling, thereby saving waiting time. Furthermore, quantum solution error data can be obtained using actual vehicles or simulation tests, and quantum parameters can be optimized (e.g., changing the annealing time from 50 ms to 40 ms while preserving solution accuracy) to verify the algorithm's correctness.

[0139] Preferably, the generated globally optimal path can also be fitted with a Bezier curve to ensure that the acceleration and turning angle change gradually, making the path smoother and more comfortable.

[0140] In this embodiment, step S05 specifically includes the following steps:

[0141] Step S501. Control command generation: Convert the generated global optimal path into the corresponding control command;

[0142] Step S502. Vehicle control execution: Transmit control commands to the actuators (such as electric motors and steering systems) to control the controlled vehicle to move safely along the planned route.

[0143] Specifically, after generating the global optimal path, the generated global optimal path is converted into control commands and provided to the actuator. The control module calculates the appropriate acceleration and steering angle, and provides the corresponding motor control signals and steering system control signals to the motor and steering system to control the vehicle speed and direction. Steps S01 to S04 are repeated until the controlled vehicle reaches the target position.

[0144] In summary, this application achieves globally optimal path planning for autonomous driving by integrating quantum annealing and classical computation. It employs a bidirectional guided quantum-classical hybrid architecture, utilizing classical algorithms to guide quantum focusing on the effective solution space and quantum feedback to optimize classical model parameters. This significantly reduces the number of road network nodes while improving computational efficiency and solution quality. It is adaptable to small- to medium-scale quantum annealers and can improve the accuracy of classical algorithms through quantum feedback iteration, reducing the exploration of invalid quantum solutions, lowering the deployment cost and resource consumption of onboard quantum hardware, and improving the real-time performance and efficiency of planning. Furthermore, by employing an improved Hamiltonian and utilizing the quantum tunneling effect to overcome local optima, global optimality can be stabilized. Combined with a multi-round annealing mechanism, it also avoids the problem of traditional optimization methods such as genetic algorithms falling into suboptimal paths. Furthermore, by constructing a dynamic Hamiltonian update closed loop of "real-time perception - trend prediction - parameter iteration", dynamic adaptability can be improved. Combined with incremental path updates, it can quickly respond to scenarios such as sudden congestion and obstacle movement, which can significantly improve response speed compared to traditional full path replanning. Furthermore, by using hard constraint energy barriers and soft constraint gradient penalties, vehicle dynamics (turning radius, acceleration), environmental safety (dynamic safety distance), and traffic rules (prohibition of wrong-way driving, traffic lights) are deeply coupled into the QUBO energy function, which can also avoid the problem of mathematical optimization but engineering infeasibility, thereby ensuring that autonomous driving can drive safely and efficiently in complex urban road networks.

[0145] This embodiment provides an unmanned global path planning system, including a processor and a memory. The memory is used to store computer programs, and the processor is used to execute the computer programs to perform the methods described above.

[0146] It is understood that the method described in this embodiment can be executed by a single device, such as a computer or server, or it can be applied to a distributed scenario where multiple devices cooperate to complete the task. In a distributed scenario, one of the multiple devices may execute only one or more steps of the method described in this embodiment, and the multiple devices interact to complete the method. The processor can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the method described in this embodiment. The memory can be implemented using read-only memory (ROM), random access memory (RAM), static storage devices, and dynamic storage devices. The memory can store the operating system and other applications. When the method described in this embodiment is implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0147] This embodiment further provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0148] Those skilled in the art will understand that the above embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0149] The above description is merely a preferred embodiment of this application and is not intended to limit the application in any way. Although this application has disclosed preferred embodiments above, it is not intended to limit the application. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of this application without departing from the content of the technical solution of this application should fall within the protection scope of the technical solution of this application.

Claims

1. An unmanned global path planning method based on quantum annealing optimization, characterized by the steps of The method comprises the following steps: Step S01. Environment information acquisition: acquiring real-time surrounding environment information of the controlled vehicle; Step S02. Quantum modeling: modeling the path planning problem as a quantum annealing model, using an improved Hamiltonian in the quantum annealing model and converting the improved Hamiltonian into a QUBO model, the global optimal path corresponding to the minimum energy value of the improved Hamiltonian, the global optimal path being obtained by solving the QUBO model, the improved Hamiltonian comprising a transverse field term, a problem coupling term, and a constraint penalty term for representing the degree of path violation, the transverse field term being used to drive the quantum bit to transition between the ground state and the excited state, the problem coupling term representing the degree of path optimization through energy level, and the parameters in the improved Hamiltonian varying according to real-time environment information, the parameters comprising the transverse field strength in the transverse field term, the problem coupling strength in the problem coupling term, and the constraint penalty coefficient in the constraint penalty term; Step S03. Hybrid quantum annealing-classical calculation: first, using a classical preprocessing algorithm to filter out candidate road network nodes to generate an initial path and perform priority sorting, then updating the parameters of the improved Hamiltonian and the QUBO model according to the surrounding environment information of the controlled vehicle, inputting the currently updated QUBO model into a quantum annealer for solving, using the initial path generated by the classical preprocessing algorithm and the priority sorting to guide the solving of the QUBO model, and optimizing the parameters of the classical preprocessing algorithm according to the solving result, and finally outputting the global optimal path by the quantum annealer; Step S04. Control execution: converting the currently generated global optimal path into a control instruction to control the controlled vehicle to travel; The calculation expression of the improved Hamiltonian is: wherein, is a transverse field term, is a transverse field strength at time t, is a transverse field; is a problem coupling term, is a problem coupling strength at time t, is a problem energy, the calculation expression is: , , and is a binary variable, represents passing through node , represents not passing through node , represents passing through node , represents not passing through node , is a Euclidean distance between node and node , represents a real-time travel time between node and node , is an energy consumption coefficient between node and node , is a weight coefficient; is a constraint penalty term, is a constraint penalty term coefficient at time t, is a constraint penalty; Transverse field strength Using an exponential decay model, the calculation expression is: where, t denotes time, is the initial transverse field strength, is the decay time constant; problem coupling strength Using a Sigmoid growth model, the calculation expression is: where, is the saturated coupling strength, is the inflection point time, is the growth rate constant.

2. The method of claim 1, wherein, The constraint penalty term comprises a hard constraint and a soft constraint, the hard constraint being to assign a specified energy penalty to a path that violates the prohibition of reverse driving or obstacle avoidance, and the soft constraint being to perform gradientized energy penalty when violating the turning radius constraint or the safety distance constraint to guide the quantum annealing to select a path that meets the constraint, the gradientized energy penalty being to generate a penalty amount based on the gradient according to the degree of violation of the turning radius constraint or the safety distance constraint. 3.The method of claim 1, wherein, Step S03 comprises: Step S301. Classical preprocessing of road network nodes: using a classical preprocessing algorithm to obtain a suboptimal feasible path, determining a candidate road network node range centered on the suboptimal feasible path, calculating the cost of all adjacent nodes from the current vehicle position, selecting the node with the minimum cost as the next target node, preferentially expanding the node with the minimum total cost, obtaining all candidate road network nodes, and performing priority sorting on each candidate road network node according to the state information of the candidate road network node, the state information of the candidate road network node comprising any one or more of path criticality, congestion sensitivity, and dynamic correlation degree, the dynamic correlation degree being determined according to the distance between the road network node and the dynamic obstacle. Step S302. Quantum solving: the quantum solver determines the solution space of the required solution according to the candidate road network node range, updates the parameters of the improved Hamiltonian according to the surrounding environment information of the controlled vehicle, and dynamically updates the Q matrix parameters of the QUBO model to obtain the current updated QUBO model. The current updated QUBO model is input into the quantum annealer for solving. In the solving process, the candidate path nodes and priorities obtained in step S301 are used to map each candidate path node to a quantum bit. Finally, the binary optimal solution is obtained. Step S303. Multi-round annealing optimization: a plurality of path solutions are obtained by using multi-round annealing. The final optimal solution is selected according to the coincidence rate of the path nodes.

4. The method of claim 3, wherein, In step S302, the Q matrix is also subjected to sparse processing based on the node adjacency relationship: A node adjacency matrix is constructed based on the road network topology structure of the environment in which the controlled vehicle is located. If there is a direct traffic link between two nodes, the nodes are determined to be adjacent nodes, otherwise they are determined to be non-adjacent nodes. The Q matrix elements corresponding to the adjacent nodes are retained, and the non-adjacent nodes are uniformly assigned a value of 0 to form a sparse matrix. The sparse matrix is stored, and a triple consisting of the row index, column index, and element value of the non-zero elements is recorded.

5. The method of claim 3, wherein, In step S302, the Q matrix parameters are dynamically updated, including: Step S321. Dynamic data exponential weighted smoothing calculation: adopt exponential weighted smoothing strategy, the weight of the data point in the window is , , is the decay time constant, and each weight is normalized. Step S322. Trend compensation correction: calculate the gradient change of the data in the window, predict the environmental trend according to the gradient change, and add a compensation term to optimize the path. The environmental trend is congestion trend. Step S323. Parameter calibration: set the initial values of the elements in the Q matrix, analyze the degree of change of each element in the Q matrix to analyze the sensitivity, select a plurality of elements that have the greatest influence on the QUBO model as the Q matrix key elements, and use a particle swarm optimization algorithm to iteratively optimize each Q matrix key element. For each particle consisting of a group of Q matrix key elements, solve the QUBO model by the quantum annealer, calculate the objective function after obtaining the path, and use it as the fitness value of the particle. After multiple iterations, the Q matrix key elements corresponding to the optimal position of the group are output. 6.The globally path planning method for unmanned vehicle based on quantum annealing optimization according to any one of claims 1-5, characterized in that, In step S03, before the quantum annealer solves, each candidate road network node is split into two or more sub-problems by K-means clustering. There are multiple overlapping nodes between the sub-problems. Each sub-problem is assigned to the quantum annealer in order, and the quantum annealing sub-problem is solved. When a local environmental change is detected, a local update of the improved Hamiltonian is triggered, so that only the changed road network nodes are used to construct a local QUBO sub-model to solve the local sub-problem during the quantum solving process. The local environmental change is that the obstacle amplitude moves less than a preset threshold or the congestion coefficient change rate is less than a preset threshold.

7. The method of claim 1-5, wherein, After obtaining the global optimal path in step S03, the processor further performs constraint verification on whether the path meets any one or more of the dynamic constraints, environmental constraints, and specified traffic rule constraints.

8. An unmanned global path planning system comprising a processor and a memory for storing a computer program, characterized in that, The processor is configured to execute the computer program to perform the global path planning method of the unmanned vehicle according to any one of claims 1-7.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program, when executed, implements the unmanned global path planning method according to any one of claims 1-7.

Images