A robot path planning method and system based on field-programmable gate array acceleration

Abstract

This invention discloses a robot path planning method and system based on field-programmable gate array (FPGA) acceleration, belonging to the field of path planning and embedded hardware acceleration technology. The method includes: constructing a path fitness function based on a two-dimensional occupancy grid map of the robot's working environment; iteratively optimizing a population of anchor-point quasi-reflective artificial protozoan optimization algorithms using a collaborative architecture of programmable logic and processing system, where each individual in the population represents a candidate robot path containing multiple path nodes, and the candidate robot path corresponding to the final globally optimal individual is taken as the optimal path; the programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to the individuals using the path fitness function; and the processing system side is used to update the individuals based on the fitness value of their corresponding candidate robot paths. This method can achieve fast, accurate, and stable path planning results in complex obstacle scenarios.

Description

Technical Field

[0001] This invention relates to a robot path planning method and system based on field-programmable gate array acceleration, belonging to the field of path planning and embedded hardware acceleration technology. Background Technology

[0002] Path planning is a key technology in autonomous navigation for mobile robots. Its purpose is to plan a feasible path for the robot from a starting point to a destination in a given environment, satisfying constraints and minimizing path costs as much as possible. This technology is widely used in industrial automation, warehousing and logistics, service robots, and unmanned delivery vehicles.

[0003] Existing path planning methods mainly include the A* algorithm and the Rapidly-exploring Random Tree (RRT) algorithm. The A* algorithm has good path search capabilities, but its computational cost is high in large-scale maps or complex obstacle environments, making it difficult to meet real-time planning requirements. The RRT algorithm is suitable for continuous spatial searches, but the generated paths often have many polygons and insufficient smoothness, and are prone to problems such as decreased convergence efficiency and unstable path quality in narrow passages or scenes with dense obstacles.

[0004] In recent years, swarm intelligence optimization algorithms have been introduced into the field of path planning, possessing a certain degree of global search capability. However, metaheuristic algorithms, such as Particle Swarm Optimization (PSO) and Grey Wolf Optimizer (GWO), still suffer from problems such as being prone to getting trapped in local optima and unstable convergence speed in complex obstacle environments. Furthermore, most existing path planning methods rely on serial CPU computation; when the population size is large or the number of iterations is high, the overall execution time increases significantly, making it difficult to meet the real-time requirements of online planning and embedded deployments.

[0005] Field-Programmable Gate Arrays (FPGAs) are characterized by strong parallel processing capabilities, low latency, and high reconfigurability, making them suitable for hardware acceleration of computationally intensive tasks in path planning. However, current technologies still lack sufficient hardware-software co-implementation for metaheuristic path planning algorithms, and there is a lack of an implementation scheme that balances path planning quality, computational efficiency, and the requirements of embedded real-time deployment. Summary of the Invention

[0006] The purpose of this invention is to provide a robot path planning method and system based on field-programmable gate array acceleration, which can achieve fast, accurate and stable path planning results in complex obstacle scenarios.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a robot path planning method based on field-programmable gate array acceleration, comprising: Obtain a two-dimensional occupancy grid map of the robot's working environment, and construct a path fitness function based on the two-dimensional occupancy grid map; Initialize the population for the multi-anchor quasi-reflective artificial protozoan optimization algorithm, where each individual in the population represents a robot candidate path containing multiple path nodes; The population is iteratively optimized using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met, and the robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, a quasi-reflection update is triggered.

[0008] In conjunction with the first aspect, further obtaining a two-dimensional occupancy grid map of the robot's working environment includes: The robot's working environment is divided into Each grid cell is marked as either passable free space or impassable obstacle. Robot candidate paths are , ,in, , They represent A fixed starting point and a fixed ending point , , , They represent x-coordinate and y-coordinate , They represent x-coordinate and y-coordinate and All are located in freely accessible spaces. … … They represent The 1st, ... … 1 path node , , They represent x-coordinate and y-coordinate … … The range of coordinate values ​​is constrained by the boundaries of the raster map.

[0009] Combining the first aspect, the path fitness function is further defined as follows: ; in, Indicates the robot's candidate paths, express fitness value, express Path length, , , They represent The The x and y coordinates of each path node, , They represent The The x and y coordinates of each path node, express The number of path nodes, Indicates the penalty coefficient. This indicates a penalty for violating the rules. , , , These represent penalties for violating boundary constraints, obstacles, and endpoint constraints, respectively.

[0010] Building upon the first aspect, furthermore, the foraging behavior of the multi-anchor quasi-reflective artificial protozoan optimization algorithm employs dynamic probability parameters. Switching between autotrophic and heterotrophic modes ,in, Indicates the current iteration number. This indicates the maximum number of iterations.

[0011] In conjunction with the first aspect, the multi-anchor quasi-reflection update mechanism further includes stagnation detection and quasi-reflection update; Stasis detection includes: Let the first The stall counter for the next iteration is If the first If the fitness value of the robot candidate path corresponding to the globally optimal individual in the next iteration does not improve, then... ,otherwise ;in, Indicates the first The stall counter for the next iteration; when At that time, a quasi-reflection update is triggered; among which, Indicates the stagnation threshold; After triggering the quasi-reflection update, the stall counter is decayed, with the decay magnitude determined by the decay parameter. control; Quasi-reflection updates include: According to the The dynamic scaling factor of the next iteration Determine the first The proportion of the worst-performing individuals participating in the quasi-reflective update in the next iteration; ,in, For indicator functions, if If it is established, then ,otherwise , Indicates the baseline ratio, Indicates the upper limit of the ratio. Indicates the growth coefficient; According to the The proportion of the worst-performing individual participating in the quasi-reflective update is determined in the next iteration. The number of worst-performing individuals participating in the quasi-reflective update at the next iteration , ,in, Indicates the number of individuals in a population. Indicates rounding down; Based on the normalization iteration progress Determine the first The globally optimal individual in the next iteration mean of elite groups Random elite individuals Nonnormalized weights of the three anchor points , , ; , , , ,in, Indicates the maximum number of iterations; right , , Normalization is performed to obtain , , Normalized weights of the three anchor points , , ; by , , for , , The probability of selecting the three anchor points, from , , Select the third of the three anchor points Anchor center of the next iteration ; use Will The robot candidate path corresponding to the worst individual is updated to a reflection candidate path: ; in, express Among the worst individuals, the first The robot candidate path corresponding to the worst individual Updated reflection candidate paths, express Among the worst individuals, the first The dimension mask corresponding to the worst individual. Indicates the first Quasi-reflection intensity coefficient of the next iteration. This represents element-wise multiplication; right Coordinates that violate boundary constraints are processed by first reflecting and then trunculating them to ensure that... Each path node is located within the boundary of the raster map.

[0012] In conjunction with the first aspect, the further hardware acceleration suitability assessment includes: The traversal of the robot candidate paths corresponding to each individual in the population is implemented in a pipeline manner, so that the fitness evaluation tasks are executed in an overlapping manner. The traversal of each path node in the candidate robot path corresponding to an individual is achieved by loop expansion, generating parallel hardware computing units for synchronously executing distance calculation and obstacle collision detection between multiple path nodes; The boundary constraint check function and obstacle collision detection function are integrated into the main processing flow of fitness assessment in an inline manner; The raster map data is stored in the on-chip block random access memory on the programmable logic side before path planning begins, so that it can be directly accessed for each fitness evaluation. The top-level function for fitness evaluation includes an actual population size parameter, which is used to dynamically specify the effective population size for fitness evaluation when the system side is called.

[0013] In conjunction with the first aspect, furthermore, the communication between the programmable logic side and the processing system side includes control paths and data paths; The control path uses an extensible interface. The processing system writes the robot candidate path data address, grid map data address, effective population size and control instructions to the programmable logic side through the extensible interface, triggering the kernel startup of the programmable logic side. The data path adopts a memory-mapped scalable interface. The programmable logic side reads robot candidate path data in batches from the double data rate synchronous dynamic random access memory through the memory-mapped scalable interface, and writes the fitness evaluation results back to the double data rate synchronous dynamic random access memory after completing the fitness evaluation. The processing system performs a cache write-back operation before triggering the programmable logic side kernel to start, and performs a cache invalidation operation before reading the fitness evaluation results, so as to maintain data consistency between the programmable logic side and the processing system side.

[0014] In conjunction with the first aspect, further iterative optimization includes the following steps: Step 1: The processing system initializes the population based on the grid map data and writes the robot candidate path data corresponding to the individuals in the population into the double data rate synchronous dynamic random access memory. Step 2: The processing system configures the programmable logic-side registers through the extensible interface, triggering the programmable logic-side kernel to start; Step 3: The programmable logic side reads robot candidate path data from the Double Data Rate Synchronous Dynamic Random Access Memory, performs fitness evaluation on each robot candidate path in parallel, including path length accumulation, boundary constraint checking and obstacle collision detection, obtains fitness evaluation results, and writes the fitness evaluation results back to the Double Data Rate Synchronous Dynamic Random Access Memory; Step 4: The processing system reads the fitness evaluation results, updates the global best individual and stagnation counters based on the fitness evaluation results, and performs elite group maintenance; Step 5: The processing system performs an update operation on the individual according to the update rules of the multi-anchor quasi-reflective artificial protozoan optimization algorithm, and writes the robot candidate path data corresponding to the updated individual back to the double data rate synchronous dynamic random access memory. Step 6: Repeat steps 2 to 5 until the maximum number of iterations is reached, and output the robot candidate path corresponding to the globally optimal individual as the optimal path.

[0015] In conjunction with the first aspect, further, the random numbers used in the fitness evaluation are generated by the random number generation unit in the programmable logic side kernel. The random number generation unit uses a 32-bit linear feedback shift register with taps set at bits 32, 22, 2, and 1 to generate the maximum length pseudo-random sequence. The initial seed of the 32-bit linear feedback shift register is loaded by the processing system side through an extensible interface.

[0016] Secondly, the present invention provides a robot path planning system based on field-programmable gate array acceleration, comprising: The data acquisition module is used to acquire a two-dimensional occupancy grid map of the robot's working environment and to construct a path fitness function based on the two-dimensional occupancy grid map; The initialization module is used to initialize the population of the multi-anchor quasi-reflective artificial protozoan optimization algorithm. Each individual in the population represents a robot candidate path containing multiple path nodes. The iterative optimization module is used to iteratively optimize the population using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met, and the robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, a quasi-reflection update is triggered.

[0017] Thirdly, the present invention provides a computer device, comprising: Storage medium: used to store computer programs; Processor: Used to execute the computer program to implement the robot path planning method based on field-programmable gate array acceleration as described in the first aspect.

[0018] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the robot path planning method based on field-programmable gate array acceleration as described in the first aspect.

[0019] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the robot path planning method based on field-programmable gate array acceleration described in the first aspect.

[0020] Compared with the prior art, the beneficial effects of the present invention are: The robot path planning method based on field-programmable gate array acceleration provided by this invention maps the path fitness evaluation task to the hardware acceleration execution of the programmable logic side, while retaining the control logic such as path update to be executed on the processing system side. This can reduce the single iteration execution time of the path planning task while ensuring the flexibility of the algorithm strategy, thereby completing more iterations within the same time budget and obtaining better path planning results.

[0021] The multi-anchor quasi-reflective update mechanism provided by this invention introduces stagnation detection and combines three anchor points—the global optimal solution, the mean of the elite population, and random elite individuals—to perform quasi-reflective guided updates on the population. This enhances the algorithm's global optimization ability and ability to escape local optima in complex multi-obstacle search spaces, improves the balance between exploration and utilization during path search, and thus improves the accuracy and convergence stability of path planning results.

[0022] This invention employs a high-level synthesis approach, combining pipeline, loop unrolling, and inline optimization strategies to transform path length accumulation, boundary checking, and obstacle collision detection into a parallel hardware pipeline structure. The pipeline architecture ensures that the execution time of each iteration remains highly consistent, and the worst-case execution time is predictable. This is beneficial for meeting the real-time constraints of online path planning for mobile robots and has good engineering deployment value in embedded real-time scenarios such as robot path planning and autonomous mobile platforms.

[0023] The hardware and software co-architecture of the programmable logic and processing system provided by this invention, compared with the pure software metaheuristic implementation, can reduce the execution time while maintaining the quality of path planning results, and has better real-time performance and configurability. Attached Figure Description

[0024] Figure 1 This is a flowchart of a robot path planning method based on field-programmable gate array acceleration provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of path planning obstacle perception and grid map abstraction provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the convergence of the MAQR-APO algorithm provided in this embodiment of the invention with other algorithms on test function F3, the third test function in the CEC2022 standard benchmark function test suite; Figure 4 This is a schematic diagram comparing the convergence of the MAQR-APO algorithm provided in this embodiment of the invention with other algorithms on test function F4, number 4 of the CEC2022 standard benchmark function test suite; Figure 5 This is a schematic diagram comparing the convergence of the MAQR-APO algorithm provided in this embodiment of the invention with other algorithms on test function F6, number 6 of the CEC2022 standard benchmark function test suite; Figure 6 This is a schematic diagram comparing the convergence of the MAQR-APO algorithm provided in this embodiment of the invention with other algorithms on test function F8, number 8 of the CEC2022 standard benchmark function test suite; Figure 7This is a schematic diagram showing the convergence comparison between the MAQR-APO algorithm provided in this embodiment of the invention and other algorithms on test function F12, number 12 of the CEC2022 standard benchmark function test suite. Detailed Implementation

[0025] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0026] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention. Unless otherwise specified, embodiments of the present invention and the technical features thereof can be combined with each other.

[0027] This invention provides a robot path planning method based on field-programmable gate array (FPGA) acceleration, comprising: Obtain a two-dimensional occupancy grid map of the robot's working environment, and construct a path fitness function based on the two-dimensional occupancy grid map; Initialize the population for the multi-anchor quasi-reflective artificial protozoan optimization algorithm, where each individual in the population represents a robot candidate path containing multiple path nodes; The population is iteratively optimized using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met, and the robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, a quasi-reflection update is triggered.

[0028] The robot path planning method based on field-programmable gate array acceleration provided in this invention maps the computationally intensive path fitness evaluation task to the programmable logic-side hardware acceleration kernel for parallel execution, while retaining high-level control logic such as population management, path update, and elite retention for execution on the processing system side. This reduces the execution time of a single iteration while ensuring the flexibility of the algorithm strategy. The invention also introduces a multi-anchor quasi-reflection update mechanism, which enhances the algorithm's global optimization capability in complex obstacle environments through stagnation detection and three-anchor quasi-reflection guidance. This effectively solves the problems of existing path planning methods having excessively long execution times in complex obstacle environments, being prone to getting trapped in local optima, and being difficult to deploy in real time on embedded platforms.

[0029] Figure 1 This is a flowchart of a robot path planning method based on field-programmable gate array acceleration provided in an embodiment of the present invention. This flowchart only shows the logical order of the method in the embodiment of the present invention. Provided there are no conflicts, different methods may be used. Figure 1 Complete the steps shown or described in the order indicated.

[0030] The robot path planning method based on field-programmable gate array acceleration provided in this invention can be applied to a terminal and can be executed by a robot path planning system based on field-programmable gate array acceleration. This system can be implemented by software and / or hardware and can be integrated into the terminal, such as any tablet computer or computer device with communication function.

[0031] This invention provides a robot path planning method based on field-programmable gate array (FPGA) acceleration, such as... Figure 1 As shown, the specific steps include the following: Step 1: Obtain a two-dimensional occupancy grid map of the robot's working environment, and construct a path fitness function based on the two-dimensional occupancy grid map; In this embodiment, as Figure 2 As shown, obtaining a two-dimensional occupancy grid map of the robot's working environment specifically includes: The robot's working environment is divided into There are 1 grid cell, each grid cell is marked as either passable free space or impassable obstacle. The value corresponding to passable free space is 0, and the value corresponding to impassable obstacle is 1.

[0032] In this embodiment, the path fitness function is: ; in, Indicates the robot's candidate paths, express fitness value, express Path length, , , They represent The The x and y coordinates of each path node, , They represent The The x and y coordinates of each path node, express The number of path nodes, Indicates the penalty coefficient. This indicates the penalty for violating constraints, and can comprehensively reflect the degree of violation of boundary constraints, obstacle constraints, and endpoint constraints. , , , These represent penalties for violating boundary constraints, obstacles, and endpoint constraints, respectively. Penalize nodes that extend beyond the map boundary in the path. Apply a penalty to nodes that fall into the obstacle grid. Penalties are imposed for deviations of the path's start and end points from set values. This applies when a candidate path simultaneously satisfies boundary constraints, obstacle constraints, and endpoint constraints. The fitness value degenerates into pure path length. When a candidate path violates a constraint, the penalty term takes a larger value, causing infeasible paths to be eliminated during the selection process.

[0033] Specifically, Set it to 1000.

[0034] The path fitness function provided in this embodiment incorporates the minimization of path length with the degree of violation of boundary constraints, obstacle constraints, and endpoint constraints into the same optimization objective. This allows the population to converge toward a short path that satisfies all constraints without increasing algorithm complexity, thus adapting to path planning tasks with varying complexity and obstacle density.

[0035] Step 2: Initialize the population for the multi-anchor quasi-reflective artificial protozoan optimization algorithm. Each individual in the population represents a candidate robot path containing multiple path nodes. In this embodiment, the robot's candidate path is , ,in, , They represent A fixed starting point and a fixed ending point , , , They represent x-coordinate and y-coordinate , They represent x-coordinate and y-coordinate and All are located in freely accessible spaces. … … They represent The 1st, ... … 1 path node , , They represent x-coordinate and y-coordinate … … The range of coordinate values ​​is constrained by the boundaries of the raster map.

[0036] The path nodes for each initial candidate path are randomly generated within the coordinate range of the raster map, ensuring that the coordinates of the path nodes satisfy the map boundary constraints.

[0037] Specifically, Set the population size to 58 and the population size to 100.

[0038] The Multi-Anchor Quasi-Reflection Artificial Protozoa Optimizer (MAQR-APO) algorithm is based on the Artificial Protozoa Optimizer (APO) algorithm. It treats each individual in the population as a candidate solution in the search space and achieves population renewal through foraging behavior, hibernation behavior, and reproductive behavior.

[0039] In this embodiment, the foraging behavior of the multi-anchor quasi-reflection artificial protozoan optimization algorithm adopts dynamic probability parameters. Switching between autotrophic and heterotrophic modes ,in, Indicates the current iteration number. This indicates the maximum number of iterations.

[0040] As the iterations gradually decrease, the MAQR-APO algorithm transitions from focusing on global exploration in the early stages to focusing on local exploration in the later stages.

[0041] To address the issues of premature convergence and local stagnation that the standard APO algorithm is prone to in complex multimodal problems, the MAQR-APO algorithm introduces a multi-anchor quasi-reflective update mechanism. During the iterative optimization process, when the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, a quasi-reflective update is triggered.

[0042] In this embodiment, the multi-anchor quasi-reflection update mechanism includes stagnation detection and quasi-reflection update.

[0043] Stasis detection includes: Let the first The stall counter for the next iteration is If the first If the fitness value of the robot candidate path corresponding to the globally optimal individual in the next iteration does not improve, then... ,otherwise ;in, Indicates the first The stall counter for the next iteration; when At that time, a quasi-reflection update is triggered; among which, Indicates the stagnation threshold; After triggering the quasi-reflection update, to avoid frequent re-triggers in the later stages of the search, the stall counter is decayed. The decay rate is determined by the decay parameter. control.

[0044] Quasi-reflection updates include: According to the The dynamic scaling factor of the next iteration Determine the first The proportion of the worst-performing individuals participating in the quasi-reflective update in the next iteration; ,in, For indicator functions, if If it is established, then ,otherwise , Indicates the baseline ratio, Indicates the upper limit of the ratio. Indicates the growth coefficient; According to the The proportion of the worst-performing individual participating in the quasi-reflective update is determined in the next iteration. The number of worst-performing individuals participating in the quasi-reflective update at the next iteration , ,in, Indicates the number of individuals in a population. Indicates rounding down; Based on the normalization iteration progress Determine the first The globally optimal individual in the next iteration mean of elite groups Random elite individuals Nonnormalized weights of the three anchor points , , ; , , , ,in, Indicates the maximum number of iterations; right , , Normalization is performed to obtain , , Normalized weights of the three anchor points , , ; by , , for , , The probability of selecting the three anchor points, from , , Select the third of the three anchor points Anchor center of the next iteration This enables the MAQR-APO algorithm to maintain strong exploration capabilities in the early stages and enhance its ability to refine the development of advantageous areas in the later stages. use Will The robot candidate path corresponding to the worst individual is updated to a reflection candidate path: ; in, express Among the worst individuals, the first The robot candidate path corresponding to the worst individual Updated reflection candidate paths, express Among the worst individuals, the first The dimension mask corresponding to the worst individual. Indicates the first Quasi-reflection intensity coefficient of the next iteration. This represents element-wise multiplication; right Coordinates that violate boundary constraints are processed by first reflecting and then trunculating them to ensure that... Each path node is located within the boundary of the raster map.

[0045] The MAQR-APO algorithm provided in this embodiment does not completely replace the standard APO algorithm. Instead, while keeping the number of function evaluations per generation constant, it optimizes the algorithm for the worst-fitting function when the stagnation condition is met. Each individual performs a MAQR update, replacing the original inefficient global random generation strategy, which improves the stability and convergence quality of path search without increasing additional computational overhead.

[0046] Step 3: Iteratively optimize the population using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met, and take the robot candidate path corresponding to the final globally optimal individual as the optimal path.

[0047] In this embodiment, the programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using a path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual.

[0048] In this embodiment, the fitness evaluation of hardware acceleration specifically includes: The traversal of the robot candidate paths corresponding to each individual in the population is implemented in a pipeline manner, so that the fitness evaluation tasks are executed in an overlapping manner. The traversal of each path node in the candidate robot path corresponding to an individual is achieved by loop expansion, generating parallel hardware computing units for synchronously executing distance calculation and obstacle collision detection between multiple path nodes; The boundary constraint check function and obstacle collision detection function are integrated into the main processing flow of fitness assessment in an inline manner; The raster map data is stored in the on-chip block random access memory on the programmable logic side before path planning begins, so that it can be directly accessed for each fitness evaluation. The top-level function for fitness evaluation includes an actual population size parameter, which is used to dynamically specify the effective population size for fitness evaluation when the system side is called.

[0049] In this embodiment, the communication between the programmable logic side and the processing system side includes a control path and a data path: The control path uses an extensible interface. The processing system writes the robot candidate path data address, grid map data address, effective population size and control instructions to the programmable logic side through the extensible interface, triggering the kernel startup of the programmable logic side. The data path adopts a memory-mapped scalable interface. The programmable logic side reads robot candidate path data in batches from the double data rate synchronous dynamic random access memory through the memory-mapped scalable interface, and writes the fitness evaluation results back to the double data rate synchronous dynamic random access memory after completing the fitness evaluation. The processing system performs a cache write-back operation before triggering the programmable logic side kernel to start, and performs a cache invalidation operation before reading the fitness evaluation results, so as to maintain data consistency between the programmable logic side and the processing system side.

[0050] Specifically, the Processing System (PS) includes Advanced RISC Machines (ARM), responsible for main iteration scheduling, population management, path updates, and result output; Programmable Logic (PL) acts as a hardware accelerator, responsible for performing parallel fitness evaluation on candidate paths; the PS and PL communicate through an Advanced eXtensible Interface (AXI), the control path uses an Advanced eXtensible Interface Lite (AXI-Lite), and the data path interacts through Double Data Rate Synchronous Dynamic Random Access Memory (DDR).

[0051] The division of tasks between the PS side and the PL side is as follows: The PS side initializes the candidate path population based on map information, stores the path node coordinates in DDR, executes the main iteration loop scheduling, writes the path data address, population size and map information address to the PL side through AXI-Lite, triggers the PL side hardware acceleration kernel to start, and reads the fitness values ​​of each path returned by the PL side from DDR, and performs global optimal update, stagnation counter management and elite group maintenance. The PL side receives the configuration parameters from the PS side, reads the node coordinate data of all candidate paths in the current generation in batches from the DDR through M-AXI, performs path length accumulation, boundary constraint check and obstacle collision detection in parallel for each candidate path, calculates the fitness value and writes it back to the DDR, and sends a completion signal to the PS side. When data is exchanged between the PS side and the PL side, the PS side performs a cache write-back operation before triggering the PL side kernel to ensure that the path data in DDR is the latest value; after the PL side completes the calculation and returns, the PS side performs a cache invalidation operation to read the latest fitness result written back to DDR by the PL side, thereby ensuring data consistency between the PS side and the PL side.

[0052] In this embodiment, the iterative optimization process specifically includes the following steps: Step 1: Initialize the grid map on the processing system side, mark the location of obstacles, set the start coordinates and end coordinates, initialize the population according to the grid map data, and write the robot candidate path data corresponding to the individuals in the population into the double data rate synchronous dynamic random access memory. Step 2: The processing system configures the programmable logic side registers through the extensible interface, writes the path data base address, map data base address and effective population size, and triggers the programmable logic side kernel to start; Step 3: The programmable logic side reads robot candidate path data from the Double Data Rate Synchronous Dynamic Random Access Memory, performs fitness evaluation in parallel on each robot candidate path, including path length accumulation, boundary constraint checking and obstacle collision detection, calculates the fitness value, obtains the fitness evaluation result, and writes the fitness evaluation result back to the Double Data Rate Synchronous Dynamic Random Access Memory; Step 4: The processing system reads the fitness evaluation results, updates the global best individual and stagnation counters based on the fitness evaluation results, and performs elite group maintenance; Step 5: The processing system performs an update operation on the individual according to the update rules of the multi-anchor quasi-reflective artificial protozoan optimization algorithm, and writes the robot candidate path data corresponding to the updated individual back to the double data rate synchronous dynamic random access memory. Step 6: Repeat steps 2 to 5 until the maximum number of iterations is reached, and output the robot candidate path corresponding to the globally optimal individual as the optimal path.

[0053] In one possible embodiment, the PL-side path fitness evaluation kernel is implemented using a High-Level Synthesis (HLS) approach: PIPELINE pipeline optimization instructions are applied to the outer population traversal loop to reduce loop start intervals and improve data throughput; UNROLL loop unrolling optimization instructions are applied to the inner loops of path nodes to generate parallel hardware resources, enabling synchronous execution of multi-node distance calculation and collision detection; INLINE optimization instructions are applied to boundary checking and collision detection auxiliary functions to reduce function call overhead and promote integration with the main pipeline. Raster map data is stored once in the PL-side Block Random Access Memory (BRAM) during system initialization, avoiding repeated readings from DDR in each iteration, thereby reducing M-AXI bus bandwidth usage. Furthermore, an actual population size parameter is introduced into the HLS top-level function interface, allowing the PS side to dynamically adjust the effective population size participating in the evaluation at runtime, improving system adaptability.

[0054] In one possible embodiment, the random numbers used in fitness evaluation are generated by a random number generation unit in the programmable logic side kernel. The random number generation unit uses a 32-bit linear feedback shift register (LFSR), with taps set at bits 32, 22, 2, and 1 to generate the maximum length pseudo-random sequence. The initial seed of the 32-bit linear feedback shift register is loaded by the processing system side through an extensible interface to ensure that different running processes have different pseudo-random sequences, meeting the randomness requirements of the path planning population initialization and random decision-making stages.

[0055] After synthesis, the generated hardware acceleration core with standard bus interface is integrated with the processing system and programmable logic into a complete system-on-a-chip. The processing system is responsible for task scheduling and control, while the programmable logic serves as the main hardware computing unit. The two communicate with each other through bus interconnection and reset management infrastructure.

[0056] The robot path planning method based on field-programmable gate array (FPGA) acceleration provided in this invention also supports general intelligent optimization computing scenarios. In these scenarios, the standard benchmark function is used as the fitness function, and a kernel-oriented unloading mode is adopted. The complete single-generation computation kernel of the MAQR-APO algorithm is unloaded to the PL side for execution. The PL side supports dynamic writing of test function identifiers, random number seeds, effective population size, and data exchange address information by the PS side at runtime via AXI-Lite, thereby supporting batch testing and rapid verification under different configuration conditions.

[0057] To verify the comprehensive performance of the robot path planning method based on field-programmable gate array (FPGA) acceleration provided in this embodiment of the invention in terms of execution efficiency and path planning quality, a hardware acceleration experiment was conducted on a development board and compared with a pure software implementation. The experiment was conducted on a 30×30 grid map, with the starting point set to (0,0) and the ending point set to (29,29). The population size was set to 100, and the number of iterations were set to 200, 500, and 1000 times, respectively. Each method was repeated 30 times, and the average execution time and optimal fitness value were recorded.

[0058] The comparison method involves comparing the execution time and path planning quality of a pure software implementation (denoted as MAQR-APO-SW) that serially executes the entire calculation process of the MAQR-APO algorithm on an ARM processor with the hardware and software co-implementation of this invention (denoted as MAQR-APO-HW).

[0059] As shown in Table 1, under the conditions of 200, 500 and 1000 iterations, the hardware-software co-implementation achieved execution time speedup ratios of 1.30, 1.29 and 1.27 compared to the pure software implementation, respectively. At the same time, the optimal fitness values ​​decreased from 57.0, 57.0 and 55.0 to 53.0, 45.0 and 45.0, respectively, demonstrating significant advantages in both execution efficiency and solution quality.

[0060] Table 1: Comparison of Software / Hardware Execution Time, Speedup Ratio, and Optimal Fitness in Path Planning Applications .

[0061] As shown in Table 1, for a given population size, the execution time of the hardware implementation remains highly consistent across different iterations, demonstrating the time determinism of the FPGA pipeline architecture. This characteristic makes the worst-case execution time of the system predictable, which is beneficial for meeting the real-time constraints of online path planning for mobile robots and has significant engineering practical value in embedded real-time scenarios such as online robot planning and autonomous mobile platforms.

[0062] To further compare the execution characteristics of the two task offloading modes in path planning scenarios, the kernel offloading mode offloads the entire single-generation computation process of the MAQR-APO algorithm to the PL-side kernel for execution. The PS side is only responsible for parameter configuration and result reading. This mode is suitable for standardized test scenarios with fixed benchmark functions and features fewer interactions between the PS and PL sides and lower scheduling overhead. The task offloading mode only offloads the computationally intensive fitness evaluation to the PL side, while the high-level control logic remains on the PS side. This mode is suitable for application scenarios such as path planning that require flexible adjustments to upper-level control strategies. The PS side can dynamically adjust map data, start-end point settings, or path constraint parameters based on real-time obstacle perception results, demonstrating the system's online reconfiguration capability. In path planning applications, the task offloading mode can maintain hardware acceleration benefits while facilitating system adaptation to path planning tasks of varying complexity and dynamism.

[0063] To further verify the effectiveness of the MAQR-APO algorithm and its FPGA hardware and software co-implementation provided in this embodiment of the invention in general optimization tasks, simulation software was used for simulation, hardware acceleration experiments were conducted on the development board, and the CEC2022 standard benchmark function test suite was selected for verification.

[0064] As shown in Table 2, the MAQR-APO algorithm is compared with mainstream algorithms such as APO, Arithmetic Optimization Algorithm (AOA), Differential Evolution (DE), Grasshopper Optimization Algorithm (GOA), GWO, PSO, and Whale Optimization Algorithm (WOA) on the CEC2022 standard benchmark test suite.

[0065] Table 2: Comparison results of MAQR-APO algorithm and mainstream algorithms on the CEC2022 standard benchmark test suite .

[0066] In Table 2, F1 to F12 represent test functions 1 to 12 in the CEC2022 standard benchmark test suite, respectively. (=) indicates that there is no significant difference between the comparison algorithm and the MAQR-APO algorithm, (+) indicates that the comparison algorithm is significantly inferior to the MAQR-APO algorithm, and (-) indicates that the comparison algorithm is significantly superior to the MAQR-APO algorithm.

[0067] As shown in Table 2, the MAQR-APO algorithm provided in this embodiment of the invention exhibits good comprehensive optimization performance on the CEC2022 standard benchmark function test suite. This indicates that the stagnation detection mechanism and multi-anchor quasi-reflection update mechanism introduced in this embodiment of the invention can improve the premature stagnation phenomenon in complex multi-peak problems and improve the solution quality and robustness of the algorithm.

[0068] As shown in Table 3, under the same population size and test function conditions, the FPGA-based hardware and software co-implementation significantly shortens the execution time compared to the pure software implementation, while maintaining the same solution quality.

[0069] Table 3: Comparison results between pure software execution and hardware accelerated execution .

[0070] In Table 3, F1 to F6 represent test functions 1 to 6 in the CEC2022 standard benchmark function test suite, respectively.

[0071] As shown in Table 3, under the same test conditions, the FPGA-based hardware and software co-implementation has better execution efficiency than the pure software implementation. The hardware and software co-implementation framework proposed in this embodiment of the invention achieves stable acceleration at the system level while maintaining the quality of understanding, indicating that the invention has optimization effects at the algorithm level and can demonstrate good acceleration capabilities and engineering deployment value at the system implementation level.

[0072] To further demonstrate the convergence behavior of the robot path planning method accelerated by field-programmable gate arrays provided in this embodiment of the invention, a representative test function was selected to plot the convergence curve. For example... Figures 3 to 7 As shown, the convergence processes of the robot path planning method based on field-programmable gate array acceleration provided in this embodiment of the invention are compared with those of other mainstream algorithms on test functions F3, F4, F6, F8, and F12 of the CEC2022 standard benchmark function test suite. (See Table 2 for details.) Figures 3 to 7 As can be seen, the robot path planning method based on field-programmable gate array acceleration provided by the embodiments of the present invention not only has a better overall performance in terms of the quality of the final solution, but also can enter the advantageous search region earlier in the convergence process and maintain a relatively stable optimization trend in subsequent iterations, thereby avoiding the premature stagnation phenomenon that may occur in the standard APO in complex multimodal problems.

[0073] This invention provides a robot path planning system based on field-programmable gate array (FPGA) acceleration, comprising: The data acquisition module is used to acquire a two-dimensional occupancy grid map of the robot's working environment and to construct a path fitness function based on the two-dimensional occupancy grid map; The initialization module is used to initialize the population of the multi-anchor quasi-reflective artificial protozoan optimization algorithm. Each individual in the population represents a robot candidate path containing multiple path nodes. The iterative optimization module is used to iteratively optimize the population using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met, and the robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, a quasi-reflection update is triggered.

[0074] The robot path planning system based on field-programmable gate array acceleration provided in this embodiment of the invention can execute the robot path planning method based on field-programmable gate array acceleration provided in this embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0075] This invention provides a computer device, comprising: Storage medium: used to store computer programs; Processor: Used to execute computer programs to implement the robot path planning method based on field-programmable gate array acceleration provided in the embodiments of the present invention.

[0076] This invention provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the robot path planning method based on field-programmable gate array acceleration provided in this invention.

[0077] This invention provides a computer program product, including a computer program that, when executed by a processor, implements the robot path planning method based on field-programmable gate array acceleration provided in this invention.

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0080] These 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 function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment 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.

[0082] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A robot path planning method based on field-programmable gate array acceleration, characterized in that, include: Obtain a two-dimensional occupancy grid map of the robot's working environment, and construct a path fitness function based on the two-dimensional occupancy grid map; Initialize the population for the multi-anchor quasi-reflective artificial protozoan optimization algorithm, where each individual in the population represents a robot candidate path containing multiple path nodes; The population is iteratively optimized using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met. The robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, quasi-reflection update is triggered. The number of worst individuals participating in quasi-reflection update is determined according to the dynamic scaling factor. The globally optimal individual, the mean of the elite group, and random elite individuals are used as anchor points. The anchor point center is selected according to the normalized weight of each anchor point. The robot candidate path corresponding to the worst individual participating in quasi-reflection update is updated as a reflection candidate path using the anchor point. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. The fitness assessment for hardware acceleration includes: The traversal of the robot candidate paths corresponding to each individual in the population is implemented in a pipeline manner, so that the fitness evaluation tasks are executed in an overlapping manner. The traversal of each path node in the candidate robot path corresponding to an individual is achieved by loop expansion, generating parallel hardware computing units for synchronously executing distance calculation and obstacle collision detection between multiple path nodes; The boundary constraint check function and obstacle collision detection function are integrated into the main processing flow of fitness assessment in an inline manner; The raster map data is stored in the on-chip block random access memory on the programmable logic side before path planning begins, so that it can be directly accessed for each fitness evaluation.

2. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, Obtaining a two-dimensional occupancy grid map of the robot's working environment includes: The robot's working environment is divided into Each grid cell is marked as either passable free space or impassable obstacle. Robot candidate paths are , ,in, , They represent A fixed starting point and a fixed ending point , , , They represent x-coordinate and y-coordinate , They represent x-coordinate and y-coordinate and All are located in freely accessible spaces. … … They represent The 1st, ... … 1 path node , , They represent x-coordinate and y-coordinate … … The range of coordinate values ​​is constrained by the boundaries of the raster map.

3. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, The path fitness function is: ； in, Indicates the robot's candidate paths, express fitness value, express Path length, , , They represent The The x and y coordinates of each path node, , They represent The The x and y coordinates of each path node, express The number of path nodes, Indicates the penalty coefficient. This indicates a penalty for violating the rules. , , , These represent penalties for violating boundary constraints, obstacles, and endpoint constraints, respectively.

4. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, The foraging behavior of the multi-anchor quasi-reflection artificial protozoan optimization algorithm adopts dynamic probability parameters. Switching between autotrophic and heterotrophic modes ,in, Indicates the current iteration number. This indicates the maximum number of iterations.

5. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, The multi-anchor quasi-reflection update mechanism includes stagnation detection and quasi-reflection update; Stasis detection includes: Let the first The stall counter for the next iteration is If the first If the fitness value of the robot candidate path corresponding to the globally optimal individual in the next iteration does not improve, then... ,otherwise ;in, Indicates the first The stall counter for the next iteration; when At that time, a quasi-reflection update is triggered; among which, Indicates the stagnation threshold; After triggering the quasi-reflection update, the stall counter is decayed, with the decay magnitude determined by the decay parameter. control; Quasi-reflection updates include: According to the The dynamic scaling factor of the next iteration Determine the first The proportion of the worst-performing individuals participating in the quasi-reflective update in the next iteration; ,in, For indicator functions, if If it is established, then ,otherwise , Indicates the baseline ratio, Indicates the upper limit of the ratio. Indicates the growth coefficient; According to the The proportion of the worst-performing individual participating in the quasi-reflective update is determined in the next iteration. The number of worst-performing individuals participating in the quasi-reflective update at the next iteration , ,in, Indicates the number of individuals in a population. Indicates rounding down; Based on the normalization iteration progress Determine the first The globally optimal individual in the next iteration mean of elite groups Random elite individuals Nonnormalized weights of the three anchor points , , ; , , , ,in, Indicates the maximum number of iterations; right , , Normalization is performed to obtain , , Normalized weights of the three anchor points , , ; by , , for , , The probability of selecting the three anchor points, from , , Select the third of the three anchor points Anchor center of the next iteration ; use Will The robot candidate path corresponding to the worst individual is updated to a reflection candidate path: ； in, express Among the worst individuals, the first The robot candidate path corresponding to the worst individual Updated reflection candidate paths, express Among the worst individuals, the first The dimension mask corresponding to the worst individual. Indicates the first Quasi-reflection intensity coefficient of the next iteration. This represents element-wise multiplication; right Coordinates that violate boundary constraints are processed by first reflecting and then trunculating them to ensure that... Each path node is located within the boundary of the raster map.

6. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, The top-level function for fitness evaluation includes the actual population size parameter, which is used to dynamically specify the effective population size for fitness evaluation when processing system-side calls.

7. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, Communication between the programmable logic side and the processing system side includes control paths and data paths; The control path uses an extensible interface. The processing system writes the robot candidate path data address, grid map data address, effective population size and control instructions to the programmable logic side through the extensible interface, triggering the kernel startup of the programmable logic side. The data path adopts a memory-mapped scalable interface. The programmable logic side reads robot candidate path data in batches from the double data rate synchronous dynamic random access memory through the memory-mapped scalable interface, and writes the fitness evaluation results back to the double data rate synchronous dynamic random access memory after completing the fitness evaluation. The processing system performs a cache write-back operation before triggering the programmable logic side kernel to start, and performs a cache invalidation operation before reading the fitness evaluation results, so as to maintain data consistency between the programmable logic side and the processing system side.

8. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, Iterative optimization includes the following steps: Step 1: The processing system initializes the population based on the grid map data and writes the robot candidate path data corresponding to the individuals in the population into the double data rate synchronous dynamic random access memory. Step 2: The processing system configures the programmable logic-side registers through the extensible interface, triggering the programmable logic-side kernel to start; Step 3: The programmable logic side reads robot candidate path data from the Double Data Rate Synchronous Dynamic Random Access Memory, performs fitness evaluation on each robot candidate path in parallel, including path length accumulation, boundary constraint checking and obstacle collision detection, obtains fitness evaluation results, and writes the fitness evaluation results back to the Double Data Rate Synchronous Dynamic Random Access Memory; Step 4: The processing system reads the fitness evaluation results, updates the global best individual and stagnation counters based on the fitness evaluation results, and performs elite group maintenance; Step 5: The processing system performs an update operation on the individual according to the update rules of the multi-anchor quasi-reflective artificial protozoan optimization algorithm, and writes the robot candidate path data corresponding to the updated individual back to the double data rate synchronous dynamic random access memory. Step 6: Repeat steps 2 to 5 until the maximum number of iterations is reached, and output the robot candidate path corresponding to the globally optimal individual as the optimal path.

9. The robot path planning method based on field-programmable gate array acceleration according to claim 1, characterized in that, The random numbers used in fitness evaluation are generated by the random number generation unit in the programmable logic side kernel. The random number generation unit uses a 32-bit linear feedback shift register with taps set at bits 32, 22, 2, and 1 to generate the maximum length pseudo-random sequence. The initial seed of the 32-bit linear feedback shift register is loaded by the processing system side through an extensible interface.

10. A robot path planning system based on field-programmable gate array acceleration, characterized in that, include: The data acquisition module is used to acquire a two-dimensional occupancy grid map of the robot's working environment and to construct a path fitness function based on the two-dimensional occupancy grid map; The initialization module is used to initialize the population of the multi-anchor quasi-reflective artificial protozoan optimization algorithm. Each individual in the population represents a robot candidate path containing multiple path nodes. The iterative optimization module is used to iteratively optimize the population using a collaborative architecture of programmable logic and processing system until the iteration termination condition is met. The robot candidate path corresponding to the final globally optimal individual is taken as the optimal path. During the iterative optimization process, a multi-anchor quasi-reflection update mechanism is adopted. When the fitness value of the robot candidate path corresponding to the globally optimal individual has not improved for several consecutive generations, quasi-reflection update is triggered. The number of worst individuals participating in quasi-reflection update is determined according to the dynamic scaling factor. The globally optimal individual, the mean of the elite group, and random elite individuals are used as anchor points. The anchor point center is selected according to the normalized weight of each anchor point. The robot candidate path corresponding to the worst individual participating in quasi-reflection update is updated to a reflection candidate path using the anchor point. The programmable logic side is used to perform hardware-accelerated fitness evaluation of the candidate robot paths corresponding to an individual using the path fitness function, and obtain the fitness value of the candidate robot paths corresponding to the individual; the processing system side is used to update the individual based on the fitness value of the candidate robot paths corresponding to the individual. The fitness assessment for hardware acceleration includes: The traversal of the robot candidate paths corresponding to each individual in the population is implemented in a pipeline manner, so that the fitness evaluation tasks are executed in an overlapping manner. The traversal of each path node in the candidate robot path corresponding to an individual is achieved by loop expansion, generating parallel hardware computing units for synchronously executing distance calculation and obstacle collision detection between multiple path nodes; The boundary constraint check function and obstacle collision detection function are integrated into the main processing flow of fitness assessment in an inline manner; The raster map data is stored in the on-chip block random access memory on the programmable logic side before path planning begins, so that it can be directly accessed for each fitness evaluation.

Images