Multi-picking-hand path planning control system of tea tender shoot picking robot

Abstract

The invention discloses a multi-picking-hand path planning control system of a tea tender shoot picking robot, and the system obtains and processes a tea tender shoot picture through a data obtaining module, and determines the coordinate information of a picking target and the initial coordinate information of each picking hand through an initialization module. An initial candidate path of each picking hand is obtained through a path preliminary planning module based on a preset reinforcement learning algorithm, collision detection is performed on the initial candidate paths through an optimization module, so that an optimized candidate path of each picking hand is determined, and the optimized candidate paths are screened by using a genetic algorithm to obtain an optimal picking path. The method improves the processing capability of a control system on complex conditions, reduces the dependence of a genetic algorithm on a fitness function and global information, greatly improves the robustness and generalization of a control algorithm, can guarantee the quality of tea tender shoots and the safety of a picking hand, and is suitable for popularization and application. And a good technical support is provided for application and popularization of mechanical picking of tea tender shoots.

Description

Technical Field

[0001] The present application relates to the technical field of tea picking robot control, and in particular to a multi-picking hand path planning control system, method and computer-readable storage medium for a tea bud picking robot. Background Art

[0002] As people pay more attention to health and understand the health properties of tea, the number of tea drinkers is gradually increasing, and the demand for tea consumption is increasing. The technological innovation of the tea industry is accelerating, the categories of tea are constantly enriched, the quality of tea is constantly improving, and the multiple connotations of tea will be brought into play. However, the aging trend of the domestic agricultural labor force is intensifying, and the difficulty in recruiting workers and the high cost of labor have become bottlenecks restricting the development of the tea industry. The research and development of unmanned picking equipment is imminent. The research and development and promotion of intelligent tea picking robots have increased the output of tea and reduced the cost of tea picking. To achieve the refined picking of tea picking robots, it is necessary to solve key problems such as precise picking control of tea buds, efficient picking, path planning and intelligent decision-making. Among them, the path planning of picking young buds is the core element of the research on tea young bud picking equipment. In order to complete the tea picking task efficiently and accurately, the picking hand needs to move accurately to the position of the identified tea buds in the workspace, which requires the optimal path of the picking hand from the starting position to the target position to be determined through path planning.

[0003] The existing path planning of picking hands usually uses bionic algorithms. Since such algorithms are based on biological methods for planning, they are inevitably subject to biological constraints. For example, the path selection in the initial state of the ant colony algorithm is completely random, and the subsequent path selection will be affected by the selection in the previous step. In some cases, the results will become unpredictable and easily fall into local optimality, resulting in problems such as long picking paths, slow speeds, and low efficiency. In addition, due to the poor robustness of machine learning algorithms, if changes occur during the movement, they need to be recalculated based on the map. All of the above defects will cause the current picking hand control system to be unable to quickly and accurately pick tea buds, and there is also a risk of damaging tea leaves or picking hands.

[0004] Therefore, there is an urgent need for a multi-picking hand path planning and control system for a tea bud picking robot that can quickly and effectively generate a picking path while also taking into account the collaboration of multiple picking hands and obstacles in the environment to ensure the picking quality of tea and the safety of the picking hands. Summary of the invention

[0005] The present application provides a multi-picking hand path planning and control system for a tea bud picking robot, which is used to solve the technical problems in the prior art that the picking robot control system is prone to fall into a local optimal solution when generating a picking path, resulting in the picking path being too long and the picking speed being slow, and the picking path safety is not tested, resulting in a high risk of damage to the picking hands and tea buds.

[0006] In a first aspect, the present application provides a multi-picking hand path planning and control system for a tea bud picking robot, comprising a data acquisition module, an initialization module, a path preliminary planning module, an optimization module and a picking hand control module connected in sequence; The data acquisition module is used to acquire the tea buds picture in the current scene and pre-process the tea buds picture; the pre-processing includes image denoising and image size adjustment; The initialization module is used to perform target detection on the preprocessed tea bud images using a preset visual detection model to determine the coordinate information of the picking target; it is also used to calibrate the positions of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; A preliminary path planning module is used to obtain an initial candidate path for each picker based on the initial coordinate information of each picker and the picking target coordinate information based on a preset reinforcement learning algorithm; The optimization module is used to perform collision detection on the initial candidate paths, determine the optimized candidate paths for each picker according to the collision detection results, and use the genetic algorithm to screen the preferred candidate paths to determine the optimal picking path for each picker; The picking hand control module is used to send picking hand control instructions to the picking robot according to the optimal picking path, so that the picking robot can complete the picking of tea buds.

[0007] In a second aspect, the present application further provides a multi-picking hand path planning control method for a tea bud picking robot, which is implemented by using the multi-picking hand path planning control system of the tea bud picking robot described in the above technical solution, and the method comprises: Get the tea buds picture in the current scene, and pre-process the tea buds picture; The preset visual detection model is used to detect the target of the preprocessed tea buds image and determine the coordinate information of the picking target; it is also used to calibrate the position of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; Based on the preset reinforcement learning algorithm, the initial candidate path of each picking hand is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information; Perform collision detection on the initial candidate paths, determine the optimized candidate paths for each picker based on the collision detection results, and use genetic algorithms to screen the preferred candidate paths to determine the optimal picking path for each picker; According to the optimal picking path, picking hand control instructions are sent to the picking robot, so that the picking robot completes the picking of tea buds.

[0008] Furthermore, the preset reinforcement learning algorithm is a Q-learning algorithm; Q-learning algorithm: Q-learning algorithm is a reinforcement learning algorithm that allows the intelligent agent to learn an optimal strategy through interaction with the environment without knowing the model of the environment, that is, what action to take in each state to obtain the maximum cumulative reward. The Q-learning algorithm is a classic algorithm proposed by Watkin. This is a model-free learning method that provides an intelligent system with a learning ability to select the optimal action in a Markov environment using the experienced action sequence. Now, the application of Q-learning in fields such as reinforcement learning is becoming more and more mature, and it has been greatly applied in path planning problems, fitness trajectory problems, etc. Related research includes the application of Q-learning algorithm in path planning of robotic arms.

[0009] The preset reinforcement learning algorithm is a Q-learning algorithm; Based on the preset reinforcement learning algorithm, the initial candidate path is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information, including: The Q function is solved with the initial coordinates of the picking hand as the initial state and the picking target coordinates as the target state, and the initial candidate path is determined based on the execution action and state value of the picking hand; The Q function is expressed by the formula: Q Π (s,a)=Q(s,a)+α[r+γmax a’ Q(s',a')-Q(s,a)] In the formula, Q(s, a) represents the Q value of taking action a in state s, α represents the learning rate, r represents the immediate reward brought by the current action, γ represents the discount factor for future rewards, and max a’ Q(s',a') represents the Q value of taking the optimal action a' in the new state s', Q Π (s,a) represents the Q value obtained by taking action a in state s according to strategy Π; Among them, strategy Π includes real-time strategy and random strategy. The real-time strategy is used to obtain the optimal solution of the current state, and the random strategy is used to obtain the optimal solution of the future state.

[0010] Based on the preset reinforcement learning algorithm, the initial candidate path is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information, including: The Q function is solved with the initial coordinates of the picking hand as the initial state and the picking target coordinates as the target state, and the initial candidate path is determined based on the execution action and state value of the picking hand; The Q function is expressed by the formula: Q Π(s,a)=Q(s,a)+α[r+γmax a’ Q(s',a')-Q(s,a)] In the formula, Q(s, a) represents the Q value of taking action a in state s, α represents the learning rate, r represents the immediate reward brought by the current action, γ represents the discount factor for future rewards, and max a’ Q(s',a') represents the Q value of taking the optimal action a' in the new state s', Q Π (s,a) represents the Q value obtained by taking action a in state s according to strategy Π; Among them, strategy Π includes real-time strategy and random strategy. The real-time strategy is used to obtain the optimal solution of the current state, and the random strategy is used to obtain the optimal solution of the future state.

[0011] Furthermore, the instant strategy is a greedy strategy, which is expressed by the formula: Π(s)=argmax a∈A Q t (s,a) In the formula, Q t represents time t, argmax a∈A Q t (s,a) means that the picking hand selects action a that maximizes the current Q value at time t, and action a is included in the A set; The random strategy is that the picking hand selects a random action a with a probability of ε each time t , select the action max(s) that maximizes the Q value with a probability of 1-ε t ,a).

[0012] Furthermore, performing collision detection on the initial candidate paths and determining the optimized candidate paths for each picking hand according to the collision detection results includes: Obstacles whose distances from the initial candidate paths are within a preset threshold range are determined based on the tea bud images; Perform intersection detection on the picking hand and the obstacle according to the preset geometric algorithm, and determine whether the picking hand and the obstacle will collide based on the intersection detection result; The initial candidate paths where the picking hand and obstacles will not collide are marked as optimized candidate paths.

[0013] Furthermore, the method of using a genetic algorithm to screen the preferred candidate paths to obtain the optimal picking path includes: Initialize an initial population containing multiple individuals, each of which represents an optimization candidate path: Taking the sum of Q values ​​as the fitness of each individual, the initial population is iteratively selected, crossed, and mutated to obtain the optimal picking path with the best fitness; Among them, the selection operation is used to determine the parent individual, the crossover operation is used to make the parent individual produce new offspring individuals, and the mutation operation is used to randomly change the picking path in the offspring individuals.

[0014] Furthermore, the preprocessing includes image denoising and image size adjustment.

[0015] Furthermore, the position calibration of the multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand includes: Place multiple calibration points with known real positions in the picking scene, use the visual device of the picking robot to obtain an image containing all the calibration points, and extract the measured coordinates of each calibration point; The visual device of the picking robot is calibrated according to the real position and measured coordinates of the calibration points, and the positions of multiple picking hands are verified using the calibrated visual device to obtain the initial coordinate information of each picking hand.

[0016] Furthermore, the preset vision is a YOLOv8 network.

[0017] In the third aspect, the present application also provides a computer-readable storage medium for storing computer-readable programs or instructions. When the program or instructions are executed by a processor, the multi-picking hand path planning control method of the tea bud picking robot described in any one of the above technical solutions can be implemented.

[0018] The beneficial effects of this application are: The present application provides a multi-picking hand path planning control system for a tea bud picking robot. The present application innovatively integrates a preset reinforcement learning algorithm, an optimized genetic algorithm and an obstacle avoidance algorithm to realize the path planning of tea bud picking hands with multiple picking hands. The preset reinforcement learning algorithm provides an initial candidate path for the genetic algorithm, collision detection is performed on the initial candidate path to obtain an optimized candidate path, and then the genetic algorithm iterates the optimized candidate path. This not only improves the preset reinforcement learning algorithm's ability to handle complex situations, but also reduces the genetic algorithm's dependence on fitness functions and global information, greatly improving the robustness and generalization of the control algorithm, while being able to ensure the quality of tea buds and the safety of pickers, providing good technical support for the application and promotion of mechanized picking of tea buds. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A system architecture diagram of an embodiment of a multi-picking hand path planning control system for a tea bud picking robot provided in this application; Figure 2 A flowchart of a multi-picking hand path planning control method for a tea bud picking robot provided in the present application in one embodiment; Figure 3 A logical diagram of the Q-learning algorithm provided in this application in one embodiment. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0021] Before introducing the embodiments of the present application, the relevant terms of the present application are explained first.

[0022] Q-learning algorithm: Q-learning algorithm is a reinforcement learning algorithm that allows the intelligent agent to learn an optimal strategy through interaction with the environment without knowing the model of the environment, that is, what action to take in each state to obtain the maximum cumulative reward. Q-learning algorithm is a classic algorithm proposed by Watkin. This is a model-free learning method that provides intelligent systems with a learning ability to select the optimal action in a Markov environment using the experienced action sequence. Now, the application of Q-learning in fields such as reinforcement learning is becoming more and more mature, and it has been greatly applied in path planning problems, fitness trajectory problems, etc. Related research includes the path planning of the Q-learning algorithm in robotic arms.

[0023] Genetic algorithm: Genetic algorithm is a random global search optimization method that simulates the phenomena of replication, crossover and mutation in natural selection and inheritance. Starting from any initial population, it generates a group of individuals that are more suitable for the environment through random selection, crossover and mutation operations, so that the population evolves to better and better areas in the search space to obtain high-quality solutions to the problem. Specifically, genetic algorithm is a search algorithm used to solve optimization problems. The overall idea of ​​the algorithm is based on the "survival of the fittest" law of Darwin's theory of biological evolution. It introduces concepts such as gene coding, chromosome crossover, gene mutation and natural selection in biology into the process of solving optimization problems. Through continuous "population evolution", it finally obtains the optimal solution to the problem.

[0024] The inventive concept of the present application is introduced below. When the robot's picking hands are used to plan the path for picking tea buds, bionic algorithms are usually applied. Such algorithms are based on biological methods for planning, so they are inevitably subject to biological constraints. For example, the path selection in the initial state of the ant colony algorithm is completely random, and the subsequent path selection will be affected by the selection in the previous step, which is easy to fall into the local optimum, resulting in problems such as long picking paths, slow speed, and low efficiency; and the robustness of the machine learning algorithm is poor. If changes occur during the movement, it needs to be recalculated based on the map, which will further reduce the efficiency of path planning. In addition, the existing technology does not perform reasonable collision detection on the optimized path, which may cause unnecessary damage to the picking hands.

[0025] In response to the above problems, the present application improves the tea picking path planning part, obtains the initial path through a preset reinforcement learning algorithm, uses an obstacle avoidance algorithm for collision avoidance detection to obtain an optimized candidate path, and finally uses the optimized genetic algorithm to iterate the initial path to obtain the optimal picking path for the pickers, so that the picking robot control system can accurately and efficiently control the pickers, and perform fast and effective mechanized tea picking operations while ensuring the safety of the pickers and tea buds.

[0026] Combination Figure 1 As shown, a specific embodiment of the present application discloses a multi-picking hand path planning control system 100 of a tea bud picking robot, comprising a data acquisition module 101, an initialization module 102, a path preliminary planning module 103, an optimization module 104 and a picking hand control module 105 connected in sequence; The data acquisition module 101 is used to acquire the tea bud pictures in the current scene and pre-process the tea bud pictures; Initialization module 102 is used to perform target detection on the preprocessed tea bud images using a preset visual detection model to determine the coordinate information of the picking target; it is also used to calibrate the positions of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; A path preliminary planning module 103 is used to obtain an initial candidate path for each picker based on the initial coordinate information of each picker and the picking target coordinate information based on a preset reinforcement learning algorithm; The optimization module 104 is used to perform collision detection on the initial candidate paths, determine the optimized candidate paths for each picker according to the collision detection results, and screen the preferred candidate paths using a genetic algorithm to determine the optimal picking path for each picker; The picking hand control module 105 is used to send picking hand control instructions to the picking robot according to the optimal picking path, so that the picking robot can complete the picking of tea buds.

[0027] Compared with the prior art, the present application acquires and processes tea bud images through a data acquisition module, determines the picking target coordinate information and the initial coordinate information of each picking hand through an initialization module, obtains the initial candidate path of each picking hand according to the initial coordinate information and the picking target coordinate information of each picking hand based on a preset reinforcement learning algorithm, performs collision detection on the initial candidate path through a path preliminary planning module, determines the optimized candidate path of each picking hand according to the collision detection result, screens the optimized candidate paths using a genetic algorithm through an optimization module to obtain the optimal picking path, and finally sends a picking hand control instruction to the picking robot according to the optimal picking path through a picking hand control module, so that the picking robot completes the picking of tea buds. This application innovatively integrates reinforcement learning algorithm, optimization genetic algorithm and obstacle avoidance algorithm to realize the path planning of tea bud picking hands with multiple picking hands. The preset reinforcement learning algorithm provides the genetic algorithm with initial candidate paths, performs collision detection on the initial candidate paths, and then the genetic algorithm iterates the optimized candidate paths. This not only improves the preset reinforcement learning algorithm's ability to handle complex situations, but also reduces the genetic algorithm's dependence on fitness functions and global information, greatly improves the robustness and generalization of the control algorithm, and can ensure the quality of tea buds and the safety of pickers at the same time, providing good technical support for the application and promotion of mechanized picking of tea buds.

[0028] The embodiment of the present application also provides a method for controlling the paths of multiple picking hands of a tea bud picking robot, which is implemented by using the path planning control system of multiple picking hands of the tea bud picking robot described in the above technical solution. Figure 2 As shown, the method includes: Step S101: obtaining a picture of young tea buds in the current scene, and preprocessing the picture of young tea buds; Step S102: using a preset visual detection model to perform target detection on the preprocessed tea bud image to determine the coordinate information of the picking target; and also used to calibrate the positions of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; Step S103: obtaining an initial candidate path for each picking hand based on the initial coordinate information of each picking hand and the picking target coordinate information based on a preset reinforcement learning algorithm; Step S104: performing collision detection on the initial candidate paths, determining the optimized candidate paths for each picker according to the collision detection results, screening the preferred candidate paths using a genetic algorithm, and determining the optimal picking path for each picker; Step S105: sending a picking hand control instruction to the picking robot according to the optimal picking path, so that the picking robot completes the picking of tea buds.

[0029] As a specific embodiment, in step S101, the picture of young tea leaves in the current scene is acquired by a binocular vision system, and the picture of young tea leaves in the current scene is acquired by two deep learning cameras with the same configuration, and then basic processing is performed on the acquired picture of young tea leaves.

[0030] In a preferred embodiment, in step S101, the preprocessing includes image denoising and image resizing.

[0031] It should be noted that in order to improve image quality and facilitate the subsequent extraction of image feature information, actual image preprocessing operations also include image enhancement, image cropping, image scaling, elimination of deformed vectors, etc.

[0032] In a preferred embodiment, the position calibration of multiple picking hands of the picking robot to obtain initial coordinate information of each picking hand includes: Place multiple calibration points with known real positions in the picking scene, use the visual device of the picking robot to obtain an image containing all the calibration points, and extract the measured coordinates of each calibration point; The visual device of the picking robot is calibrated according to the real position and measured coordinates of the calibration points, and the positions of multiple picking hands are verified using the calibrated visual device to obtain the initial coordinate information of each picking hand.

[0033] It should be noted that in practice, the nine-point calibration method is generally used to locate and calibrate the coordinates of the picking hand.

[0034] In a specific embodiment, a target detection algorithm, such as a target detection algorithm based on deep learning, is used to detect the position of the tea buds in the image and locate their two-dimensional pixel coordinates. According to the two-dimensional pixel coordinates of the tea buds in the image, combined with the internal and external parameters of the camera, three-dimensional reconstruction can be performed to derive the three-dimensional world coordinates of the tea bud picking point and the top. Through geometric calculation or triangulation methods, the spatial distance, angle or other relevant parameters between the tea bud picking point and the top can be calculated based on the three-dimensional world coordinates.

[0035] In a preferred embodiment, in step S102, the preset vision is a YOLOv8 network. The tea buds are positioned by the YOLOv8 network, and the tea buds are marked with rectangles to determine the target location of the tea buds. The YOLO algorithm achieves fast target detection by converting the target detection task into a single network forward propagation, so it has high real-time performance. Compared with other target detection algorithms, YOLO can achieve faster detection speed while maintaining high accuracy. In addition, YOLO can effectively handle the detection of dense targets by using the global receptive field of the convolutional layer and the design of the anchor frame, avoiding the problems of repeated detection or missed detection that may occur in other algorithms, and can meet the actual use requirements of the tea bud picking process.

[0036] In a specific embodiment, the picking target coordinate information is obtained based on the YOLO v8 network, including: rectangular marks obtained through target detection to determine the two-dimensional pixel coordinates of the tea bud picking point and the top of the tea bud.

[0037] In a preferred embodiment, in step S103, the preset reinforcement learning algorithm is a Q-learning algorithm; Based on the preset reinforcement learning algorithm, the initial candidate path is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information, including: The Q function is solved with the initial coordinates of the picking hand as the initial state and the picking target coordinates as the target state, and the initial candidate path is determined based on the execution action and state value of the picking hand; The Q function is expressed by the formula: Q Π (s,a)=Q(s,a)+α[r+γmax a’ Q(s',a')-Q(s,a)] In the formula, Q(s, a) represents the Q value of taking action a in state s, α represents the learning rate, r represents the immediate reward brought by the current action, γ represents the discount factor for future rewards, and max a’ Q(s',a') represents the Q value of taking the optimal action a' in the new state s', Q Π (s,a) represents the Q value obtained by taking action a in state s according to strategy Π; Among them, strategy Π includes real-time strategy and random strategy. The real-time strategy is used to obtain the optimal solution of the current state, and the random strategy is used to obtain the optimal solution of the future state.

[0038] In a preferred embodiment, the real-time strategy is a greedy strategy, which is expressed as follows: Π(s)=argmax a∈A Q t (s,a) In the formula, Q t represents time t, argmax a∈A Q t (s,a) means that the picking hand selects action a that maximizes the current Q value at time t, and action a is included in the A set; The random strategy is that the picking hand selects a random action a with a probability of ε each time t , select the action max(s) that maximizes the Q value with a probability of 1-ε t ,a).

[0039] By dividing the acquisition of Q value into two strategies: immediate strategy and non-immediate strategy (random strategy), we can get the Q-table at each state, and get the initial path according to the transformation of the Q-table. The logic block diagram is as follows Figure 3 shown.

[0040] Execute action a t Then enter s t+1 state, will (s t ,a t ,r t ,s t+1 ) are stored in Data, the data in Data are converted into initial paths of different segments, and then the obtained initial paths are combined and iterated through a genetic algorithm.

[0041] It should be noted that, in addition to the basic Q-learning algorithm, the preset reinforcement learning algorithm can also be a combination of the Q-learning algorithm and other bionic algorithms, such as a combination of the Q-learning algorithm with an ant colony algorithm and a particle swarm algorithm. In some embodiments, it can also be a combination of the Q-learning algorithm and a deep learning method. The above optional methods need to be adaptively selected according to the actual scene requirements and the actual computing power resources of the picking robot.

[0042] In a preferred embodiment, in step S104, performing collision detection on the initial candidate paths and determining the optimized candidate paths for each picking hand according to the collision detection results includes: Obstacles whose distances from the initial candidate paths are within a preset threshold range are determined based on the tea bud images; Perform intersection detection on the picking hand and the obstacle according to the preset geometric algorithm, and determine whether the picking hand and the obstacle will collide based on the intersection detection result; The initial candidate paths where the picking hand and obstacles will not collide are marked as optimized candidate paths.

[0043] As a specific embodiment, for each point or line segment on the candidate path, it is necessary to detect whether they intersect with the collision boundary of any obstacle. For each obstacle, its collision boundary or collision volume needs to be defined. Collision detection can be implemented through mathematical geometry algorithms, such as intersection detection between rays and bounding boxes, intersection detection between line segments and cylinders, etc. If the distance between the picking hand and the obstacle is less than the preset threshold distance, it is judged that a collision will occur. At this time, the optimized candidate path is discarded and the next optimized candidate path is detected.

[0044] In a preferred embodiment, the method of using a genetic algorithm to screen the preferred candidate paths to obtain the optimal picking path includes: Initialize an initial population containing multiple individuals, each of which represents an optimization candidate path: Taking the sum of Q values ​​as the fitness of each individual, the initial population is iteratively selected, crossed, and mutated to obtain the optimal picking path with the best fitness; Among them, the selection operation is used to determine the parent individual, the crossover operation is used to make the parent individual produce new offspring individuals, and the mutation operation is used to randomly change the picking path in the offspring individuals.

[0045] The method of this embodiment obtains the initial candidate path through the Q-learning algorithm, avoids the influence of biological characteristics in the bionics algorithm in the initial stage, and effectively reduces the uncertainty of initial path acquisition. After random collision detection, the candidate path is iterated through the optimized genetic algorithm to avoid unpredictable error information caused by the machine learning algorithm itself from entering the path planning stage and affecting subsequent calculations.

[0046] As a specific embodiment, in step S105, the picking hand control module usually uses PLC (programmable logic controller) to control the mechanical arm of the picking robot. PLC is an electronic device specially used for industrial automation and control. It can receive sensor signals or other inputs, and output control signals to actuators after logical operations, thereby realizing the control of equipment such as mechanical arms. In the application of mechanical arms, PLC is usually connected with sensors, actuators and other devices to form a complete system to realize precise control of parameters such as position, speed, torque and so on of the mechanical arm. In addition to PLC, control instructions can also be generated by logic chips such as embedded chips and single-chip microcomputers.

[0047] This embodiment also provides a computer-readable storage medium, on which a multi-picking hand path planning control program of a tea bud picking robot is stored. When the multi-picking hand path planning control program of the tea bud picking robot is executed by a processor, the steps in the above embodiment can be implemented.

[0048] The multi-picking hand path planning and control system of the tea bud picking robot provided in the present embodiment acquires and processes tea bud images through a data acquisition module, determines picking target coordinate information and the initial coordinate information of each picking hand through an initialization module, obtains an initial candidate path for each picking hand according to the initial coordinate information and the picking target coordinate information of each picking hand based on a preset reinforcement learning algorithm, performs collision detection on the initial candidate paths through a path preliminary planning module, determines an optimized candidate path for each picking hand according to the collision detection results, screens the optimized candidate paths using a genetic algorithm through an optimization module to obtain an optimal picking path, and finally sends a picking hand control instruction to the picking robot according to the optimal picking path through a picking hand control module, so that the picking robot completes the picking of tea buds.

[0049] This application innovatively integrates reinforcement learning algorithm, optimization genetic algorithm and obstacle avoidance algorithm to realize the path planning of tea bud picking hands with multiple pickers. The preset reinforcement learning algorithm provides the genetic algorithm with initial candidate paths, performs collision detection on the initial candidate paths, and then the genetic algorithm iterates the optimized candidate paths. This not only improves the preset reinforcement learning algorithm's ability to handle complex situations, but also reduces the genetic algorithm's dependence on fitness functions and global information, greatly improves the robustness and generalization of the control algorithm, and can ensure the quality of tea buds and the safety of pickers, providing technical support for the application and promotion of mechanized picking of tea buds.

[0050] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0051] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-picking hand path planning control system for a tea bud picking robot, characterized in that: It includes a data acquisition module, an initialization module, a path preliminary planning module, an optimization module and a picking hand control module which are connected in sequence; The data acquisition module is used to acquire the tea bud pictures in the current scene and pre-process the tea bud pictures; An initialization module is used to perform target detection on the preprocessed tea bud images using a preset visual detection model to determine the coordinate information of the picking target; It is also used to calibrate the positions of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; A preliminary path planning module is used to obtain an initial candidate path for each picking hand based on the initial coordinate information of each picking hand and the picking target coordinate information based on a preset reinforcement learning algorithm; the preset reinforcement learning algorithm is a Q-learning algorithm; Based on the preset reinforcement learning algorithm, the initial candidate path is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information, including: The Q function is solved with the initial coordinates of the picking hand as the initial state and the picking target coordinates as the target state, and the initial candidate path is determined based on the execution action and state value of the picking hand; The Q function is expressed by the formula: Q Π (s,a)=Q(s,a)+α[r+γmax a’ Q(s’,a’)-Q(s,a)] In the formula, Q(s, a) represents the Q value of taking action a in state s, α represents the learning rate, r represents the immediate reward brought by the current action, γ represents the discount factor for future rewards, and max a’ Q(s',a') represents the Q value of taking the optimal action a' in the new state s', Q Π (s,a) represents the Q value obtained by taking action a in state s according to strategy Π; Among them, strategy Π includes real-time strategy and random strategy. The real-time strategy is used to obtain the optimal solution of the current state, and the random strategy is used to obtain the optimal solution of the future state. The optimization module is used to perform collision detection on the initial candidate paths, determine the optimized candidate paths for each picker according to the collision detection results, screen the preferred candidate paths using a genetic algorithm, and determine the optimal picking path for each picker; the collision detection on the initial candidate paths and determining the optimized candidate paths for each picker according to the collision detection results include: Obstacles whose distances from the initial candidate paths are within a preset threshold range are determined based on the tea bud images; Perform intersection detection on the picking hand and the obstacle according to the preset geometric algorithm, and determine whether the picking hand and the obstacle will collide based on the intersection detection result; Marking the initial candidate path where the picking hand and the obstacle will not collide as the optimized candidate path; The picking hand control module is used to send picking hand control instructions to the picking robot according to the optimal picking path, so that the picking robot can complete the picking of tea buds.

2. A multi-picking hand path planning control method for a tea bud picking robot, characterized in that: The multi-picking hand path planning control system of the tea bud picking robot according to claim 1 is implemented, comprising: Get the tea buds picture in the current scene, and pre-process the tea buds picture; The preset visual detection model is used to detect the target of the preprocessed tea buds image and determine the coordinate information of the picking target; it is also used to calibrate the position of multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand; Based on a preset reinforcement learning algorithm, according to the initial coordinate information of each picking hand and the picking target coordinate information, an initial candidate path for each picking hand is obtained; the preset reinforcement learning algorithm is a Q-learning algorithm; Based on the preset reinforcement learning algorithm, the initial candidate path is obtained according to the initial coordinate information of each picking hand and the picking target coordinate information, including: The Q function is solved with the initial coordinates of the picking hand as the initial state and the picking target coordinates as the target state, and the initial candidate path is determined based on the execution action and state value of the picking hand; The Q function is expressed by the formula: Q Π (s,a)=Q(s,a)+α[r+γmax a’ Q(s’,a’)-Q(s,a)] In the formula, Q(s, a) represents the Q value of taking action a in state s, α represents the learning rate, r represents the immediate reward brought by the current action, γ represents the discount factor for future rewards, and max a’ Q(s',a') represents the Q value of taking the optimal action a' in the new state s', Q Π (s,a) represents the Q value obtained by taking action a in state s according to strategy Π; Among them, strategy II includes an instant strategy and a random strategy, the instant strategy is used to obtain the optimal solution of the current state, and the random strategy is used to obtain the optimal solution of the future state; performing collision detection on the initial candidate path, determining the optimized candidate path of each picker according to the collision detection result, and screening the preferred candidate path using a genetic algorithm to determine the optimal picking path for each picker; performing collision detection on the initial candidate path, and determining the optimized candidate path of each picker according to the collision detection result, includes: Obstacles whose distances from the initial candidate paths are within a preset threshold range are determined based on the tea bud images; Perform intersection detection on the picking hand and the obstacle according to the preset geometric algorithm, and determine whether the picking hand and the obstacle will collide based on the intersection detection result; Marking the initial candidate path where the picking hand and the obstacle will not collide as the optimized candidate path; According to the optimal picking path, picking hand control instructions are sent to the picking robot, so that the picking robot completes the picking of tea buds.

3. The multi-picking hand path planning control method of the tea bud picking robot according to claim 2 is characterized in that: The real-time strategy is a greedy strategy, which is expressed as follows: Π(s)=argmax a∈A Q t (s,a) In the formula, Q t represents time t, argmax a∈A Q t (s,a) means that the picking hand selects action a that maximizes the current Q value at time t, and action a is included in the A set; The random strategy is that the picking hand selects a random action a with a probability of ε each time t , select the action max(s) that maximizes the Q value with a probability of 1-ε t ,a).

4. The multi-picking hand path planning control method of the tea bud picking robot according to claim 2 is characterized in that: The method of using a genetic algorithm to screen the preferred candidate paths and determine the optimal picking path for each picker includes: Initialize an initial population containing multiple individuals, each of which represents an optimization candidate path: Taking the sum of Q values ​​as the fitness of each individual, the initial population is iteratively selected, crossed, and mutated to obtain the optimal picking path with the best fitness; Among them, the selection operation is used to determine the parent individual, the crossover operation is used to make the parent individual produce new offspring individuals, and the mutation operation is used to randomly change the picking path in the offspring individuals.

5. The multi-picking hand path planning control method of the tea bud picking robot according to claim 2 is characterized in that: The preprocessing includes image denoising and image resizing.

6. The multi-picking hand path planning control method of the tea bud picking robot according to claim 2 is characterized in that: The position calibration of the multiple picking hands of the picking robot to obtain the initial coordinate information of each picking hand includes: Place multiple calibration points with known real positions in the picking scene, use the visual device of the picking robot to obtain an image containing all the calibration points, and extract the measured coordinates of each calibration point; The visual device of the picking robot is calibrated according to the real position and measured coordinates of the calibration points, and the positions of multiple picking hands are verified using the calibrated visual device to obtain the initial coordinate information of each picking hand.

7. The multi-picking hand path planning control method of the tea bud picking robot according to claim 2 is characterized in that: The preset vision is the YOLOv8 network.

8. A computer-readable storage medium, characterized in that: Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the multi-picking hand path planning control method of the tea bud picking robot described in any one of claims 2 to 7.

Images