Intelligent path planning method and device for lunar miniature robot based on knowledge guidance

Abstract

The invention provides an intelligent path planning method and device for a lunar miniature robot based on knowledge guidance, and relates to the technical field of extraterrestrial detection unmanned systems, and the method comprises the steps: obtaining an actual image and a target image which are shot by the lunar miniature robot and are preprocessed; the preprocessing comprises down-sampling processing and semantic segmentation processing; and inputting the actual image and the target image into the reinforcement learning decision model, and outputting to obtain a path plan for enabling the robot to move from the current pose to the pose of shooting the image most similar to the target image. The training process of the reinforcement learning decision model comprises the following steps: performing deep learning pre-training on an initial neural network by using a pre-training data set constructed by priori knowledge to obtain an initial decision model; and based on the initial decision model, performing fine tuning training by adopting a reinforcement learning method to obtain a reinforcement learning decision model. According to the scheme, a human-like local path planning strategy can be learned from visual perception, and a photographing task is intelligently completed.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned extraterrestrial exploration systems, and in particular to a method and device for intelligent path planning of a lunar micro-robot based on knowledge guidance. Background Art

[0002] In exploration missions such as lunar exploration, Mars exploration, and asteroid exploration, autonomous navigation and path planning of extraterrestrial exploration unmanned systems are key links and have important engineering significance. The mobility performance of the probe is closely related to the terrain and soil characteristics. Since it is difficult to accurately simulate the fine terrain of the actual landing area in advance, autonomous perception and path planning are required.

[0003] In related technologies, the positioning performance of a robot is usually affected by many factors, which makes it difficult to achieve accurate positioning of itself. Moreover, due to positioning errors, even if the posture is pre-planned, large posture deviations are likely to occur during the robot's movement to the target position, greatly increasing the difficulty of its path planning.

[0004] Based on this, there is an urgent need for a knowledge-guided intelligent path planning method and device for lunar micro-robots to solve the above technical problems. Summary of the invention

[0005] The embodiment of the present invention provides a method and device for intelligent path planning of a lunar micro-robot based on knowledge guidance, which can dynamically and autonomously plan a local path to reach the most suitable target point with a desired posture to complete a photo-taking task.

[0006] In a first aspect, an embodiment of the present invention provides a knowledge-guided lunar micro-robot path intelligent planning method, comprising:

[0007] Acquire the actual image and the target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing;

[0008] The actual image and the target image are input into a reinforcement learning decision model, and a path planning is output to move the robot from a current position to a position for capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning an initial decision model, and the initial decision model is obtained by deep learning pre-training of an initial neural network using a pre-training data set constructed using prior knowledge, wherein the prior knowledge is that all images contain the target to be captured and the robot cannot return to its original position after performing multiple steps of action.

[0009] In a second aspect, an embodiment of the present invention further provides a knowledge-guided lunar micro-robot path intelligent planning device, comprising:

[0010] An acquisition module is used to acquire the actual image and target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing;

[0011] A planning module is used to input the actual image and the target image into a reinforcement learning decision model, and output a path planning for moving the robot from a current posture to a posture for capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning an initial decision model, and the initial decision model is obtained by deep learning pre-training of an initial neural network using a pre-training data set constructed based on prior knowledge, wherein the prior knowledge is that all images contain a target to be captured and the robot cannot return to its original position after performing multiple steps of action.

[0012] In a third aspect, an embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method described in any embodiment of this specification is implemented.

[0013] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, enables the computer to execute the method described in any embodiment of this specification.

[0014] The embodiment of the present invention provides a method and device for intelligent path planning of a lunar micro-robot based on knowledge guidance. The method first introduces prior knowledge in the sampling stage and generates training samples using the traditional precise positioning path planning method; secondly, the constructed decision model is pre-trained using a deep learning method; then, based on the knowledge guidance information provided by the pre-trained model, the model is fine-tuned and trained using a reinforcement learning method; finally, a decision model is obtained that can intelligently and autonomously complete the photo-taking task based on the target image and the actual image information. In the absence of a complete extraterrestrial environment model and the inability to accurately navigate and position, this method can learn a human-like local path planning strategy from visual perception and intelligently complete the photo-taking task. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0016] Figure 1 It is a flow chart of a method for intelligent path planning of a lunar micro-robot based on knowledge guidance provided by an embodiment of the present invention;

[0017] Figure 2 is a flow chart of a learning method provided by an embodiment of the present invention;

[0018] Figure 3 is a hardware architecture diagram of an electronic device provided by an embodiment of the present invention;

[0019] Figure 4 It is a structural diagram of a knowledge-guided lunar micro-robot path intelligent planning device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0021] As mentioned earlier, the mobility performance of the probe is closely related to the terrain and soil characteristics. Since it is difficult to accurately simulate the fine terrain and topography of the actual landing area in advance, autonomous perception and path planning are required. During extraterrestrial mobile patrols, the robot's positioning performance will usually be affected in some way, making it difficult to achieve accurate positioning.

[0022] Based on this, the concept of the present invention is to dynamically and autonomously plan the local path based on the perception-judgment-decision-execution method, combined with sensor information with limited precision, to reach the most suitable target point with the expected posture, so as to take and send back photos, and intelligently and autonomously complete the photo-taking task.

[0023] The specific implementation of the above concept is described below.

[0024] Please refer to Figure 1 The embodiment of the present invention provides a method for intelligent path planning of a lunar micro-robot based on knowledge guidance, the method comprising:

[0025] Step 100, obtaining the actual image and the target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing;

[0026] Step 102: input the actual image and the target image into a reinforcement learning decision model, and output a path planning for moving the robot from a current position to a position for capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning an initial decision model, and the initial decision model is obtained by deep learning pre-training an initial neural network using a pre-training data set constructed based on prior knowledge, wherein the prior knowledge is that all images contain the target to be captured and the robot cannot return to its original position after performing multiple steps of action.

[0027] In the embodiment of the present invention, prior knowledge is first introduced in the sampling stage, and training samples are generated using the traditional precise positioning path planning method; secondly, the constructed decision model is pre-trained using the deep learning method; then, based on the knowledge guidance information provided by the pre-trained model, the model is fine-tuned and trained using the reinforcement learning method; finally, a decision model is obtained that can intelligently and autonomously complete the photo-taking task based on the target image and the actual image information. In the absence of a complete extraterrestrial environment model and the inability to accurately navigate and position, this method can learn a human-like local path planning strategy from visual perception and intelligently complete the photo-taking task. .

[0028] Described below Figure 1 How the various steps are performed.

[0029] First, with respect to step 100, the actual image and the target image taken by the lunar micro-robot after preprocessing are acquired.

[0030] In the embodiment of the present invention, the lunar microrobot will use the camera it carries to capture RGB images. In order to better simulate the environment of extraterrestrial exploration, the following experimental scene is designed, including: robot, camera, extraterrestrial environment, and lander.

[0031] Specifically including: Robot motion simulation: It can realistically simulate the motion of the robot's parallel multi-link drive mechanism, and realize different motion actions such as forward, backward, left turn, and right turn through the differential rotation control of the left and right drive motors; Camera imaging simulation: The robot only contains a pair of Realsense D455 cameras with a front-to-back symmetrical configuration. The camera has a resolution of 1280×800, a field of view of 90°, and an angle of 20° between the line of sight axis and the horizontal. The camera can obtain image information of the environment in real time during the movement of the robot; Lander geometry simulation: The lander adopts a simplified model, which mainly simulates the geometric shape of the lander. The texture material is mainly based on color block mapping, which is appropriately simplified to improve the computational efficiency of the simulator.

[0032] The tiny robot on the moon acquires images of all targets to be photographed in the above environment.

[0033] Then, for step 102, the actual image and the target image are input into the reinforcement learning decision model, and a path planning is outputted to enable the robot to move from the current posture to the posture of capturing an image most similar to the target image.

[0034] In an embodiment of the present invention, the reinforcement learning decision model is obtained by fine-tuning the initial decision model, and the initial decision model is obtained by deep learning pre-training of the initial neural network using a pre-training data set constructed by prior knowledge, wherein the prior knowledge is that all images contain the target to be photographed and the robot cannot return to the original position after performing multiple steps of action.

[0035] In an embodiment of the present invention, the process of determining pre-training data includes: randomly generating a target posture of the robot, and photographing a target image of the target to be photographed at the target posture; determining whether an initial error between the randomly generated initial posture and the target posture is less than a preset error threshold; if so, regenerating the initial posture until the initial error is greater than the error threshold; based on prior knowledge, utilizing an error-driven action sequence generation and screening mechanism, executing effective actions from the initial posture to the target posture, collecting and processing state data, generating state-action pairs, and finally summarizing all state-action pairs to construct a pre-training data set.

[0036] Specifically, the pre-training data set requires that given a target image, the robot moves to a certain position so that the image taken by the robot camera is similar to the target image. The similarity between the image actually taken by the robot camera and the target image includes the pixel number error of the segmented image, the cosine distance error, and the Euclidean distance error. The pixel number error is used to represent the overall similarity of the target object captured by the two images; the cosine distance error is used to represent the similarity between the shooting angles of the two images; and the Euclidean distance error is used to represent the similarity between the shooting positions of the two images.

[0037] The core idea of ​​generating pre-training samples is to introduce prior knowledge, use the precise posture information of the robot in the simulation environment, and plan an optimal path from the starting point to the end point to meet various constraints such as kinematic constraints. The main prior knowledge includes that the entire decision-making process needs to ensure that the target object being photographed is always within the field of view of the robot camera, and the robot cannot return to its original position after sequentially performing multiple steps.

[0038] In the embodiment of the present invention, the process of simulating the multi-component action of the robot moving from the initial posture to the posture most similar to the target posture according to the prior knowledge and establishing the pre-training data set includes the following steps:

[0039] S1, randomly generate the target posture and initial posture of the robot, so that the initial error (i.e. the error between the current posture of the robot and the target posture) is greater than the threshold. Initialize the robot to the initial posture, use the image taken by the robot at the target position as the target image, and record the initial error;

[0040] S2, repeat the following process until the generated action sequence cannot further reduce the error:

[0041] (1) Randomly generate a set of action sequences containing multiple meta-actions according to prior knowledge;

[0042] (2) obtaining the latest posture of the robot by simulating all the elementary actions in the action sequence and calculating the latest simulation error;

[0043] (3) Determine whether the latest simulation error is not less than the actual error of the previous step. If so, the robot remains stationary, captures the current image and saves it, and terminates the current round;

[0044] (4) If the simulation error is less than the actual error of the previous step (i.e., there is an effective action that can reduce the error), the robot actually executes the first meta-action in the action sequence, updates the robot posture and actual error, and captures and saves the current image;

[0045] (5) preprocessing and concatenating the current image and the target image, using the preprocessed image and the target image as states, and constructing and recording corresponding state-action pairs;

[0046] S3, determine whether the number of state-action pairs currently collected has reached a preset threshold. If so, the entire data collection process ends; if not, a new round of data collection starts again, and steps S1-S2 are repeated. All collected state-action pairs are summarized to construct a pre-training dataset.

[0047] Specifically, if the above initial error exceeds the error threshold, by traversing and simulating the execution of the next n possible action sequences, evaluate whether this set of action sequences can make the current posture closer to the target posture. If so, then the first meta-action in the action sequence is actually executed. After each action is performed, the robot records the RGB image taken by the camera it carries, and then downsamples and semantically segments the image taken by the current camera to obtain a semantic segmentation image, which effectively identifies each key part. The spliced ​​image of the current semantic segmentation image and the target semantic segmentation image and the corresponding action actually performed by the robot together constitute a state-action pair in the training data set. Then, in the new state, repeat the action sequence generation and error judgment, and repeat this cycle until the error can no longer be reduced (the image taken by the robot camera is similar to the target image), such as Figure 2 shown.

[0048] For example, randomly generate the target pose T g , and obtain the image f taken by the camera at the target posture g , f g The robot’s initial pose is randomly given and its difference from the target pose T is calculated. g If the error between them is less than the preset threshold ε, the initial posture of the robot is re-given until the condition is met. The initial posture T1 and the initial error E1 are recorded, and the image f1 taken at the initial posture T1 is saved.

[0049] Furthermore, a set of action sequences consisting of n meta-actions is randomly generated. Get the robot's position T after executing the action sequence i+1 , calculate the error E between the target pose and the target pose i+1 If E i+1 ≥E i , the robot remains stationary and obtains the image f taken at the current position i+1 , and save and record this set of pre-training samples: Input [f g ,f i ], the output is -1, and the action is -1, which means the robot is stationary; if E i+1 <E i , the robot performs the first meta-action in the action sequence Get the image f taken at the current position i+1 , record this set of pre-training samples: input [f g ,f i ], the output is

[0050] The above method is used for multiple rounds of processing to finally obtain a pre-training data set.

[0051] It is worth mentioning that, considering the requirements of both network lightweight and performance, the neural network structure is designed based on ResNet18. The input of the network consists of the target image after semantic segmentation and the current image after semantic segmentation. The position change of the robot relative to the target is represented by a meta-action as the output of the network. Considering that in real physical environments, robots often have difficulty in turning in place due to slipping, a set of meta-actions consisting of 7 meta-actions is designed. The numbers represent forward movement, backward movement, left forward movement, right forward movement, left backward movement, right backward movement and stillness respectively, for example, it can be forward 20cm, backward 20cm, forward 20cm+turn left 10°, forward 20cm+turn right 10°, backward 20cm+turn left 10°, backward 20cm+turn right 10°, stillness and taking picture to save.

[0052] In an embodiment of the present invention, after obtaining the pre-training data set, the initial decision model is trained in the following manner: the pre-training data set is input into the initial neural network after parameter optimization using the cross entropy loss function and the Adam optimization algorithm, and the image-based action decision knowledge is learned to obtain the initial decision model.

[0053] The initial decision model trained in this way can be regarded as a kind of knowledge, that is, based on the pre-training data, finding out how to move from the current image to the target image can narrow the posture difference. This rule can be regarded as a kind of knowledge.

[0054] Furthermore, the decision model is trained in the following manner: the parameters of the initial decision model are used as the initial parameters of the decision model, and training is performed in a reinforcement learning environment; when the initial decision model chooses to stop moving, the number of steps of the executed action is recorded, and the cumulative reward of this trajectory is calculated according to the preset reward function. Among them, the reward function includes an immediate reward and a terminal reward: the immediate reward is used to punish invalid or inefficient actions; the terminal reward is used to evaluate the action accuracy, task difficulty, and the error between the actual strategy execution and the estimate; according to the calculation result of the reward function, the parameters of the initial decision model are updated, and the reward curve is drawn; when the reward curve meets the preset convergence condition, the training is terminated and the reinforcement learning decision model is obtained.

[0055] Specifically, the “knowledge” learned by the initial decision model is used to guide the fine-tuning process and improve the model performance. g The image segmented from the current captured image is used as the network input, and the output is the meta-action. When the policy network outputs the "save" action (-1), the program ends, the number of steps of the action is counted, and the image similarity λ is calculated. The reward function R contains the immediate reward R I and terminal reward R T Two parts. The purpose of setting the process reward is to avoid invalid actions, so that after executing a set of meta-action combinations, the player returns to the position before the action was executed. If such a combination occurs, such as the actions are "forward-backward", "left front + backward", "right front + backward", "left back + forward", "right back + forward", then the immediate reward is -1, otherwise the immediate reward is 0.

[0056] In the embodiment of the present invention, the terminal reward is calculated by the following formula:

[0057]

[0058] Where Δs is the position error at the end; Δα is the posture error at the end; step is the number of meta-actions executed; and γ is the preset number of meta-actions executed.

[0059] e in the terminal reward function -Δs cosΔα evaluates the action accuracy, Δs0 / (1+Δα0) evaluates the difficulty of the task, and min(step,γ) / max(step,γ) evaluates the error between the actual strategy execution and the estimated one.

[0060] In summary, the network trained by the method provided in this embodiment can accurately identify various feature parts and make efficient and autonomous decisions to complete the task of autonomous photography.

[0061] like Figure 3 , Figure 4 As shown, the embodiment of the present invention provides a knowledge-guided lunar micro-robot path intelligent planning device. The device embodiment can be implemented by software, or by hardware or a combination of software and hardware. From the hardware level, Figure 3 As shown in FIG. 1 , a hardware architecture diagram of an electronic device in which a knowledge-guided lunar micro-robot path intelligent planning device is provided in an embodiment of the present invention is provided. Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, the electronic device in the embodiment may also include other hardware, such as a forwarding chip responsible for processing messages, etc. Taking software implementation as an example, Figure 4 As shown, as a device in a logical sense, the CPU of the electronic device in which it is located reads the corresponding computer program in the non-volatile memory into the memory and runs it. This embodiment provides a knowledge-guided lunar micro-robot path intelligent planning device, including:

[0062] The acquisition module 400 is used to acquire the actual image and the target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing;

[0063] The planning module 402 is used to input the actual image and the target image into the reinforcement learning decision model, and output a path planning for moving the robot from the current posture to the posture of capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning the initial decision model, and the initial decision model is obtained by deep learning pre-training of the initial neural network through a pre-training data set constructed by prior knowledge, and the prior knowledge is that all images contain the target to be captured and the robot cannot return to the original position after performing multiple steps.

[0064] In an embodiment of the present invention, the pre-training data is determined in the following manner: randomly generating a target posture of the robot, and photographing the target image of the target to be photographed at the target posture; determining whether the initial error between the randomly generated initial posture and the target posture is less than a preset error threshold; if so, regenerating the initial posture until the initial error is greater than the error threshold; based on prior knowledge, utilizing an error-driven action sequence generation and screening mechanism, executing effective actions from the initial posture to the target posture, collecting and processing state data, generating state-action pairs, and finally summarizing all state-action pairs to construct a pre-training data set.

[0065] In an embodiment of the present invention, multiple sets of meta-actions of a robot moving from the initial posture to the posture most similar to the target posture are simulated according to prior knowledge to establish the pre-training data set, including: S1, randomly generating an initial posture in which the error between the robot posture and the target posture is greater than a threshold; initializing the robot to the initial posture, using an image taken by the robot at the target position as the target image, and recording the initial error; S2, repeating the following process until the generated action sequence cannot further reduce the error: (1) randomly generating a set of action sequences including multiple meta-actions according to prior knowledge; (2) obtaining the latest posture of the robot by simulating and executing all the meta-actions in the action sequence, and calculating the latest simulation error; (3) judging whether the latest simulation error is not less than If the simulation error is less than the actual error of the previous step, the robot remains still, captures and saves the current image, and terminates the current round; (4) If the simulation error is less than the actual error of the previous step, the robot executes the first meta-action in the action sequence, updates the robot posture and actual error, captures and saves the current image; (5) The current image and the target image are preprocessed and spliced, and the corresponding state-action pairs are constructed and recorded as states; S3, determines whether the number of currently collected state-action pairs reaches a preset threshold. If the threshold is reached, the entire data collection process is terminated; if the threshold is not reached, a new round of data collection is restarted, and steps S1-S2 are repeated to summarize all collected state-action pairs and construct a pre-training data set.

[0066] In the embodiment of the present invention, the elementary actions include: forward action, backward action, left forward action, right forward action, left backward action, right backward action and standing still.

[0067] In an embodiment of the present invention, the initial decision model is obtained by training in the following manner: the pre-training data set is input into an initial neural network after parameter optimization using a cross entropy loss function and an Adam optimization algorithm, image-based action decision knowledge is learned, and an initial decision model is obtained.

[0068] In an embodiment of the present invention, the reinforcement learning decision model is obtained by training in the following manner: the parameters of the initial decision model are used as the initial parameters of the decision model, and training is performed in a reinforcement learning environment; when the initial decision model chooses to stop moving, the number of steps of the executed action is recorded, and the cumulative reward of this trajectory is calculated according to a preset reward function. Among them, the reward function includes an immediate reward and a terminal reward: the immediate reward is used to punish invalid or inefficient actions; the terminal reward is used to evaluate the accuracy of the action, the difficulty of the task, and the error between the actual strategy execution and the estimate; according to the calculation result of the reward function, the parameters of the initial decision model are updated, and the reward curve is drawn; when the reward curve meets the preset convergence condition, the training is terminated and the reinforcement learning decision model is obtained.

[0069] In the embodiment of the present invention, the terminal reward is calculated by the following formula:

[0070]

[0071] Where Δs is the position error at the end; Δα is the posture error at the end; step is the number of meta-actions executed; and γ is the preset number of meta-actions executed.

[0072] It is to be understood that the structure illustrated in the embodiment of the present invention does not constitute a specific limitation on a path intelligent planning device for a lunar micro-robot based on knowledge guidance. In other embodiments of the present invention, a path intelligent planning device for a lunar micro-robot based on knowledge guidance may include more or fewer components than shown in the figure, or combine some components, or split some components, or arrange the components differently. The components shown in the figure may be implemented in hardware, software, or a combination of software and hardware.

[0073] The information interaction, execution process and other contents between the modules in the above-mentioned device are based on the same concept as the embodiment of the method of the present invention. For the specific contents, please refer to the description in the embodiment of the method of the present invention, and no further description is given here.

[0074] An embodiment of the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, a knowledge-guided lunar micro-robot path intelligent planning method in any embodiment of the present invention is implemented.

[0075] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the processor executes a knowledge-guided lunar micro-robot path intelligent planning method in any embodiment of the present invention.

[0076] Specifically, a system or device equipped with a storage medium can be provided, on which software program code that implements the functions of any of the above-mentioned embodiments is stored, and a computer (or CPU or MPU) of the system or device can be enabled to read and execute the program code stored in the storage medium.

[0077] In this case, the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, and thus the program code and the storage medium storing the program code constitute a part of the present invention.

[0078] The storage medium embodiments for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program code can be downloaded from a server computer by a communication network.

[0079] In addition, it should be clear that the functions of any of the above embodiments can be implemented not only by executing the program code read by the computer, but also by enabling an operating system operating on the computer to complete part or all of the actual operations based on instructions from the program code.

[0080] In addition, it can be understood that the program code read from the storage medium is written to a memory provided in an expansion board inserted into the computer or to a memory provided in an expansion module connected to the computer, and then based on the instructions of the program code, a CPU installed on the expansion board or expansion module is enabled to perform part or all of the actual operations, thereby realizing the functions of any of the above-mentioned embodiments.

[0081] It should be noted that, in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the presence of other identical factors in the process, method, article or device including the elements.

[0082] A person of ordinary skill in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: ROM, RAM, magnetic disk or optical disk, etc., various media that can store program codes.

[0083] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge-guided intelligent path planning method for lunar microrobots, characterized in that: include: Acquire the actual image and the target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing; The actual image and the target image are input into a reinforcement learning decision model, and a path planning is output to move the robot from a current position to a position for capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning an initial decision model, and the initial decision model is obtained by deep learning pre-training of an initial neural network using a pre-training data set constructed using prior knowledge, wherein the prior knowledge is that all images contain the target to be captured and the robot cannot return to its original position after performing multiple steps of action.

2. The method according to claim 1, characterized in that The pre-training data set is determined in the following manner: Randomly generate a target posture of the robot, and photograph the target image of the target to be photographed under the target posture; Determine whether an initial error between the randomly generated initial posture and the target posture is less than a preset error threshold; if so, regenerate the initial posture until the initial error is greater than the error threshold; Based on prior knowledge, the error-driven action sequence generation and screening mechanism is used to perform effective actions from the initial posture to the target posture, collect and process state data, generate state-action pairs, and finally summarize all state-action pairs to construct a pre-training dataset.

3. The method according to claim 2, characterized in that The state data is collected and processed to generate state-action pairs, and finally all state-action pairs are aggregated to construct a pre-training dataset, including: S1, randomly generate an initial posture whose error between the robot posture and the target posture is greater than a threshold; initialize the robot to the initial posture, use the image taken by the robot at the target position as the target image, and record the initial error; S2, repeat the following process until the generated action sequence cannot further reduce the error: (1) Randomly generate a set of action sequences containing multiple meta-actions according to prior knowledge; (2) obtaining the latest posture of the robot by simulating all the elementary actions in the action sequence and calculating the latest simulation error; (3) Determine whether the latest simulation error is not less than the actual error of the previous step. If so, the robot remains stationary, captures the current image and saves it, and terminates the current round; (4) If the simulation error is less than the actual error of the previous step, the first meta-action in the action sequence is executed, the robot posture and actual error are updated, and the current image is captured and saved; (5) preprocessing and concatenating the current image and the target image, using the preprocessed image and the target image as states, and constructing and recording corresponding state-action pairs; S3, determine whether the number of state-action pairs currently collected reaches the preset threshold. If so, the entire data collection process ends; if not, a new round of data collection starts again, and steps S1-S2 are repeated to summarize all collected state-action pairs and construct a pre-training data set.

4. The method according to claim 3, characterized in that The elementary actions include: forward action, backward action, left forward action, right forward action, left backward action, right backward action and standing still.

5. The method according to claim 1, characterized in that The initial decision model is trained in the following way: The pre-trained data set is input into the initial neural network after parameter optimization using the cross entropy loss function and the Adam optimization algorithm to learn image-based action decision knowledge and obtain an initial decision model.

6. The method according to claim 5, characterized in that The reinforcement learning decision model is trained in the following way: The parameters of the initial decision model are used as the initial parameters of the decision model and trained in a reinforcement learning environment; When the initial decision model chooses to stop moving, the number of steps of the executed action is recorded, and the cumulative reward of this trajectory is calculated according to the preset reward function; wherein the reward function includes immediate reward and terminal reward: the immediate reward is used to punish invalid or inefficient actions; the terminal reward is used to evaluate the action accuracy, task difficulty and the error between the actual strategy execution and the estimation; According to the calculation result of the reward function, the parameters of the initial decision model are updated, and a reward curve is drawn; when the reward curve meets the preset convergence condition, the training is terminated and the reinforcement learning decision model is obtained.

7. The method according to claim 6, characterized in that The terminal reward is calculated by the following formula: Where Δs is the position error at the end; Δα is the posture error at the end; step is the number of meta-actions executed; and γ is the preset number of meta-actions executed.

8. A knowledge-guided intelligent path planning device for lunar microrobots, characterized in that: include: An acquisition module is used to acquire the actual image and target image taken by the lunar micro-robot after preprocessing; the preprocessing includes downsampling processing and semantic segmentation processing; A planning module is used to input the actual image and the target image into a reinforcement learning decision model, and output a path planning for moving the robot from a current posture to a posture for capturing an image most similar to the target image; wherein the reinforcement learning decision model is obtained by fine-tuning an initial decision model, and the initial decision model is obtained by deep learning pre-training of an initial neural network using a pre-training data set constructed based on prior knowledge, wherein the prior knowledge is that all images contain a target to be captured and the robot cannot return to its original position after performing multiple steps of action.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

Images