A visual navigation method for mobile robots based on voxel spatial attention mechanism

Abstract

The present invention proposes a mobile robot visual navigation method based on voxel space attention mechanism, including: obtaining RGB image data and Depth depth data based on visual sensors, designing voxel grids, converting two-dimensional pixels into three-dimensional point cloud data, and projecting them into three-dimensional voxel space; designing voxel space attention modules, processing voxel grids, calculating the attention weight of each voxel to select the area of ​​interest and determine the navigation target position; designing an attention weight decoding module, decoding the final output of the attention model into the action to be performed by the mobile robot at the next moment, and realizing the navigation motion planning of the robot; and optimizing parameters based on a deep reinforcement learning framework to realize the robot visual navigation decision. The present invention realizes motion planning for visual navigation in indoor scenes. Compared with traditional methods, the voxel space attention mechanism is used to help guide the robot to find the target object, which reduces the training time and improves the generalization of navigation.

Description

Technical Field

[0001] The present invention belongs to the field of visual navigation motion planning technology, and in particular to a mobile robot visual navigation method based on voxel space attention mechanism. Background Art

[0002] Visual navigation is a technology that uses cameras as the main sensor to enable robots to navigate autonomously in an environment. The robot analyzes the real-time images captured by the camera, identifies and understands its surrounding environment, and plans a path to the target. With the rapid development of science and technology, the demand for intelligent robots in daily life and industrial applications is increasing. Especially in complex indoor environments, how to achieve efficient and safe navigation has become an important research topic. Visual navigation, as a technology that uses visual information to guide robots to navigate autonomously, has attracted widespread attention due to its potential in environmental understanding and decision-making.

[0003] However, although visual navigation technology has made significant progress, it still faces many challenges in practical applications, especially in complex indoor environments. Visual navigation systems usually need to process a large amount of high-dimensional data, and extracting effective information from these data is a major challenge. How to effectively identify and distinguish navigation-related features from complex visual inputs is the key to improving navigation efficiency and accuracy. In addition, processing these large amounts of data requires efficient algorithms and powerful computing power, which is particularly challenging on resource-constrained robotic platforms. Most existing visual navigation models are trained in specific datasets and predefined environments, but their performance often drops significantly when applied to new environments. How to design a visual navigation model that can effectively generalize to unseen environments is a hot topic in current research. This involves aspects such as model training methods, data diversity, and algorithm robustness.

[0004] Voxel grids, as a 3D spatial representation technology, offer several significant advantages in advanced visual navigation systems, especially for complex environments that require highly accurate and reliable navigation capabilities. Voxel grids discretize space into tiny cubic units, which can capture and store detailed structural information in the environment while maintaining high resolution. Voxel grids provide an inherent 3D data structure for depth perception, which enables visual navigation systems to perform spatial analysis more accurately. Summary of the invention

[0005] In response to the problems existing in the above-mentioned prior art, the present invention proposes a mobile robot visual navigation method based on voxel spatial attention mechanism, which understands and captures the structural information of the environment through voxel grids to store a large amount of spatial information, including the shape, size, position of objects and the relative relationship between spaces, thereby providing rich environmental information for machine vision; in addition, the voxel spatial attention mechanism is used to help guide the robot to discover the target object, thereby reducing training time and improving navigation generalization.

[0006] In order to achieve the above technical objectives, the present invention provides the following technical solutions:

[0007] A mobile robot visual navigation method based on voxel space attention mechanism, which specifically includes:

[0008] S1. For the mobile robot in the scene, at any time t during the navigation process, obtain RGB image data and Depth data based on its visual sensor; design a voxel conversion algorithm, based on the camera's internal and external parameters, convert the RGB image data and Depth data into three-dimensional point cloud data, and project it into a three-dimensional voxel space to obtain the voxel matrix of the mobile robot at any time t with three-dimensional point cloud information Similarly, for a given navigation task target’s RGBD data, the navigation target voxel matrix is ​​obtained in the same way: ;

[0009] S2. Design a voxel spatial attention module to monitor the voxel matrix of the mobile robot at any time t. and the navigation target voxel matrix Processing is performed to calculate the attention weight of each voxel in the voxel grid A ; Based on attention weight A Select the region of interest and get the predicted position of the navigation target in the voxel grid;

[0010] S3. Design an attention weight decoding module to decode the final output of the voxel space attention module into the action to be performed by the mobile robot in the next time step, so as to realize the navigation motion planning of the robot;

[0011] S4. Based on the deep reinforcement learning framework, set the reward function and loss function, optimize the parameters of the two modules designed in steps S2 and S3, and realize the optimal navigation motion planning of the robot in the scene.

[0012] Furthermore, step S1 specifically includes:

[0013] S11, combining the RGB image data and the Depth data at time t into an RGB-D image of size (128, 128, 4), which carries two-dimensional pixel information and depth information;

[0014] S12. By combining the intrinsic parameters of the visual sensor and the acquired pixel and depth information, a three-dimensional point cloud is constructed using the mapping relationship between two-dimensional and three-dimensional space; in the image coordinate system, the three-dimensional coordinates corresponding to a pixel point (u, v) and its depth value d in the camera coordinate system are recorded as (X, Y, Z); according to the camera imaging principle, the conversion relationship between the two-dimensional pixel coordinates and the three-dimensional point cloud data is obtained as follows:

[0015] ;

[0016] ;

[0017] ;

[0018] Where s represents the unit conversion scaling factor between pixels and millimeters, K is the intrinsic parameter matrix of the depth camera, and is the image focal length, and is the center of the depth image; parameter and The units of are all pixels, and the units of the 3D point cloud coordinates (X, Y, Z) to be solved are millimeters;

[0019] S13, according to the conversion relationship in step S12, solve the three-dimensional point cloud coordinates to achieve the mapping between the two-dimensional pixel coordinates, that is, RGB-D data, and the three-dimensional point cloud data. The formula is expressed as:

[0020] ;

[0021] S14, voxelize the 3D point cloud data obtained in step S13 with a resolution of 32*32*32; first calculate the minimum bounding box of the 3D point cloud data; then divide the minimum bounding box into voxel spaces with a resolution of L; traverse each voxel, if a 3D space point exists, store 1, otherwise store 0; finally obtain a voxel matrix , which is the input of the voxel space attention module at time t.

[0022] Furthermore, the voxel spatial attention module designed in step S2 is specifically: a convolutional neural network structure for three-dimensional voxel data processing, which includes:

[0023] An input preprocessing module is used to receive three-dimensional input data with a predetermined number of channels; the input preprocessing module includes three three-dimensional convolutional layers, and the three-dimensional convolutional layers use multi-size convolutional kernels to perform feature extraction, respectively extracting and The voxel features are obtained and ;

[0024] Voxel cross attention machine module, and is the input and outputs the attention weight A , and use it to update the voxel features ;

[0025] Specifically include: using a three-dimensional convolution layer to process and , to obtain the query matrix Q and the key matrix K, and then process it with a three-dimensional convolution layer Get the value matrix V; and set the learnable hyperparameters , then we get the attention weight A The formula is expressed as:

[0026] ;

[0027] in, is the activation function;

[0028] Finally, the updated voxel features are output .

[0029] Further, in step S2, the attention weight A Select the region of interest and get the predicted position of the navigation target in the voxel grid as follows:

[0030] According to the calculated attention weight A Find the area where the target object is located: First define the sum operation, A Each voxel in The dimensional features are summed, and then based on the summation results of each voxel feature, the maximum value and its corresponding voxel index position are found , which is the predicted position of the navigation target; the mathematical expression of this process is: ,in is the attention weight A Median component; select voxel index position Central location Construct the subsequent loss function.

[0031] Furthermore, step S3 is specifically as follows:

[0032] Design an attention weight decoding module to decode the final output of the voxel space attention module into the action to be performed by the mobile robot in the next time step, so as to realize the navigation motion planning of the robot; the attention weight decoding module is composed of 3 3D convolutional layers and 2 fully connected layers to update the voxel features As input, the probability values ​​of the robot's five discrete navigation actions are output, and the action with the largest probability value is selected as the robot's navigation action at the next moment; the five discrete actions include: walking forward, moving backward, turning left, turning right, and stopping.

[0033] Furthermore, step S4 specifically includes:

[0034] S41. Set the reward function, record the reward for each moment as -β, set the robot to perform only one action at each moment, and the reward for successful navigation is ; The criterion for whether the target position is reached is: when the instance of the target position is less than 1 meter away from the robot's camera and within the robot's field of view, and the attention weight decoding module outputs a stop action, the robot is considered to have reached the target position and successfully completed the navigation task; the reward function of the robot's navigation at each moment is expressed by the formula:

[0035] ;

[0036] S42. Use the Smooth L1 loss function to compare the predicted value and the true value of the target object’s area location, and add the attention weight A of Regularize the loss to prevent overfitting; the formula of the loss function is expressed as:

[0037] ;

[0038] in, is the difference between the predicted value and the true value of the target position, that is, , , Represent the predicted and true positions of the target in the voxel grid, respectively;

[0039] Regularization loss:

[0040] ;

[0041] in, yes The regularized weight coefficient is a hyperparameter; is the attention weight A Medium weight, is the index;

[0042] S43. At the same time, the traditional reinforcement learning model Asynchronous Advantage Actor-Critic is introduced, that is, the loss function of A3C , together with the Smooth L1 loss and the regularization loss, constitute the total loss function of the mobile robot visual navigation model based on the voxel space attention mechanism:

[0043] ;

[0044] in, It is a hyperparameter that represents the importance of each item in the loss function;

[0045] S44. For the total loss function ,Based on the stochastic gradient descent algorithm, the parameters of the voxel space attention module and the attention weight decoding module are iteratively optimized to realize the robot visual navigation decision.

[0046] More specifically, step S44 is as follows:

[0047] During the visual navigation simulation process, the target position is set multiple times for navigation tasks, and the reward function and loss function of each navigation task are recorded; with the minimization of the loss function as the direction, the parameters of the voxel space attention module and the attention weight decoding module are optimized until the overall reward function curve of all tasks converges, completing the training process and obtaining the optimized voxel space attention module and attention weight decoding module; the two optimized modules are deployed on the mobile robot to realize the robot's visual navigation decision.

[0048] Based on the above technical solution, the present invention has the following beneficial effects:

[0049] A mobile robot visual navigation method based on voxel space attention mechanism proposed in the present invention can realize motion planning for visual navigation in complex indoor environment scenes; compared with the traditional motion planning method for visual navigation based on deep reinforcement learning, this method discretizes the space into small cubic units through voxel grids, which can capture and store structural information in the environment in detail while maintaining a high resolution; at the same time, combined with the attention mechanism, it helps guide the robot to discover the target object, thereby improving the training efficiency; under the framework of deep learning, combined with the reward function and the loss function, the training results are further optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A flowchart of a mobile robot visual navigation method based on voxel spatial attention mechanism proposed by the present invention;

[0051] Figure 2 A schematic diagram of a currently observed scene obtained in the present invention;

[0052] Figure 3 This is a network structure diagram of the voxel space attention module in the present invention;

[0053] Figure 4 This is the network structure diagram of the attention weight decoding module in the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0055] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0056] like Figure 1 As shown, the present invention proposes a mobile robot visual navigation method based on voxel space attention mechanism, which specifically includes:

[0057] S1. For the mobile robot in the scene, at any time t during the navigation process, obtain RGB image data and Depth data based on its visual sensor; design a voxel conversion algorithm, based on the camera's internal and external parameters, convert the RGB image data and Depth data into three-dimensional point cloud data, and project it into a three-dimensional voxel space to obtain the voxel matrix of the mobile robot at any time t with three-dimensional point cloud information Similarly, for a given navigation task target’s RGBD data, the navigation target voxel matrix is ​​obtained in the same way: ;

[0058] As a preferred implementation, step S1 specifically includes:

[0059] S11, the RGB image data and Depth data at time t are combined into an RGB-D image of size (128, 128, 4), which carries two-dimensional pixel information and depth information; the RGB image data and Depth data are respectively as follows Figure 2 As shown in (a) and (b);

[0060] S12. By combining the intrinsic parameters of the visual sensor and the acquired pixel and depth information, a three-dimensional point cloud is constructed using the mapping relationship between two-dimensional and three-dimensional space; in the image coordinate system, the three-dimensional coordinates corresponding to a pixel point (u, v) and its depth value d in the camera coordinate system are recorded as (X, Y, Z); according to the camera imaging principle, the conversion relationship between the two-dimensional pixel coordinates and the three-dimensional point cloud data is obtained as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] Where s represents the unit conversion scaling factor between pixels and millimeters, K is the intrinsic parameter matrix of the depth camera, and is the image focal length, and is the center of the depth image; parameter and The units of are all pixels, and the units of the 3D point cloud coordinates (X, Y, Z) to be solved are millimeters;

[0065] S13, according to the conversion relationship in step S12, solve the three-dimensional point cloud coordinates to achieve two-dimensional pixel coordinates, i.e., RGB-D data, and three-dimensional point cloud data (such as Figure 2 The mapping between (as shown in (c) in the figure) is expressed as:

[0066] ;

[0067] S14, voxelize the three-dimensional point cloud data obtained in step S13 (such as Figure 2 As shown in (d) in the figure), first calculate the minimum bounding box of the 3D point cloud data; then divide the minimum bounding box into voxel spaces with a resolution of L; traverse each voxel, if a 3D space point exists, store 1, otherwise store 0; finally obtain a voxel matrix , which is the input of the voxel space attention module at time t.

[0068] S2. Design a voxel spatial attention module to process the voxel grid and calculate the attention weight of each voxel in the voxel grid. A ; Based on attention weight A Select the region of interest and find the predicted location of the navigation target in the voxel grid;

[0069] As a preferred implementation, the voxel spatial attention module designed in step S2 is specifically: a convolutional neural network structure for three-dimensional voxel data processing, such as Figure 3 As shown, it includes:

[0070] An input preprocessing module is used to receive three-dimensional input data with a predetermined number of channels; the input preprocessing module includes three three-dimensional convolutional layers, and the three-dimensional convolutional layers use multi-size convolutional kernels to perform feature extraction, respectively extracting and The voxel features are obtained and ;

[0071] Voxel cross attention machine module, and is the input and outputs the attention weight A , and use it to update the voxel features ;

[0072] Specifically include: using a three-dimensional convolution layer to process and , to obtain the query matrix Q and the key matrix K, and then process it with a three-dimensional convolution layer Get the value matrix V; and set the learnable hyperparameters , then we get the attention weight A The formula is expressed as:

[0073] ;

[0074] in, is the activation function;

[0075] Finally, the updated voxel features are output .

[0076] After calculating the self-attention weights through the above-mentioned voxel space attention module, based on the attention weights A Select the region of interest and get the predicted position of the navigation target in the voxel grid as follows:

[0077] First, define the sum operation. A Each voxel in The dimensional features are summed, and then based on the summation results of each voxel feature, the maximum value and its corresponding voxel index position are found , which is the predicted position of the navigation target; the mathematical expression of this process is: ,in is the attention weight A The present invention selects the voxel index position Central location Construct the subsequent loss function.

[0078] S3. Design an attention weight decoding module to decode the final output of the voxel space attention module into the action to be performed by the mobile robot in the next time step, so as to realize the navigation motion planning of the robot;

[0079] As a preferred implementation, step S3 is specifically as follows:

[0080] Design the attention weight decoding module, such as Figure 4As shown in Figure 1, the final output of the voxel spatial attention module is decoded into the action to be performed by the mobile robot in the next time step to realize the navigation motion planning of the robot. The attention weight decoding module is composed of 3 3D convolutional layers and 2 fully connected layers to update the voxel features. As input, output the probability value of the robot's navigation of 5 discrete actions , select the action with the largest probability value as the navigation action of the robot at the next moment; the five discrete actions include: walking forward 0.25 meters, moving backward 0.25 meters, turning left 90 degrees, turning right 90 degrees, and stopping (that is, reaching the target position).

[0081] S4. Based on the deep reinforcement learning framework, the reward function and loss function are set, and the parameters of the two modules designed in steps S2 and S3 are optimized to achieve the optimal navigation motion planning of the robot in the scene;

[0082] As a preferred implementation, step S4 specifically includes:

[0083] S41. Set the reward function, record the reward for each moment as -β, set the robot to perform only one action at each moment, and the reward for successful navigation is ; The criterion for whether the target position is reached is: when the instance of the target position is less than 1 meter away from the robot's camera and within the robot's field of view, and the attention weight decoding module outputs a stop action, the robot is considered to have reached the target position and successfully completed the navigation task; the reward function of the robot's navigation at each moment is expressed by the formula:

[0084] ;

[0085] In this embodiment, we want the robot to spend as little time as possible to complete the navigation task, so the total reward for a successful navigation task is ;

[0086] S42. Use the Smooth L1 loss function to compare the predicted value and the true value of the target object’s area location, and add the attention weight A of Regularize the loss to prevent overfitting; the formula of the loss function is expressed as:

[0087] Smooth L1 loss:

[0088] ;

[0089] in, is the difference between the predicted value and the true value of the target position, that is, , , Represent the predicted and true positions of the target in the voxel grid, respectively;

[0090] Regularization loss:

[0091] ;

[0092] in, yes The regularized weight coefficient is a hyperparameter; is the attention weight A Medium weight, is the index;

[0093] S43. At the same time, the traditional reinforcement learning model Asynchronous Advantage Actor-Critic is introduced, that is, the loss function of A3C , together with the Smooth L1 loss and the regularization loss, constitute the total loss function of the mobile robot visual navigation model based on the voxel space attention mechanism:

[0094] ;

[0095] in, It is a hyperparameter that represents the importance of each item in the loss function;

[0096] S44. For the total loss function , based on the stochastic gradient descent algorithm, the parameters of the voxel space attention module and the attention weight decoding module are iteratively optimized to realize the robot's visual navigation decision; specifically: in the visual navigation simulation process, the target position is set multiple times for navigation tasks, and the reward function and loss function of each navigation task are recorded; with the minimization of the loss function as the direction, the parameters of the voxel space attention module and the attention weight decoding module are optimized until the overall reward function curve of all tasks converges, the training process is completed, and the optimized voxel space attention module and attention weight decoding module are obtained; the two optimized modules are deployed on the mobile robot to realize the robot's visual navigation decision. In this embodiment, the global optimum is achieved through the training of multiple navigation tasks.

[0097] In summary, in the indoor visual navigation working scenario, by adopting the visual navigation method proposed in the present invention, the robot can execute navigation motion planning and smoothly reach the location of the target object, and the training time is significantly shortened.

[0098] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0099] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A mobile robot visual navigation method based on voxel spatial attention mechanism, characterized in that: Specifically include: S1. For the mobile robot in the scene, at any time t during the navigation process, the RGB image data and Depth depth data are obtained based on its visual sensor; Design a voxel conversion algorithm to convert RGB image data and Depth data into 3D point cloud data based on camera intrinsic and extrinsic parameters, and project it into 3D voxel space to obtain the voxel matrix of the mobile robot at any time t with 3D point cloud information. Similarly, for a given navigation task target’s RGBD data, the navigation target voxel matrix is ​​obtained in the same way: ; S2. Design a voxel spatial attention module to monitor the voxel matrix of the mobile robot at any time t. and the navigation target voxel matrix Processing is performed to calculate the attention weight of each voxel in the voxel grid A ; Based on attention weight A Select the region of interest and get the predicted position of the navigation target in the voxel grid; S3. Design an attention weight decoding module to decode the final output of the voxel space attention module into the action to be performed by the mobile robot in the next time step, so as to realize the navigation motion planning of the robot; S4. Based on the deep reinforcement learning framework, set the reward function and loss function, optimize the parameters of the two modules designed in steps S2 and S3, and realize the optimal navigation motion planning of the robot in the scene.

2. A mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 1, characterized in that: Step S1 specifically includes: S11, combining the RGB image data and the Depth data at time t into an RGB-D image of size (128, 128, 4), which carries two-dimensional pixel information and depth information; S12. By combining the intrinsic parameters of the visual sensor and the acquired pixel and depth information, a three-dimensional point cloud is constructed using the mapping relationship between two-dimensional and three-dimensional space; in the image coordinate system, the three-dimensional coordinates corresponding to a pixel point (u, v) and its depth value d in the camera coordinate system are recorded as (X, Y, Z); according to the camera imaging principle, the conversion relationship between the two-dimensional pixel coordinates and the three-dimensional point cloud data is obtained as follows: ； ； ； Where s represents the unit conversion scaling factor between pixels and millimeters, K is the intrinsic parameter matrix of the depth camera, and is the image focal length, and is the center of the depth image; parameter and The units of are all pixels, and the units of the 3D point cloud coordinates (X, Y, Z) to be solved are millimeters; S13, according to the conversion relationship in step S12, solve the three-dimensional point cloud coordinates to achieve the mapping between the two-dimensional pixel coordinates, that is, RGB-D data, and the three-dimensional point cloud data. The formula is expressed as: ； S14, voxelize the 3D point cloud data obtained in step S13 with a resolution of 32*32*32; first calculate the minimum bounding box of the 3D point cloud data; then divide the minimum bounding box into voxel spaces with a resolution of L; traverse each voxel, if a 3D space point exists, store 1, otherwise store 0; finally obtain a voxel matrix , which is the input of the voxel space attention module at time t.

3. The mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 1, characterized in that: The voxel spatial attention module designed in step S2 is specifically: a convolutional neural network structure for three-dimensional voxel data processing, which includes: An input preprocessing module is used to receive three-dimensional input data with a predetermined number of channels; the input preprocessing module includes three three-dimensional convolutional layers, and the three-dimensional convolutional layers use multi-size convolutional kernels to perform feature extraction, respectively extracting and The voxel features are obtained and ; Voxel cross attention machine module, and is the input and output is the attention weight A , and use it to update the voxel features ; Specifically include: using a three-dimensional convolution layer to process and , to obtain the query matrix Q and the key matrix K, and then process it with a three-dimensional convolution layer Get the value matrix V; and set the learnable hyperparameters , then we get the attention weight A The formula is: ； in, is the activation function; Finally, output the updated voxel features .

4. A mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 3, characterized in that: In step S2, the attention weight A Select the region of interest and get the predicted position of the navigation target in the voxel grid as follows: According to the calculated attention weight A Find the area where the target object is located: First define the sum operation, A Each voxel in The dimensional features are summed, and then based on the summation results of each voxel feature, the maximum value and its corresponding voxel index position are found , which is the predicted position of the navigation target; the mathematical expression of this process is: ,in is the attention weight A Median component; select voxel index position Central location Construct the subsequent loss function.

5. The mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 1, characterized in that: Step S3 is specifically as follows: Design an attention weight decoding module to decode the final output of the voxel space attention module into the action to be performed by the mobile robot in the next time step, so as to realize the navigation motion planning of the robot; the attention weight decoding module is composed of 3 3D convolutional layers and 2 fully connected layers to update the voxel features As input, output the probability values ​​of the robot's five discrete navigation actions, and select the action with the largest probability value as the robot's navigation action at the next moment; The five discrete actions include: walking forward, moving backward, turning left, turning right, and stopping.

6. The mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 1, characterized in that: Step S4 specifically includes: S41. Set the reward function, record the reward for each moment as -β, set the robot to perform only one action at each moment, and the reward for successful navigation is ; The criterion for whether the target position is reached is: when the instance of the target position is less than 1 meter away from the robot's camera and within the robot's field of view, and the attention weight decoding module outputs a stop action, the robot is considered to have reached the target position and successfully completed the navigation task; the reward function of the robot's navigation at each moment is expressed by the formula: ； S42. Use the Smooth L1 loss function to compare the predicted value and the true value of the target object’s area location, and add the attention weight A of Regularize the loss to prevent overfitting; the formula of the loss function is: ； in, is the difference between the predicted value and the true value of the target position, that is, , , Represent the predicted and true positions of the target in the voxel grid, respectively; Regularization loss: ； in, yes The regularized weight coefficient is a hyperparameter; is the attention weight A Medium weight, is the index; S43. At the same time, the traditional reinforcement learning model Asynchronous Advantage Actor-Critic is introduced, that is, the loss function of A3C , together with the Smooth L1 loss and the regularization loss, constitute the total loss function of the mobile robot visual navigation model based on the voxel space attention mechanism: ； in, It is a hyperparameter that represents the importance of each item in the loss function; S44. For the total loss function ,Based on the stochastic gradient descent algorithm, the parameters of the voxel space attention module and the attention weight decoding module are iteratively optimized to realize the robot visual navigation decision.

7. A mobile robot visual navigation method based on voxel spatial attention mechanism according to claim 6, characterized in that: Step S44 is specifically as follows: During the visual navigation simulation process, the target position is set multiple times for navigation tasks, and the reward function and loss function of each navigation task are recorded; with the minimization of the loss function as the direction, the parameters of the voxel space attention module and the attention weight decoding module are optimized until the overall reward function curve of all tasks converges, completing the training process and obtaining the optimized voxel space attention module and attention weight decoding module; the two optimized modules are deployed on the mobile robot to realize the robot's visual navigation decision.

Images