An automated truck loading and unloading capacity sensing and stacking planning method

Abstract

The application discloses an automatic loading and unloading truck capacity sensing and stacking planning method, and belongs to the technical field of automatic loading and unloading trucks. The truck cargo point cloud data graph and truck position information are obtained through a robot laser scanning sensor and an ultra-wideband technology; an edge correction is performed on the obtained point cloud data graph through a star map denoising method, so that the purpose of removing noise and deleting useless areas is achieved; the corrected point cloud graph is subjected to Delaunay subdivision, volume superposition is performed on the three prisms formed by the projection of the N triangular facet convex hulls subjected to the subdivision; the initial data set of reinforcement learning is continuously optimized through gradient descent, so that the optimal space utilization rate is obtained; and the optimal stacking quantity is calculated. The application is accurate in sensing and positioning of the truck position, accurate in judgment of the truck capacity and cargo volume, small in error range, and gradually close to the true value with the increase of the convex hull capacity; meanwhile, the optimal planning of the stacking mode can be realized in a short time, so that the space utilization rate is maximized, and the waste of space resources is reduced as much as possible.

Description

Technical Field

[0001] An automated loading and unloading capacity sensing and palletizing planning method belongs to the field of automated loading and unloading technology. Background Technology

[0002] Automated loading and unloading robot systems primarily stem from enterprises' demands for automation and intelligence in the loading and unloading process. It involves the integrated application of multiple technological fields, such as mechanical engineering, robotics, computer vision, sensing technology, and navigation technology. Advances and developments in these technologies have provided the foundation and support for the realization of automated loading and unloading robot systems. Currently, with the rapid development of the logistics industry and the continuous optimization of enterprise cargo loading and unloading supply chains, identifying truck locations, improving loading efficiency, increasing space utilization, reducing operating costs, and mitigating human resource risks have become key concerns for enterprises. The commonly used technologies for automated loading and unloading robot systems are as follows:

[0003] (1) Computer vision technology: using cameras for image recognition and feature extraction to determine the robot's position. However, visual recognition is easily affected by factors such as light, background changes and occlusion, and the recognition accuracy may sometimes be insufficient.

[0004] (2) Weight sensing technology: The truck's capacity is sensed by measuring the weight of the load by installing weighing sensors on the bottom of the truck. However, weight sensing may not be able to accurately identify the weight of a single object, and it needs to be compared with an external database, which limits its real-time performance and accuracy.

[0005] (3) Optimization algorithm: The palletizing planning problem is solved by optimization algorithm. However, optimization algorithm may require a long computation time and it is difficult to find the global optimal solution in complex scenarios.

[0006] To address these issues, further research and improvement of the technology and functions of automated loading and unloading robot systems are needed to enhance the robots' ability to perceive and adapt to complex environments, making their design more flexible, precise, and time-efficient. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an automated loading and unloading capacity sensing and palletizing planning method, which can accurately sense and locate the position of the truck, accurately judge the truck capacity and cargo volume, with a small error range, and the judged value gradually approaches the true value as the convex shell capacity increases.

[0008] The technical solution adopted by this invention to solve its technical problem is: the automated loading and unloading vehicle capacity sensing and palletizing planning method, characterized by including the following steps:

[0009] S1 sets up the initial dataset for reinforcement learning and obtains point cloud data of truck cargo and truck location information through robot laser scanning sensors and ultra-wideband technology.

[0010] S2 uses star map denoising to perform edge correction on the obtained point cloud data map, thereby removing noise and deleting useless areas.

[0011] The S3-corrected point cloud map is divided into Delaunay sections. The volume of the triangular prism formed by the projection of the N triangular facets onto the convex shells of the section is superimposed to obtain the desired sample.

[0012] S4 uses different stacking methods for N convex hull regions, continuously optimizes the initial dataset for reinforcement learning through gradient descent, updates the optimal policy, and obtains the optimal space utilization.

[0013] The optimal palletizing strategy is the palletizing method for the convex shell area with the best space utilization in S5, and the optimal number of pallets is calculated.

[0014] Preferably, the method further includes the robot being in state k at time t, and the action taken at this time being... The expected value function of the cumulative return at this point is:

[0015] ;

[0016] in, For cumulative returns, , Let t represent the state and action at time t. This represents the expected cumulative return.

[0017] Preferably, a reward function is set in the action space. To determine the optimal cargo quantity planning, the reward function is used. for:

[0018] ;

[0019] in, The estimated reward for state k at time t; The state at the next moment; This indicates that the initial state at time t is k. Let a represent the action performed at time t. This indicates that, given the current state is k and action a is taken, the policy is followed. The expected value of the cumulative rewards that can be obtained. For conditional expectation operators.

[0020] Preferably, the method further includes the robot using ultra-wideband technology to achieve high-precision judgment and positioning of the truck's location, including the following steps:

[0021] The S101 is equipped with an ultra-wideband transmitter on the truck to send short pulse signals.

[0022] The S102 robot, loading platform, and control box are equipped with sensors featuring ultra-wideband receivers to receive ultra-wideband signals.

[0023] After receiving the ultra-wideband signal, the S103 robot calculates the signal propagation time difference from the transmitter to the receiver. In conjunction with known receiver information, triangulation was used to determine the truck's location;

[0024] The distance between the truck and the base station is:

[0025] ;

[0026] The truck's location coordinates are: The location coordinates of the three base stations are , i=1,2,3.

[0027] Preferably, the truck cargo point cloud data map consists of three parts: the identified target, the identified background, and image noise. The truck cargo point cloud data map is as follows:

[0028] ;

[0029] in, For point cloud data pixels grayscale value, To highlight the target, To identify the grayscale values ​​of the background, This is an image noise signal;

[0030] The background grayscale value is identified by each data point following the extended definition:

[0031] ;

[0032] ;

[0033] in, To identify the grayscale values ​​of the background, For cloud data graph pixels grayscale value, , for The grayscale values ​​before and after.

[0034] Preferably, the method further includes the processing coefficients and center points of the Gaussian filter. The neighborhood is:

[0035] ;

[0036] ;

[0037] in, , This represents the distance between the grayscale point and the center position of the Gaussian filter. This represents the distance between the grayscale point and the center of the Gaussian filter.

[0038] Gaussian filter center point The method of using the neighborhood median instead of the difference is as follows:

[0039] ;

[0040] ;

[0041] in, For point cloud data pixels grayscale value, After data expansion go through The gray value after neighborhood median processing The final output grayscale value, where N is the center point. The median within the neighborhood, where p and q represent the number of distinct data extensions.

[0042] Preferably, the method further includes establishing a point cloud data triangular network comprising N triangular facets and a convex hull. The steps for establishing the triangular network are as follows:

[0043] S301 Settings For the planar layer, for the defined set of triangular element convex hull regions, find the largest inscribed circle and rotate the triangular element convex hull around any vertex.

[0044] S302 settings Planar layer, insert data points Find all data points The circumcircle, and remove the cavity region formed between the circumcircles;

[0045] S303 to Set of triangular convex shell regions in planar layers Planar layer data points A layout process was implemented to form a new triangular network;

[0046] S304 Fill the new triangular network with the corresponding point cloud data and delete the corresponding point cloud data. Repeat step S302 until all point cloud data has been traversed and inserted, and finally form a point cloud data triangular network containing N triangular facets and a convex hull.

[0047] Preferably, the method further includes projecting the formed triangular convex shell to form a triangular prism, wherein the base area and height of the i-th triangular prism are calculated using the corresponding triangular convex shell, and the calculation method is as follows:

[0048] ;

[0049] ;

[0050] , , The locations of the point cloud data for the vertices of the three triangular convex hulls. For the defined initial elevation data, , , These are the elevation data of the point cloud data of the three vertices of the j-th triangular convex hull;

[0051] The volume of the superimposed convex hull of N triangular facets is:

[0052] ;

[0053] in, , For the first The base area and height of a convex triangular prism This represents the error increment of the convex hull.

[0054] Preferably, the method further includes setting the maximum load limit as follows:

[0055] ;

[0056] in, This indicates rounding down. , These represent the truck's capacity and the volume of a single cargo item, respectively.

[0057] Preferably, the method further includes the gradient descent formula for the optimal spatial rate of change as follows:

[0058] ;

[0059] in, Let be the space utilization return value of the i-th palletizing method, i.e., the i-th short-term optimal strategy. Let N be the error increment of the convex shell, and N be the number of stacks. For the i-th palletizing method, For the i-th palletizing method, the quantity of goods carried in the convex shell area under the j-dimensional angle is given. Let i be the reward function for the i-th palletizing method;

[0060] The parallel optimization parameters for gradient descent are:

[0061] ;

[0062] in, Let N be the number of pallets to be stacked, and let N be the parallel optimization parameters for the i-th type. The gradient of the short-term optimal policy change. Let represent the quantity of goods carried in the three-dimensional convex shell region of the i-th palletizing method. Maximum cargo limit;

[0063] All N convex shell areas of the truck body adopt a stacking method based on optimal space utilization, and the optimal stacking quantity is calculated.

[0064] ;

[0065] in, To determine the optimal number of trucks for palletizing. For the volume of a single cargo, for The capacity of the corresponding three-dimensional convex hull region.

[0066] Compared with the prior art, the beneficial effects of this invention are:

[0067] This invention first uses a robot's laser scanning sensor to obtain and correct point cloud data of the truck and cargo. Simultaneously, it employs ultra-wideband triangulation to accurately determine the truck's position. N different triangular convex hulls are obtained through Delaunay partitioning, and their volumes are superimposed by projection to obtain the truck's capacity and cargo volume. Then, gradient descent is used to continuously update the reinforcement learning state dataset, and parallel optimization parameters are set to learn the optimal strategy. Finally, the optimal palletizing method with the best space utilization is obtained, thus solving the problem of optimal palletizing planning for trucks loaded with cargo using a robotic system. This method provides accurate truck position perception and localization, accurate judgment of truck capacity and cargo volume with a small error range, and the judged values ​​gradually approach the true values ​​as the convex hull capacity increases. Furthermore, it can achieve optimal palletizing planning in a short time, maximizing space utilization and minimizing space resource waste. Attached Figure Description

[0068] Figure 1 This is a flowchart analyzing the capacity-aware and palletizing planning method based on gradient reinforcement learning Delaunay partitioning.

[0069] Figure 2 It is a frequency domain distribution map of point cloud data noise;

[0070] Figure 3 This is a plan view of the truck bed and the Delaunay section.

[0071] Figure 4 It is a curve showing the error variation between the predicted and actual truck capacity values ​​based on Delaunay subdivision;

[0072] Figure 5 It is a structural model diagram of the overall system. Detailed Implementation

[0073] The present invention will be further described below with reference to specific embodiments. However, those skilled in the art should understand that the detailed description given here with reference to the accompanying drawings is for better explanation. The structure of the present invention necessarily exceeds the limited embodiments described herein. Some equivalent alternatives or common means will not be described in detail here, but still fall within the protection scope of this application. Figures 1-5 This is the preferred embodiment of the present invention, which is described below in conjunction with the accompanying drawings. Figures 1-5 The present invention will be further described below.

[0074] like Figure 1 As shown: An automated loading and unloading capacity sensing and palletizing planning method includes the following steps:

[0075] S1 sets up the initial dataset for reinforcement learning and obtains point cloud data of truck cargo and truck location information through robot laser scanning sensors and ultra-wideband technology.

[0076] S2 uses star map denoising to perform edge correction on the obtained point cloud data map, thereby removing noise and deleting useless areas.

[0077] The S3-corrected point cloud map is divided into Delaunay sections. The volume of the triangular prism formed by the projection of the N triangular facets onto the convex shells of the section is superimposed to obtain the desired sample.

[0078] S4 uses different stacking methods for N convex hull regions, continuously optimizes the initial dataset for reinforcement learning through gradient descent, updates the optimal policy, and obtains the optimal space utilization.

[0079] The optimal palletizing strategy is the palletizing method for the convex shell area with the best space utilization in S5, and the optimal number of pallets is calculated.

[0080] Given a dataset of reinforcement learning state representations Here, S consists of a four-tuple: C represents the truck's cargo compartment capacity, V represents the volume of a single cargo item, U represents the space utilization rate, and Q represents the quantity of cargo carried by the truck. During each state transition, the dataset S is updated internally. The space utilization rate reward value obtained each time satisfies... So the new The maximum space utilization reward value for this state transition is denoted as the short-time optimal strategy. In finding the optimal strategy, the goal is to maximize the expected cumulative reward function corresponding to that strategy. If the robot's state at time t is k, the action taken at that time is denoted as... Then, the expected cumulative return function is defined as follows:

[0081] ;

[0082] in, For cumulative returns, , Let t represent the state and action at time t. This represents the expected cumulative return.

[0083] Define the robot's state-action space Different palletizing methods on trucks and truck location Corresponding short-time optimal strategy The output is as follows: The space of the truck is divided into scattered cells or continuous regions, each cell or region representing a different palletizing method. Clearly, the size of the action space depends on the short-time optimal strategy. Different palletizing methods and truck sizes. A reward function is set in the motion space. To determine the optimal cargo quantity planning. During the data update process of taking different actions in different states, if the cumulative expected value function... The larger The larger the value, the more it incentivizes short-term optimal strategies. Update the state dataset . The calculation formula is as follows:

[0084] ;

[0085] in, The estimated reward for state k at time t; The state at the next moment; This indicates that the initial state at time t is k. Let a represent the action performed at time t. This indicates that, given the current state is k and action a is taken, the policy is followed. The expected value of the cumulative rewards that can be obtained. For conditional expectation operators.

[0086] The robot laser scanning sensor is a hardware device carried by the robot, which can indirectly acquire target point cloud data. This invention does not provide a detailed explanation of the sensor's operating principles. Instead, it provides a specific explanation of the truck position information determination using ultra-wideband technology.

[0087] The robot uses ultra-wideband technology to achieve high-precision judgment and positioning of trucks indoors or in complex environments. This includes the following steps:

[0088] S101 Installation of Ultra-Wideband Transmitter: An ultra-wideband transmitter is installed on the truck to send short pulse signals;

[0089] S102 is equipped with an ultra-wideband receiver: the robot, loading platform, and control box are equipped with sensors with ultra-wideband receivers to receive ultra-wideband signals;

[0090] S103 Calculates the time difference: After receiving the ultra-wideband signal, the robot calculates the signal propagation time difference from the transmitter to the receiver. By combining the known receiver information, triangulation is used to determine the truck's location.

[0091] Set the location coordinates of the three base stations as follows i=1, 2, 3; the truck's position coordinates are The distance between the truck's location and the base station satisfies:

[0092] ;

[0093] The truck's location coordinates are: The location coordinates of the three base stations are , i=1,2,3.

[0094] After obtaining point cloud data images of the truck and individual cargo using laser scanning sensors on robot hardware, star map denoising is applied to process the point cloud data image containing a large number of points. This achieves edge correction, noise removal, and deletion of useless regions to facilitate subsequent calculations. The star map denoising method performs preliminary processing using a Gaussian filter with data expansion, and uses the center point of the Gaussian filter as the basis for further processing. The median value within the neighborhood is used as the standard value. Points in areas with a large difference in grayscale value from the standard value are replaced with the median value and then deleted. After this process, the overall difference in the neighborhood of the point cloud data is significantly reduced, which reduces the error in data acquisition and avoids affecting subsequent subdivision calculations.

[0095] The point cloud data map of trucks and cargo consists of three parts: target identification, background identification, and image noise. This data can be represented by the following model:

[0096] ;

[0097] in, For point cloud data pixels grayscale value, To highlight the target, To identify the grayscale values ​​of the background, This represents the image noise signal.

[0098] Data augmentation removes areas of concentrated noise with significant differences by expanding the grayscale image segmentation regions that identify the background. Figure 2 The frequency domain distribution of noise in point cloud data is given. The background grayscale value of each data point follows the following extended definition:

[0099] ;

[0100] ;

[0101] in, To identify the grayscale values ​​of the background, For cloud data graph pixels grayscale value, , for The grayscale values ​​before and after.

[0102] After expanding the grayscale segmentation region, the center point of the identification target in the truck and cargo point cloud data map is overlapped with the center position of the Gaussian filter. Pooling and convolutional layer parameters are then set, and overlap-processing convolution operations are performed to obtain a grayscale image where the data sample points are close to the center position of the Gaussian filter. This facilitates the comparison of differences between points in different regions. Let x and y be the distances from any grayscale point in the grayscale image to the center position of the Gaussian filter, then the processing coefficients and center point of the Gaussian filter are... The neighborhood is:

[0103] ;

[0104] ;

[0105] in, , This represents the distance between the grayscale point and the center position of the Gaussian filter. This represents the distance between the grayscale point and the center position of the Gaussian filter. When... When the sample size is small, the changes in sample points are significant, and the point cloud map becomes more compact and clear.

[0106] Next, the point cloud map can be processed using extended Gaussian filtering. When the extended point cloud map data points... and When the grayscale values ​​of standard points within a neighborhood differ significantly, these are the areas we need to focus on processing. For cloud map data points with large differences... Using Gaussian filter center point The median of the neighborhood is used instead of deletion to reduce this dissimilarity. (Gaussian filter center point) The neighborhood median is used instead of the difference calculation formula:

[0107] ;

[0108] ;

[0109] in, For point cloud data pixels grayscale value, After data expansion go through The gray value after neighborhood median processing The final output grayscale value, where N is the center point. The median within the neighborhood, where p and q represent the number of distinct data extensions.

[0110] Figure 3 A comparison of the truck capacity point cloud data before and after correction is presented. It can be seen that by processing the point cloud data containing a large number of points using the star map denoising method, image correction of abnormal difference values ​​is achieved, thus realizing the purpose of noise reduction and cloud map optimization.

[0111] Based on the revised point cloud data, a triangular element convex hull is defined, which contains a portion of the point cloud data, forming a region set. All triangular elements must adhere to two conditions: first, the Delaunay triangulation is unique and invariant, and the circumcircle of any triangular element cannot contain any redundant points; second, among the possible triangulations formed by the scattered point set, the minimum angle of the triangle formed by the Delaunay triangulation is the largest.

[0112] Based on the above rules, a triangular network of point cloud data containing a convex hull with N triangular facets is constructed. The steps for constructing this triangular network are as follows:

[0113] S301 Settings For the planar layer, for the defined set of triangular element convex hull regions, find the largest inscribed circle and rotate the triangular element convex hull around any vertex.

[0114] S302 settings Planar layer, insert data points Find all data points The circumcircle, and remove the cavity region formed between the circumcircles;

[0115] S303 to Set of triangular convex shell regions in planar layers Planar layer data points A layout process was implemented to form a new triangular network;

[0116] S304 Fill the new triangular network with the corresponding point cloud data and delete the corresponding point cloud data. Repeat step S302 until all point cloud data has been traversed and inserted, and finally form a point cloud data triangular network containing N triangular facets and a convex hull.

[0117] Projecting the resulting triangular convex shell, we form a triangular prism. The base area and height of the i-th triangular prism are then calculated using the corresponding triangular convex shell, as shown in the following formula:

[0118] ;

[0119] ;

[0120] in, , , The locations of the point cloud data for the vertices of the three triangular convex hulls. For the defined initial elevation data, , , These are the elevation data of the point cloud data of the three vertices of the j-th triangular convex hull.

[0121] The volume of the superimposed convex hull of N triangular facets is:

[0122] ;

[0123] in, , For the first The base area and height of a convex triangular prism This represents the error increment of the convex hull.

[0124] By overlaying volumes, the original volume of the target in the point cloud data map is obtained, namely the truck capacity and the volume of a single cargo, which are respectively set as... , . Figure 4 The error curves between the predicted and actual truck capacity values ​​based on Delaunay subdivision are presented. As can be seen from the image, the truck capacity perception of this scheme is more accurate, and the principle of cargo volume perception is consistent with that of this scheme.

[0125] Given the maximum stacking limit is Let the stacking set of N triangular convex hull regions be . ,like Given a pallet set ;like The number of triangular convex shells in the Delaunay section is ;like If the initial data is not maintained, calculate the maximum cargo capacity that can be loaded based on the truck capacity and individual cargo volume obtained in step S303.

[0126] ;

[0127] in, This indicates rounding down to the nearest integer.

[0128] Gradient descent searches for the minimum along the negative gradient, that is, it searches for the rate of change of the ratio of the area occupied by the stacked goods in each triangular convex hull region to the total area of ​​the convex hull. The minimum value. The gradient descent formula for the optimal rate of change in space:

[0129] ;

[0130] in, Let be the space utilization return value of the i-th palletizing method, i.e., the i-th short-term optimal strategy. Let N be the error increment of the convex shell, and N be the number of stacks. For the i-th palletizing method, For the i-th palletizing method, the quantity of goods carried in the convex shell area under the j-dimensional angle is given. Let be the reward function for the i-th palletizing method.

[0131] Continuously optimize by iterating over parameters. Each new If satisfied Then the reward function Increase and update the robot's state and motion space. Data and reinforcement learning state dataset This will incentivize the robot to continue learning the optimal strategy.

[0132] Parallel optimization parameters during gradient descent:

[0133] ;

[0134] in, Let N be the number of pallets to be stacked, and let N be the parallel optimization parameters for the i-th type. The gradient of the short-term optimal policy change. Let represent the quantity of goods carried in the three-dimensional convex shell region of the i-th palletizing method. Maximum cargo limit;

[0135] By optimizing parameters in parallel using gradient descent, the rate of change of the area ratio of the regions is continuously minimized, thereby maximizing the iterative optimization of the spatial change rate of the optimal learning strategy. This allows the short-time optimal strategy to gradually stabilize, and the short-time optimal strategy at this point represents the optimal space utilization for the entire state. The N convex hull regions of the truck bed are all stacked using a method based on optimal space utilization, and their optimal stacking quantity is calculated.

[0136] ;

[0137] in, To determine the optimal number of trucks for palletizing. For the volume of a single cargo, for The capacity of the corresponding three-dimensional convex hull region.

[0138] In summary, we can now establish a structured model diagram of a system that integrates truck location determination, capacity sensing, and palletizing planning. Figure 5 A structural model diagram of the overall system is provided.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. An automated loading and unloading vehicle capacity sensing and palletizing planning method, characterized in that: Includes the following steps: S1 sets up the initial dataset for reinforcement learning and obtains point cloud data of truck cargo and truck location information through robot laser scanning sensors and ultra-wideband technology. S2 uses star map denoising to perform edge correction on the obtained point cloud data map, thereby removing noise and deleting useless areas. The S3-corrected point cloud map is divided into Delaunay sections. The volume of the triangular prism formed by the projection of the N triangular facets onto the convex shells of the section is superimposed to obtain the desired sample. S4 uses different stacking methods for N convex hull regions, continuously optimizes the initial dataset for reinforcement learning through gradient descent, updates the optimal policy, and obtains the optimal space utilization. The optimal palletizing strategy is the palletizing method for the convex shell area with the best space utilization in S5, and the optimal number of pallets is calculated. The method further includes establishing a point cloud data triangular network containing N triangular facets and a convex hull. The steps for establishing the triangular network are as follows: S301 Settings For the planar layer, for the defined set of triangular element convex hull regions, find the largest inscribed circle and rotate the triangular element convex hull around any vertex. S302 settings Planar layer, insert data points Find all data points The circumcircle, and remove the cavity region formed between the circumcircles; S303 to Set of triangular convex shell regions in planar layers Planar layer data points A layout process was implemented to form a new triangular network; S304 Fill the new triangular network with the corresponding point cloud data and delete the corresponding point cloud data. Repeat step S302 until all point cloud data has been traversed and inserted, and finally form a point cloud data triangular network containing N triangular facets and a convex hull. The method further includes projecting the formed triangular convex shell to form a triangular prism. The base area and height of the i-th triangular prism are calculated using the corresponding triangular convex shell, as follows: ； ； in, These are the position coordinates of the point cloud data of the three vertices of the triangular convex hull. For the defined initial elevation data, , , These are the elevation data of the point cloud data of the three vertices of the j-th triangular convex hull; The volume of the superimposed convex hull of N triangular facets is: ； in, , For the first The base area and height of a convex triangular prism This represents the error increment of the convex hull. The method also includes a maximum cargo loading limit of: ； in, This indicates rounding down. , These represent the truck's capacity and the volume of a single cargo item, respectively. The method also includes the gradient descent formula for the optimal spatial rate of change: ； in, Let be the space utilization return value of the i-th palletizing method, i.e., the i-th short-term optimal strategy. This represents the space utilization return value for the (i-1)th palletizing method. Let N be the error increment of the convex shell, and N be the number of stacks. For the i-th palletizing method, For the i-th palletizing method, the quantity of goods carried in the convex shell area under the j-dimensional angle is given. Let i be the reward function for the i-th palletizing method; The parallel optimization parameters for gradient descent are: ; in, Let be the parallel optimization parameters for the i-th type. Here, N represents the number of pallets to be stacked, and the parameter is the parallel optimization parameter for the (i-1)th type. The gradient of the short-term optimal policy change. Let represent the quantity of goods carried in the three-dimensional convex shell region of the i-th palletizing method. Maximum cargo limit; All N convex shell areas of the truck body adopt a stacking method based on optimal space utilization, and the optimal stacking quantity is calculated. ； in, To determine the optimal number of trucks for palletizing. For the volume of a single cargo, for The capacity of the corresponding three-dimensional convex hull region.

2. The automated loading and unloading vehicle capacity sensing and palletizing planning method according to claim 1, characterized in that: The method further includes the robot being in state k at time t, and the action taken at that time being... The expected value function of the cumulative return at this point is: ； in, For cumulative returns, , Let t represent the state and action at time t. This represents the expected cumulative return.

3. The automated loading and unloading vehicle capacity sensing and palletizing planning method according to claim 2, characterized in that: Set the reward function in the action space To determine the optimal cargo quantity planning, the reward function is used. for: ； in, The estimated reward for state k at time t; The state at the next moment; This indicates that the initial state at time t is k. Let a represent the action performed at time t. This indicates that, given the current state is k and action a is taken, the policy is followed. The expected value of the cumulative rewards that can be obtained. For conditional expectation operators.

4. The automated loading and unloading vehicle capacity sensing and palletizing planning method according to claim 1, characterized in that: The method also includes the robot using ultra-wideband technology to achieve high-precision judgment and positioning of the truck's location, including the following steps: S101 is equipped with an ultra-wideband transmitter on the truck to send short pulse signals; The S102 robot, loading platform, and control box are equipped with sensors featuring ultra-wideband receivers to receive ultra-wideband signals. After receiving the ultra-wideband signal, the S103 robot calculates the signal propagation time difference from the transmitter to the receiver. In conjunction with known receiver information, triangulation was used to determine the truck's location; The distance between the truck and the base station is: ； The truck's location coordinates are: The location coordinates of the three base stations are , i=1,2,3.

5. The automated loading and unloading vehicle capacity sensing and palletizing planning method according to claim 1, characterized in that: The truck cargo point cloud data map consists of three parts: target identification, background identification, and image noise. The truck cargo point cloud data map is as follows: ； in, For point cloud data pixels grayscale value, To highlight the target, To identify the grayscale values ​​of the background, This is an image noise signal; The background grayscale value is identified by each data point following the extended definition: ； ； in, To identify the grayscale values ​​of the background, For cloud data graph pixels grayscale value, , for The grayscale values ​​before and after.

6. The automated loading and unloading vehicle capacity sensing and palletizing planning method according to claim 5, characterized in that: The method also includes the processing coefficients and center point of the Gaussian filter. The neighborhood is: ； ； in, , This represents the distance between the grayscale point and the center position of the Gaussian filter. This represents the distance between the grayscale point and the center position of the Gaussian filter. These are the coefficients after Gaussian filtering. It refers to the neighborhood range; Gaussian filter center point The method of using the neighborhood median instead of the difference is as follows: ； ； in, For point cloud data pixels grayscale value, After data expansion go through The gray value after neighborhood median processing The final output grayscale value, where N is the center point. The median within the neighborhood, where p and q represent the number of distinct data extensions.

Images