Flexible plugging method and device for charging mechanical arm of electric vehicle

Through the combination of 3D cameras and reinforcement learning models, the calculation capability limitation problem of the automatic charging robot arm in the identification and plug-in process of electric vehicle charging interfaces is solved, and high-precision flexible plug-in is realized to adapt to the plug-in of different types of charging interfaces demand, improving charging efficiency and reliability.

CN118990487APending Publication Date: 2024-11-22山东华科信息技术有限公司 +6
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411221440.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-02
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing automatic charging robot arm has computing capability limitations in the identification and flexible plug-in process of electric vehicle charging interfaces, resulting in positioning errors and plug-in difficulties. Especially for the differences in specifications of different types of charging interfaces, it is difficult to achieve high-precision plug-in.

Method used

Point cloud data is collected through a 3D camera, noise reduction filtering and downsampling are performed, and the charging interface target point cloud is identified using DBSCAN clustering, combined with the charging interface model matching relationship, 6D poses are obtained, and reinforcement learning model is used to fuse visual and force feedback control. The plug-in process realizes flexible plug-in of the robot arm.

Benefits of technology

It effectively reduces positioning errors during the plug-in process, improves the plug-in accuracy and reliability of the charging robot arm to different types of charging interfaces, and ensures the stability and automation of the charging process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118990487A_ABST
    Figure CN118990487A_ABST
Patent Text Reader

Abstract

The invention relates to a flexible plugging method and device for a charging mechanical arm of an electric vehicle. The method comprises the following steps: collecting point clouds at an automobile charging interface, carrying out noise reduction filtering, downsampling and DBSCAN clustering on the point clouds to segment point cloud clusters, and identifying a target point cloud cluster forming the charging interface according to a first composite feature of the point cloud clusters; obtaining a target conversion relation of the target point cloud cluster relative to the charging interface model by utilizing a corresponding matching relation between the target point cloud cluster and the charging interface model, and obtaining a 6D pose of the target point cloud cluster relative to a mechanical arm coordinate system based on a conversion relation between a charging interface model coordinate system and the mechanical arm coordinate system and the target conversion relation; inserting according to a 6D pose planning path; and in the process of plugging the charging head, a reinforcement learning model for generating an action and an expected docking result by fusing and analyzing a docking video and force parameter feedback is used for controlling the plugging process, so that smooth plugging of the charging mechanical arm is realized.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of charging mechanical arm control, and in particular to a method and device for compliant plugging of a charging mechanical arm of an electric vehicle. Background Art

[0002] With the popularity of new energy electric vehicles and the increasing market size, the automatic charging of electric vehicles has gradually become a research hotspot. At present, the plugging and unplugging of electric vehicles is mainly done by manpower. In some commercial and public places, logistics and warehousing, there are many electric working vehicles or electric passenger cars. There is an urgent need for an "automatic charging robot arm" to complete these complicated plugging and unplugging tasks to improve charging efficiency and save equipment and management costs.

[0003] For the automatic charging robot arm, there are two difficulties in smooth plugging: one is the identification and positioning of the vehicle charging interface, and the other is how to achieve smooth plugging. Identification and positioning through the charging interface point cloud is an effective way. Due to cost considerations, the computing performance of the robot arm's control device is limited. Due to the computing power of the robot arm's control device, high-precision point cloud analysis and calculation processes are difficult to achieve. Generally, the point cloud is sampled, which inevitably introduces more errors during positioning, and it is impossible to directly perform plug-in control based on the estimated positioning results; in addition, the charging interface of electric vehicles is divided into DC charging interface and AC charging interface. For different types of vehicles, the specifications of the charging interface may also be different, which brings difficulties to the smooth plugging and unplugging of the robot arm. Summary of the invention

[0004] In order to solve the above technical problem or at least partially solve the above technical problem, the present invention provides a method and device for flexible plugging of an electric vehicle charging robot arm.

[0005] In a first aspect, the present invention provides a method for compliant plugging of an electric vehicle charging robot arm, comprising:

[0006] The point cloud at the car charging interface is collected by a 3D camera, and the point cloud is subjected to noise reduction filtering and downsampling to obtain a sparse point cloud; the sparse point cloud is clustered by DBSCAN to segment the point cloud clusters contained in the sparse point cloud, and the target point cloud clusters constituting the charging interface are identified according to the first composite feature of the point cloud cluster; the target transformation relationship of the target point cloud cluster relative to the charging interface model is obtained by using the corresponding matching relationship between the target point cloud cluster and the charging interface model, and then the 6D pose of the target point cloud cluster relative to the robotic arm coordinate system is obtained based on the transformation relationship and the target transformation relationship between the charging interface model coordinate system and the robotic arm coordinate system;

[0007] The initial docking target position is determined based on the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system. The Dijkstra algorithm is used for path planning. A collision-free path is planned from the current position of the robot arm to the initial docking target position. The robot arm is moved to the vehicle charging port for docking according to the planned path.

[0008] During the process of plugging the charging head, the camera is used to collect the docking video of the charging plug and the car's charging interface, and the six-axis force sensor is used to collect the force parameters of the robotic arm during the plugging process. The docking video and force parameter feedback are fused and analyzed to generate a reinforcement learning model of the action and expected docking results to control the plugging process and achieve smooth plugging of the charging robotic arm.

[0009] Furthermore, the sparse point cloud is clustered by DBSCAN to segment the point cloud clusters contained in the sparse point cloud, and the target point cloud clusters constituting the charging interface are identified according to the first composite feature of the point cloud clusters, including: defining the minimum number of points included in each cluster, defining the neighborhood range of a point in each cluster; initializing all points in the sparse point cloud as unvisited points;

[0010] Traverse the unvisited points. For each unvisited point, if the number of points in the neighborhood of the unvisited point is greater than the minimum number of included points, the unvisited point is a core point. Create a new cluster, recursively add the points in the direct and indirect neighborhoods to the created new cluster, and change the state to visited. Finally, output the point cloud clustering representing the structure.

[0011] For each point cloud cluster, the principal component analysis method is used to extract shape features; the maximum range of the point cloud data of each point cloud cluster in three-dimensional space is calculated as the size feature; the surface features are calculated using the normal information and curvature of the point cloud data of the point cloud cluster; the shape features, size features and surface features of the point cloud cluster are used as the first composite features to match the corresponding first composite features of the charging interface model, and the target point cloud cluster that constitutes the charging interface is screened out from all point cloud clusters.

[0012] Furthermore, the target conversion relationship between the target point cloud clustering and the charging interface model is obtained by using the corresponding matching relationship between the target point cloud clustering and the charging interface model, including:

[0013] The charging interface model is sampled uniformly on a large scale to obtain several sampling points;

[0014] For each sampling point, find the nearest neighbor to it;

[0015] Calculate the distance between the sampling point and its nearest neighbor, the angle between the projection line of the sampling point and its nearest neighbor projected onto any two fixed orthogonal planes of the charging interface model coordinate system and the axis of the non-intersecting line in the corresponding two orthogonal planes, and the angle between the normal line of the sampling point and its nearest neighbor; construct the second composite feature between the sampling point and its nearest neighbor through the above parameters;

[0016] For the target point cloud clustering constituting the charging interface, all possible sampling points and the nearest neighboring points of the sampling points are selected, and the second composite features of all possible sampling points and the nearest neighboring points of the sampling points are calculated. The corresponding sampling points and the nearest neighboring points of the sampling points in the charging interface model are matched and found from the hash table or storage tree of the second composite features of the sampling points and the nearest neighboring points of the sampling points in the charging interface model.

[0017] For a number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, a transformation matrix between the two is calculated, so as to obtain a number of transformation matrices that define the position transformation relationship between the number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, wherein the transformation matrix includes a rotation matrix and a translation matrix;

[0018] The rotation angles and translation distances corresponding to all transformation matrices are used to vote for the rotation angles and translation distances within the set posture space range, and the voting results of the rotation angles and translation distances within the set posture space range are clustered by meanshift. The rotation angles and translation distances corresponding to the main cluster with a large number of votes are selected to obtain the target transformation matrix of the target point cloud cluster relative to the charging interface model.

[0019] Furthermore, a hash table or a storage tree is constructed using the second composite feature; the hash table uses the second composite feature as a key and the sampling point and the nearest neighbor of the sampling point of the corresponding feature on the charging interface model as a value; the storage tree includes a five-layer structure, the root node of the first layer stores the distance between the sampling point and the nearest neighbor of the sampling point, the second layer tree nodes and the third layer tree nodes store the two angles between the projection line of the sampling point and the axis of the non-intersecting line in the corresponding two orthogonal planes after the line connecting the sampling point and the nearest neighbor of the sampling point is projected onto any two fixed orthogonal planes of the charging interface model coordinate system, the fourth layer tree nodes store the angle between the sampling point and the normal of the nearest neighbor of the sampling point, and the leaf nodes of the fifth layer store the sampling point and the nearest neighbor of the second composite feature stored in the tree node where the corresponding leaf node is located and the root node.

[0020] Furthermore, the reinforcement learning model includes: a docking image feature encoder based on a convolutional network, a Softmax layer, a fully connected network, a weight generation network and a docking image feature decoder based on a transposed convolutional network, wherein the docking image feature encoder receives a current docking video frame sequence [P1, P2....P n ], the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points and the force parameters corresponding to each docking video frame in the current docking video frame sequence, the distance moved by the robot arm toward the vehicle charging interface, the action and the type encoding of the corresponding charging interface type are input into the fully connected network, and the fully connected network obtains the predicted next docking video frame The weight generation network generates n groups of weights based on the attention points corresponding to each docking video frame in the next docking video frame sequence, and each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted results are input into the docking image feature decoder to predict each docking video frame in the next docking video frame sequence; the action predicted by the fully connected network is provided to the charging robot arm for execution; wherein the type code is used as a condition to enable the fully connected network to perform action control for the corresponding type of charging interface.

[0021] Furthermore, the fully connected network is trained by the DDQN algorithm, and two fully connected networks are initialized as the main Q network and the target Q network respectively;

[0022] The following process is repeated until the algorithm converges or the set number of training rounds is reached:

[0023] According to the current state s, the ε-greedy strategy is used to select action a; the current state s includes: the attention point in the docking video frame, the force parameters corresponding to the docking video frame, the distance and type code of the robot arm moving toward the car charging port; the ε-greedy strategy refers to randomly selecting an action with a certain probability ε, and selecting the action that maximizes the Q value with a probability of 1-ε; the action a includes the index of the moving direction and the index of the moving distance, where the moving direction includes 8 directions: up, down, left, right, upper left, lower left, upper right, and lower right;

[0024] Execute action a and observe the reward r and new state s' of the feedback of the charging robot's moving action;

[0025] The experience you will gain<s,a,r,s'> Stored in the experience pool;

[0026] Randomly extract a batch of experience data from the experience pool for training;

[0027] For each sample<s,a,r,s'> , calculate the target Q value:

[0028] Use the main Q network to calculate the Q values ​​of all possible actions in state s';

[0029] Use the target Q network to calculate the maximum Q value corresponding to the action selected by the main Q network in state s';

[0030] The target Q value calculation formula is: Q target =r+γ*max(Q target (s', a')), where a' is the value that makes Q main (s', a') is the largest action, γ is the discount factor, and r is the reward:

[0031]

[0032] Among them, R is the reward score constant, and L is the average distance threshold;

[0033] Calculate the loss function;

[0034] Based on the loss obtained from the loss function, the parameters of the main Q network are updated using the gradient descent method;

[0035] Every certain number of steps, the parameters of the main Q network are copied to the target Q network.

[0036] Furthermore, the loss function is as follows: LOSS = αLoss1 + βLoss2 + κLoss3; wherein α, β, and κ are the corresponding weight coefficients, and Loss1 represents the predicted next docking video frame. And the next real docking video frame P n+1 Loss2 represents the mean Euclidean distance between the predicted attention points and the true attention points, and Loss3 represents the mean square error of the Q value calculated by the main Q network and the target Q network.

[0037] Furthermore, in the reinforcement learning model, the docking image feature encoder, the Softmax layer, the weight generation network and the docking image feature decoder based on the transposed convolutional network follow the following constraints: the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points are input into the weight generation network to generate n groups of weights, each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted results are input into the docking image feature decoder to restore each docking video frame in the current docking video frame sequence.

[0038] In a second aspect, the present invention provides a flexible control device for an electric vehicle charging robot arm, comprising: at least one processing unit, wherein the processing unit, a storage unit, an execution unit and a collection unit are interconnected through a bus unit, and the storage unit stores a computer program. When the computer program is executed by the processing unit, the flexible plug-in method for the electric vehicle charging robot arm is implemented.

[0039] In a third aspect, the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for flexible plugging of an electric vehicle charging robot arm is implemented.

[0040] The above technical solution provided by the embodiment of the present invention has the following advantages compared with the prior art:

[0041] The present application obtains point cloud clusters of relevant structures at the automobile charging interface by filtering, downsampling, and clustering the point cloud, matches the first composite feature of the point cloud clusters with the first composite feature of the preset charging interface model, and can identify the target point cloud clusters that constitute the automobile charging interface, and obtains the target conversion relationship of the target point cloud clusters relative to the charging interface model by using the corresponding matching relationship between the target point cloud clusters and the second composite feature of the charging interface model, and then obtains the 6D pose of the target point cloud clusters relative to the robot arm coordinate system. In the process of plugging the charging head, the docking video of the charging plug and the automobile charging interface is collected by the camera, and the force parameters of the robot arm during the plugging process are collected by the six-axis force sensor. The reinforcement learning model of the action and the expected docking result is generated by the fusion analysis of the docking video and the force parameter feedback to control the action during the plugging process, so as to realize the smooth plugging of the charging robot arm. In the plugging process, a reinforcement learning model of fusion analysis of visual and tactile information is provided, and the action adjustment is controlled by mutual assistance of visual and force information, so as to provide sufficient parameter basis for the action adjustment of the charging plug of the complex charging interface, and ensure the accuracy of the predicted action. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0043] 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, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0044] Figure 1 A flow chart of a method for compliant plugging of an electric vehicle charging robot arm provided by an embodiment of the present invention;

[0045] Figure 2 A flowchart of performing DBSCAN clustering on a sparse point cloud to segment each cluster contained in the sparse point cloud, and identifying a target point cloud cluster constituting a charging interface according to a first composite feature of each cluster, provided in an embodiment of the present invention;

[0046] Figure 3 A flowchart for obtaining 6D pose estimation of target point cloud clustering relative to a robotic arm coordinate system provided by an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of a storage tree provided by an embodiment of the present invention;

[0048] Figure 5 A schematic diagram of a reinforcement learning model provided by an embodiment of the present invention;

[0049] Figure 6 A schematic diagram of a device for implementing a flexible plugging method for an electric vehicle charging robot arm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] It should be noted that, in this article, the terms "include", "comprises" or any other variations 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, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0052] Example 1

[0053] like Figure 1 As shown, the technology of the present invention realizes a flexible plugging method for an electric vehicle charging mechanical arm, and the steps include:

[0054] The purpose of the patent of this invention is to provide a flexible plug-in and pull-out strategy for a charging robot arm based on force feedback. Through a charging socket search algorithm based on a multi-layer perceptron MLP and a force-position hybrid controller, visual positioning errors are compensated, and admittance control is used to achieve precise plug-in and pull-out of the charging gun. The electric vehicle charging method with flexible control completely simulates the manual charging process, can move freely in complex outdoor environments, has good adaptability, can automatically charge electric vehicles, and can smoothly insert the charging gun into the charging socket, avoiding damage to the vehicle.

[0055] The main steps of the compliance control method involved in this patent are as follows:

[0056] The point cloud at the car charging port is collected by a 3D camera, and the point cloud is subjected to noise reduction filtering and downsampling processing to obtain a sparse point cloud.

[0057] The point cloud only contains the point cloud data of the physical structure near the car charging port. For computers or control devices, the car charging port cannot be identified by the point cloud alone. In order to avoid noise interference and reduce the subsequent calculation amount, the point cloud is subjected to denoising filtering and downsampling to obtain a sparse point cloud. The denoising filtering removes points that do not have enough neighbors within a specified radius, and the downsampling is performed by voxel downsampling.

[0058] Perform DBSCAN clustering on the sparse point cloud to segment the clusters contained in the sparse point cloud, and identify the target point cloud clusters constituting the charging interface according to the first composite feature of each cluster;

[0059] In the specific implementation process, Figure 2 As shown, including:

[0060] Define the minimum number of points included in each cluster, define the neighborhood range of a point in each cluster, define the minimum number of points included in each cluster, and define the neighborhood range of a point in each cluster;

[0061] Initialize all points in the sparse point cloud as unvisited points;

[0062] Traverse the unvisited points. For each unvisited point, if the number of points in the neighborhood of the unvisited point is greater than the minimum number of included points, the unvisited point is a core point. Create a new cluster, recursively add the points in the direct and indirect neighborhoods to the created new cluster, and change the state to visited. Finally, output the point cloud clustering representing the structure.

[0063] For each point cloud cluster, the principal component analysis method is used to extract shape features; the maximum range of the point cloud data of each point cloud cluster in three-dimensional space is calculated as the size feature; the surface features are calculated using the normal information and curvature of the point cloud data of the point cloud cluster; the shape features, size features and surface features of the point cloud cluster are used as the first composite features to match the corresponding first composite features of the charging interface model, and the target point cloud cluster that constitutes the charging interface is screened out from all point cloud clusters.

[0064] The charging interface model is a model pre-created for various types of automobile charging interface structures. In order to achieve the matching of the charging interface model with the target point cloud cluster, a first composite feature is pre-calculated for each charging interface model. The charging interface model is also used to estimate the 6D pose of the target point cloud cluster.

[0065] For the target point cloud clusters that constitute the charging interface, the target transformation relationship between the target point cloud clusters and the charging interface model is obtained by using the corresponding matching relationship between the target point cloud clusters and the charging interface model. Then, based on the transformation relationship and the target transformation relationship between the charging interface model coordinate system and the robotic arm coordinate system, the 6D pose estimation of the target point cloud clusters relative to the robotic arm coordinate system is obtained, as shown in Figure 3 As shown, including:

[0066] The charging interface model is sampled uniformly on a large scale to obtain several sampling points.

[0067] For each sampling point of the charging interface model, find the nearest neighboring point to it.

[0068] Calculate the distance between the sampling point and the nearest neighbor of the sampling point, the angle between the projection line of the sampling point and the nearest neighbor of the sampling point after projecting onto any two fixed orthogonal planes of the charging interface model coordinate system and the axis of the non-intersecting line in the corresponding two orthogonal planes, and the angle between the normal of the sampling point and the nearest neighbor of the sampling point; construct a second composite feature between the sampling point and the nearest neighbor of the sampling point through the above parameters; construct a hash table or storage tree using the second composite feature; the hash table uses the second composite feature as the key and the sampling point and the nearest neighbor of the sampling point of the corresponding feature on the charging interface model as the value; such as Figure 4 As shown, the storage tree includes a five-layer structure, the root node of the first layer stores the distance between the sampling point and the nearest neighbor of the sampling point, the second layer tree nodes and the third layer tree nodes store the two angles between the projection line after the sampling point and the nearest neighbor line of the sampling point are projected onto any two fixed orthogonal planes of the charging interface model coordinate system and the axes of the non-intersecting lines in the corresponding two orthogonal planes, the fourth layer tree nodes store the angle between the sampling point and the normal of the nearest neighbor of the sampling point, and the leaf nodes of the fifth layer store the sampling points and the nearest neighbor of the second composite feature stored in the tree nodes and root nodes corresponding to the leaf nodes.

[0069] For the target point cloud clustering that constitutes the charging interface, all possible sampling points and the nearest neighboring points of the sampling points are selected, and the second composite features of all possible sampling points and the nearest neighboring points of the sampling points are calculated; the corresponding sampling points and the nearest neighboring points of the sampling points in the charging interface model are matched and found from the hash table or storage tree using the second composite features of all possible sampling points and the nearest neighboring points of the sampling points.

[0070] For a number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, a transformation matrix between the two is calculated, thereby obtaining a number of transformation matrices that define the position transformation relationship between a number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, wherein the transformation matrix includes a rotation matrix and a translation matrix.

[0071] The rotation angles and translation distances corresponding to all transformation matrices are used to vote for the rotation angles and translation distances within the set posture space range, and the voting results of the rotation angles and translation distances within the set posture space range are clustered by meanshift. The rotation angles and translation distances corresponding to the main cluster with a large number of votes are selected to obtain the target transformation matrix of the target point cloud cluster relative to the charging interface model.

[0072] Based on the target transformation matrix of the target point cloud clustering relative to the charging interface model, the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system is obtained by utilizing the transformation relationship between the charging robot arm coordinate system and the charging interface model coordinate system.

[0073] Since the obtained 6D pose has inevitable errors, it is not enough to guide the robot arm to directly plug the charger plug into the car charging port. The initial docking target position is determined according to the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system, and the Dijkstra algorithm is used for path planning to plan a collision-free path from the current position of the robot arm to the initial docking target position.

[0074] According to the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system, an initial docking target position is set for the robot arm, that is, the coordinate of the docking point with the charging port. When calculating the motion trajectory of the robot arm, an offset is given to the target point to ensure a certain fault tolerance during docking. The size of the offset can be adjusted according to actual needs to ensure that the robot arm can dock the charging port more accurately.

[0075] When planning the path, the Dijkstra algorithm is used for path planning according to the position and size of the obstacles, and the motion trajectory of the robot is generated according to the planned path. Among them, the target detection method is used to detect the obstacles on the robot's motion path, and the detected obstacle information is converted into coordinates in the robot's coordinate system.

[0076] Finally, the generated motion trajectory is converted into a command sequence for the robot to execute. Through the robot controller, the various joints of the robot are driven to move according to the command sequence. During the movement, the position and posture of the robot are monitored in real time to ensure that the robot moves according to the predetermined trajectory. During the movement of the robot, the changes in the surrounding environment are constantly detected, especially the movement of obstacles or new obstacles. If new obstacles or changes in the environment are found, the path planning and trajectory generation are re-performed, and the movement of the robot is fine-tuned according to the real-time feedback data of the robot.

[0077] During the process of plugging the charging head, the docking video of the charging plug and the car charging interface is collected through the camera, and the force parameters of the robot arm during the plugging process are collected through the six-axis force sensor. The reinforcement learning model that generates the action and expected docking results by fusion analysis of the docking video and force parameter feedback is used to control the plugging process and realize the smooth plugging of the charging robot arm. The smooth plugging strategy of the charging robot arm is adopted by combining video and force feedback. The charging socket search algorithm based on multi-layer perceptron and force-position hybrid controller is used to compensate for the visual positioning error, and the admittance control is used to realize the precise and smooth plugging of the charging gun, and the posture of the robot arm is adjusted in real time during the insertion process.

[0078] like Figure 5 As shown, the reinforcement learning model includes: a docking image feature encoder based on a convolutional network, a Softmax layer, a fully connected network, a weight generation network and a docking image feature decoder based on a transposed convolutional network, wherein the docking image feature encoder receives a current docking video frame sequence [P1, P2....P n ], the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points and the force parameters corresponding to each docking video frame in the current docking video frame sequence, the distance moved by the robot arm toward the vehicle charging interface, the action and the type encoding of the corresponding charging interface type are input into the fully connected network, and the fully connected network obtains the predicted next docking video frame The corresponding action and the attention point corresponding to each docking video frame in the next docking video frame sequence, the weight generation network generates n groups of weights based on the attention point corresponding to each docking video frame in the next docking video frame sequence, each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted result is input into the docking image feature decoder to predict each docking video frame in the next docking video frame sequence; the action predicted by the fully connected network is provided to the charging robot arm for execution; wherein the type code is used as a condition to enable the fully connected network to perform action control for the corresponding type of charging interface. An exemplary weight generation network includes: stacked convolutional layers, normalization layers, activation function layers and maximum pooling layers, and the stacked solution structure is finally linked to the sigmoid activation function.

[0079] In the reinforcement learning model, the docking image feature encoder, the Softmax layer, the weight generation network, and the docking image feature decoder based on the transposed convolutional network follow the following constraints: the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points are input into the weight generation network to generate n groups of weights, each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted result is input into the docking image feature decoder to restore each docking video frame in the current docking video frame sequence and add the attention points. That is, the docking image feature encoder, the Softmax layer, the weight generation network, and the docking image feature decoder based on the transposed convolutional network use the encoding and decoding of the current docking video frame sequence to add a set number of attention points for training, and the loss function used for training is expressed as: the weighted sum of Loss4 and Loss5, Loss4 represents the Euclidean distance between the restored docking video frame and the real docking video frame, and Loss5 represents the mean of the Euclidean distance between the annotated attention points and the predicted attention points.

[0080] The fully connected network is trained using the DDQN algorithm. During the training process, the parameters of the docking image feature encoder, the Softmax layer, the weight generation network, and the docking image feature decoder based on the transposed convolutional network are locked. The specific process includes:

[0081] Initialize two fully connected networks as the main Q network and the target Q network respectively;

[0082] The following process is repeated until the algorithm converges or the set number of training rounds is reached:

[0083] According to the current state s, the ε-greedy strategy is used to select action a; wherein the current state s includes: the attention point in the docking video frame, the force parameter corresponding to the docking video frame, the distance the robot moves toward the car charging interface, and the type code. Since the type code is introduced to distinguish different types of charging interfaces, it supports the introduction of the experience of different types of charging interfaces into the experience pool, so that the fully connected network can learn the control methods of different types of charging interfaces during the training process. In the actual control process, after the type code is input into the fully connected network, the parameter distribution of the fully connected network meets the action requirements when the corresponding type of charging interface is plugged in under the constraint of the type code; the ε-greedy strategy refers to randomly selecting an action under a certain probability ε, and selecting the action that maximizes the Q value under the probability of 1-ε; action a includes the index of the moving direction and the index of the moving distance, wherein the moving direction includes 8 directions: up, down, left, right, upper left, lower left, upper right, and lower right;

[0084] Execute action a and observe the reward r and new state s' of the feedback of the charging robot's moving action;

[0085] The experience you will gain<s,a,r,s'> Stored in the experience pool;

[0086] Randomly extract a batch of experience data from the experience pool for training;

[0087] For each sample<s,a,r,s'> , calculate the target Q value:

[0088] Use the main Q network to calculate the Q values ​​of all possible actions in state s';

[0089] Use the target Q network to calculate the maximum Q value corresponding to the action selected by the main Q network in state s';

[0090] The target Q value calculation formula is: Q target =r+γ*max(Q target (s', a')), where a' is the value that makes Q main (s', a') is the largest action, γ is the discount factor, and r is the reward:

[0091]

[0092] Among them, R is the reward score constant, and L is the average distance threshold;

[0093] The loss function is calculated as follows: LOSS = αLoss1 + βLoss2 + κLoss3; where α, β, and κ are the corresponding weight coefficients, and Loss1 represents the predicted next docking video frame. And the next real docking video frame P n+1Loss2 represents the mean Euclidean distance between the predicted attention points and the true attention points, and Loss3 represents the mean square error of the Q value calculated by the main Q network and the target Q network.

[0094] Based on the loss obtained from the loss function, the parameters of the main Q network are updated using the gradient descent method;

[0095] Every certain number of steps, the parameters of the main Q network are copied to the target Q network.

[0096] When the end of the charging robot arm approaches the charging port, the position of the robot arm is fine-tuned through feedback from the six-axis force sensor to compensate for the error in the initial visual positioning. If the sensed force does not match expectations (for example, there is an obstacle or a deviation in the position of the socket), the algorithm will make corresponding adjustments. During the process of inserting the charging gun, the sensor will monitor the relative position and force of the robot arm and the charging socket in real time.

[0097] Example 2

[0098] See also Figure 6 As shown, an embodiment of the present invention provides a device for realizing flexible control of an electric vehicle charging mechanical arm, including: the processing unit, storage unit, execution unit and acquisition unit are interconnected through a bus unit, and the storage unit can be used to store software programs, computer executable programs and modules, such as the software programs, computer executable programs and modules corresponding to the flexible plugging method of an electric vehicle charging mechanical arm in an embodiment of the present invention. The processing unit implements the above-mentioned flexible plugging method of an electric vehicle charging mechanical arm by running the software programs, computer executable programs and modules stored in the storage unit, including:

[0099] The point cloud at the charging interface of the vehicle is collected by the 3D camera of the acquisition unit, and the point cloud is subjected to noise reduction filtering and downsampling to obtain a sparse point cloud; the sparse point cloud is subjected to DBSCAN clustering to segment the point cloud clusters contained in the sparse point cloud, and the target point cloud cluster constituting the charging interface is identified according to the first composite feature of the point cloud cluster; the target transformation relationship of the target point cloud cluster relative to the charging interface model is obtained by using the corresponding matching relationship between the target point cloud cluster and the charging interface model, and then based on the transformation relationship and the target transformation relationship between the charging interface model coordinate system and the robot arm coordinate system, the 6D pose of the target point cloud cluster relative to the robot arm coordinate system is obtained;

[0100] The initial docking target position is determined based on the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system. The Dijkstra algorithm is used for path planning. A collision-free path is planned from the current position of the robot arm to the initial docking target position. The robot arm is moved to the vehicle charging port for docking according to the planned path.

[0101] During the process of plugging in the charging head, the camera of the acquisition unit collects the docking video of the charging plug and the vehicle charging interface, and the six-axis force sensor of the acquisition unit collects the force parameters of the robotic arm during the plugging process. The reinforcement learning model of the action and expected docking results is generated by fusion analysis of the docking video and force parameter feedback to control the execution unit plugging process and realize smooth plugging of the charging robotic arm.

[0102] Of course, the computer program stored in the storage unit of the flexible control device of the electric vehicle charging robot arm provided in the embodiment of the present invention is not limited to the method operation as described above, and can also execute the related operations in the flexible plugging method of the electric vehicle charging robot arm provided in any embodiment of the present invention.

[0103] Example 3

[0104] An embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. When the computer program is executed, the method for compliant plugging of an electric vehicle charging mechanical arm is implemented, including:

[0105] The point cloud at the car charging interface is collected by a 3D camera, and the point cloud is subjected to noise reduction filtering and downsampling to obtain a sparse point cloud; the sparse point cloud is clustered by DBSCAN to segment the point cloud clusters contained in the sparse point cloud, and the target point cloud clusters constituting the charging interface are identified according to the first composite feature of the point cloud cluster; the target transformation relationship of the target point cloud cluster relative to the charging interface model is obtained by using the corresponding matching relationship between the target point cloud cluster and the charging interface model, and then the 6D pose of the target point cloud cluster relative to the robotic arm coordinate system is obtained based on the transformation relationship and the target transformation relationship between the charging interface model coordinate system and the robotic arm coordinate system;

[0106] The initial docking target position is determined based on the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system. The Dijkstra algorithm is used for path planning. A collision-free path is planned from the current position of the robot arm to the initial docking target position. The robot arm is moved to the vehicle charging port for docking according to the planned path.

[0107] During the process of plugging the charging head, the camera is used to collect the docking video of the charging plug and the car's charging interface, and the six-axis force sensor is used to collect the force parameters of the robotic arm during the plugging process. The docking video and force parameter feedback are fused and analyzed to generate a reinforcement learning model of the action and expected docking results to control the plugging process and achieve smooth plugging of the charging robotic arm.

[0108] A computer-readable storage medium provided in an embodiment of the present invention stores a computer program which is not limited to the method operations described above, but can also execute related operations in a method for compliant plugging of an electric vehicle charging robot arm provided in any embodiment of the present invention.

[0109] In the embodiments provided by the present invention, it should be understood that the disclosed structures and methods can be implemented in other ways. For example, the structural embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, structures or units, which can be electrical, mechanical or other forms.

[0110] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0111] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0112] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A flexible plug-in method for an electric vehicle charging robot arm, characterized in that: include: The point cloud at the car charging port is collected by a 3D camera, and the point cloud is subjected to noise reduction filtering and downsampling to obtain a sparse point cloud; Perform DBSCAN clustering on the sparse point cloud to segment the point cloud clusters contained in the sparse point cloud, and identify the target point cloud cluster that constitutes the charging interface based on the first composite feature of the point cloud cluster; use the corresponding matching relationship between the target point cloud cluster and the second composite feature of the charging interface model to obtain the target transformation relationship of the target point cloud cluster relative to the charging interface model; based on the transformation relationship and the target transformation relationship between the charging interface model coordinate system and the robotic arm coordinate system, obtain the 6D pose of the target point cloud cluster relative to the robotic arm coordinate system; The initial docking target position is determined based on the 6D pose of the target point cloud clustering relative to the charging robot arm coordinate system. The Dijkstra algorithm is used for path planning. A collision-free path is planned from the current position of the robot arm to the initial docking target position. The robot arm is moved to the vehicle charging port for docking according to the planned path. During the process of plugging in the charging head, the camera is used to collect the docking video of the charging plug and the car's charging interface, and the six-axis force sensor is used to collect the force parameters of the robotic arm during the plugging process. The reinforcement learning model of the action and expected docking results is generated by fusion analysis of the docking video and force parameter feedback to control the action during the plugging process and realize smooth plugging of the charging robotic arm.

2. The flexible plug-in method of the electric vehicle charging robot arm according to claim 1, characterized in that: Perform DBSCAN clustering on the sparse point cloud to segment each point cloud cluster contained in the sparse point cloud, and identify the target point cloud cluster constituting the charging interface according to the first composite feature of the point cloud cluster, including: defining the minimum number of points included in each cluster, defining the neighborhood range of a point in each cluster; initializing all points in the sparse point cloud as unvisited points; Traverse the unvisited points. For each unvisited point, if the number of points in the neighborhood of the unvisited point is greater than the minimum number of included points, the unvisited point is a core point. Create a new cluster, recursively add the points in the direct and indirect neighborhoods to the created new cluster, and change the state to visited. Finally, output the point cloud clustering representing the structure. For each point cloud cluster, the principal component analysis method is used to extract shape features; the maximum range of the point cloud data of each point cloud cluster in three-dimensional space is calculated as the size feature; the surface features are calculated using the normal information and curvature of the point cloud data of the point cloud cluster; the shape features, size features and surface features of the point cloud cluster are used as the first composite features to match the corresponding first composite features of the charging interface model, and the target point cloud cluster that constitutes the charging interface is screened out from all point cloud clusters.

3. The flexible plug-in method of an electric vehicle charging robot arm according to claim 1, characterized in that: The target conversion relationship between the target point cloud cluster and the charging interface model is obtained by using the corresponding matching relationship between the target point cloud cluster and the second composite feature of the charging interface model, including: The charging interface model is sampled uniformly on a large scale to obtain several sampling points; For each sampling point, find the nearest neighbor to it; Calculate the distance between the sampling point and its nearest neighbor, the angle between the projection line of the sampling point and its nearest neighbor projected onto any two fixed orthogonal planes of the charging interface model coordinate system and the axis of the non-intersecting line in the corresponding two orthogonal planes, and the angle between the normal line of the sampling point and its nearest neighbor; construct the second composite feature between the sampling point and its nearest neighbor through the above parameters; For the target point cloud clustering constituting the charging interface, all possible sampling points and the nearest neighboring points of the sampling points are selected, and the second composite features of all possible sampling points and the nearest neighboring points of the sampling points are calculated. The corresponding sampling points and the nearest neighboring points of the sampling points in the charging interface model are matched and found from the hash table or storage tree of the second composite features of the sampling points and the nearest neighboring points of the sampling points in the charging interface model. For a number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, a transformation matrix between the two is calculated, so as to obtain a number of transformation matrices that define the position transformation relationship between the number of matching sampling points and the nearest neighboring points of the sampling points in the target point cloud cluster and the charging interface model, wherein the transformation matrix includes a rotation matrix and a translation matrix; The rotation angles and translation distances corresponding to all transformation matrices are used to vote for the rotation angles and translation distances within the set posture space range, and the voting results of the rotation angles and translation distances within the set posture space range are clustered by meanshift. The rotation angles and translation distances corresponding to the main cluster with a large number of votes are selected to obtain the target transformation matrix of the target point cloud cluster relative to the charging interface model.

4. The flexible plug-in method of the electric vehicle charging robot arm according to claim 3 is characterized in that: A hash table or a storage tree is constructed using the second composite feature; the hash table uses the second composite feature as a key and the sampling point and the nearest neighbor of the sampling point of the corresponding feature on the charging interface model as a value; the storage tree includes a five-layer structure, wherein the root node of the first layer stores the distance between the sampling point and the nearest neighbor of the sampling point, the second layer tree nodes and the third layer tree nodes store the two angles between the projection line after the sampling point and the nearest neighbor line of the sampling point are projected onto any two fixed orthogonal planes of the charging interface model coordinate system and the axes of the non-intersecting lines in the corresponding two orthogonal planes, the fourth layer tree nodes store the angle between the sampling point and the normal of the nearest neighbor of the sampling point, and the fifth layer leaf nodes store the sampling point and the nearest neighbor of the second composite feature stored in the tree node where the corresponding leaf node is located and the root node.

5. The flexible plug-in method of the electric vehicle charging robot arm according to claim 1, characterized in that: The reinforcement learning model includes: a docking image feature encoder based on a convolutional network, a Softmax layer, a fully connected network, a weight generation network and a docking image feature decoder based on a transposed convolutional network, wherein the docking image feature encoder receives a current docking video frame sequence [P1, P2....P n ], the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points and the force parameters corresponding to each docking video frame in the current docking video frame sequence, the distance moved by the robot arm toward the vehicle charging interface, the action and the type encoding of the corresponding charging interface type are input into the fully connected network, and the fully connected network obtains the predicted next docking video frame The weight generation network generates n groups of weights based on the attention points corresponding to each docking video frame in the next docking video frame sequence, and each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted results are input into the docking image feature decoder to predict each docking video frame in the next docking video frame sequence; the action predicted by the fully connected network is provided to the charging robot arm for execution; wherein the type code is used as a condition to enable the fully connected network to perform action control for the corresponding type of charging interface.

6. The flexible plug-in method of the electric vehicle charging robot arm according to claim 5, characterized in that: The fully connected network is trained by the DDQN algorithm, and two fully connected networks are initialized as the main Q network and the target Q network respectively; The following process is repeated until the algorithm converges or the set number of training rounds is reached: According to the current state s, the ε-greedy strategy is used to select action a; the current state s includes: the attention point in the docking video frame, the force parameters corresponding to the docking video frame, the distance and type code of the robot arm moving toward the car charging port; the ε-greedy strategy refers to randomly selecting an action with a certain probability ε, and selecting the action that maximizes the Q value with a probability of 1-ε; the action a includes the index of the moving direction and the index of the moving distance, where the moving direction includes 8 directions: up, down, left, right, upper left, lower left, upper right, and lower right; Execute action a and observe the reward r and new state s' of the feedback of the charging robot's moving action; The experience you will gain<s,a,r,s'> Stored in the experience pool; Randomly extract a batch of experience data from the experience pool for training; For each sample<s,a,r,s'> , calculate the target Q value: Use the main Q network to calculate the Q values ​​of all possible actions in state s'; Use the target Q network to calculate the maximum Q value corresponding to the action selected by the main Q network in state s'; The target Q value calculation formula is: Q target =r+γ*max(Q target (s', a')), where a' is the value that makes Q main (s', a') is the largest action, γ is the discount factor, and r is the reward: Among them, R is the reward score constant, and L is the average distance threshold; Calculate the loss function; Based on the loss obtained from the loss function, the parameters of the main Q network are updated using the gradient descent method; Every certain number of steps, the parameters of the main Q network are copied to the target Q network.

7. The flexible plug-in method of the electric vehicle charging robot arm according to claim 6, characterized in that: The loss function is as follows: LOSS = αLoss1 + βLoss2 + κLoss3; where α, β, and κ are the corresponding weight coefficients, and Loss1 represents the predicted next docking video frame. And the next real docking video frame P n+1 Loss2 represents the mean Euclidean distance between the predicted attention points and the true attention points, and Loss3 represents the mean square error of the Q value calculated by the main Q network and the target Q network.

8. The method for flexible plugging of an electric vehicle charging robot arm according to claim 5, characterized in that: In the reinforcement learning model, the docking image feature encoder, the Softmax layer, the weight generation network and the docking image feature decoder based on the transposed convolutional network follow the following constraints: the docking image feature encoder cooperates with the Softmax layer to extract the attention points in each docking video frame in the current docking video frame sequence, and the extracted attention points are input into the weight generation network to generate n groups of weights, each group of weights is weighted with the corresponding docking video frame features extracted by the docking image feature encoder, and the weighted results are input into the docking image feature decoder to restore each docking video frame in the current docking video frame sequence.

9. A compliant control device for an electric vehicle charging robot arm, characterized in that: include: At least one processing unit, the processing unit, storage unit, execution unit and acquisition unit are interconnected through a bus unit, the storage unit stores a computer program, and when the computer program is executed by the processing unit, the flexible plug-in method of the electric vehicle charging robot arm as described in any one of claims 1-8 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, the flexible plug-in method of the electric vehicle charging robot arm as described in any one of claims 1-8 is implemented.

Citation Information

Cited By

  • Method and system for recognizing and grabbing wire harness by robot based on point cloud clustering

    CN120791808A

  • A method and system for robot identification and grasping wire harness based on point cloud clustering

    CN120791808B

  • Two-section type plug-in method for charging mechanical arm of electric vehicle

    CN121062515A