Dual-arm robot motion planning method and system for complex manipulation tasks

Abstract

The present invention discloses a dual-arm robot motion planning method and system for complex operation tasks. The method decouples the dual-arm robot collision detection into dual-arm self-collision detection and environmental obstacle detection, constructs a collision distance learning framework for high-degree-of-freedom dual-arm robots, uses a paired collision distance learning model to achieve high-precision self-collision detection of the dual-arm robot, and uses a collision distance inference learning model to achieve collision detection between the dual-arm robot and environmental obstacles. On the basis of maintaining the high-precision self-collision detection accuracy of the dual-arm robot, the system is enabled to have the ability to handle dynamic obstacles; designs a batch parallel expansion mechanism based on the artificial potential field method to achieve efficient path search of the dual-arm robot in complex environments; constructs a path replanning module based on the collision distance learning framework to monitor and update environmental obstacle change information in real time. The dual-arm robot can adaptively adjust the motion path, ensuring that the dual-arm robot can adapt to complex operation tasks.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot motion planning based on learning models, and in particular relates to a dual-arm robot motion planning method and system for complex operation tasks. Background Art

[0002] Dual-arm robotic systems with kinematic redundancy are increasingly becoming a key transformative technology for complex manipulation tasks, such as aerospace component assembly and sterile drug dispensing. While dual-arm systems offer greater flexibility and coordination, their high degree of freedom also presents significant motion planning challenges. Currently, motion planning for dual-arm robots typically requires over 20 seconds to perform collaborative tasks of varying complexity, and in some extreme scenarios, even over 100 seconds, severely limiting their ability to respond in real time in dynamic environments. Compared to single-arm robotic systems, dual-arm systems face three core challenges in efficient motion planning: 1) The significant increase in the configuration space dimension increases the complexity of the feasible path search process; 3) The left and right robotic arms share a portion of the working area, increasing the risk of self-collision. The complex mechanical structure makes collision detection more frequent and the computational burden heavier, further slowing trajectory generation efficiency. To address these issues, existing research mainly optimizes along two paths: one is to design a parallel path search mechanism for high-dimensional configuration space to improve the search efficiency of collision-free paths; the other is to build a high-performance collision detection model based on the characteristics of the dual-arm structure, thereby significantly improving the overall planning speed and system responsiveness.

[0003] Researchers are currently attempting to integrate learning-based collision detection models into traditional planning algorithms to improve the efficiency of robot motion planning. However, learning-based methods face the following challenges when applied to dual-arm robot systems: 1) The collision distance function in the high-dimensional dual-arm configuration space of dual-arm motion exhibits a complex, highly non-convex structure, making it difficult for existing learning-based models to accurately infer the self-collision state of the dual arms. This reduces the reliability of the estimation and ultimately the efficiency of motion planning. 2) Most learning-based motion planning methods are designed for static environments, where any changes in the obstacle configuration render the pre-trained collision model invalid or require retraining.

[0004] Therefore, in response to the above problems and challenges, the present invention proposes a dual-arm robot motion planning method and system for complex operation tasks. Summary of the Invention

[0005] In response to the above technical problems, the present invention provides a dual-arm robot motion planning method and system for complex operation tasks.

[0006] The technical solution adopted by the present invention to solve the technical problem is:

[0007] A dual-arm robot motion planning method for complex manipulation tasks comprises the following steps:

[0008] S100: Build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. Detect the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and an obstacle. Obtain the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and the obstacle.

[0009] S200: Initialize the path search tree and set the start and end joint configurations. Generate joint configuration points through batch random sampling. In each iteration, select a strategy based on the generated random probability and the value of the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as a candidate expansion node; otherwise, a candidate expansion node is generated based on the artificial potential field method.

[0010] S300: Match the nearest neighbor tree node for each candidate extension node to form an extension edge. Through equidistant interpolation discretization processing, the interpolation point is input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each extension edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends;

[0011] S400: The acquired dual-arm robot path is sent to the dual-arm robot for execution. During the movement, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurred to replan and perform real-time motion path repair to ensure that when the environmental obstacles change, the dual-arm robot adaptively adjusts the motion path.

[0012] Preferably, the pre-trained paired collision distance learning model in S100 detects the collision distance between the left robotic arm and the right robotic arm, including:

[0013] S110: Set any set of dual-arm robot joint angles Input to the pairwise collision distance learning model, output Paired collision distance between different links of the left and right arms ,in, , Represents the left and right arms Connecting rods, , are all integers;

[0014] S120: From The minimum distance is selected from the paired collision distances between the different links of the left and right arms of the group as the left robotic arm and right robotic arm The collision distance between:

[0015] .

[0016] Preferably, the paired collision distance learning model pre-trained in S100 includes five parts. The first part is to configure the dual-arm robot joints As the initial input, and use the position encoding function Encode the joint angles of the dual-arm robot and increase the frequency of the input values, where yes dimensional matrix, the output of the first part will be used as the input of the second part;

[0017] The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part.

[0018] The third part includes a fully connected layer, an activation function Sigmoid, and a DropOut layer. The output of the third part will serve as the input of the fourth part.

[0019] The fourth part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the fourth part will serve as the input of the fifth part.

[0020] The fifth part includes the fully connected layer, and the output features are Paired collision distance between different links of the left and right arms The learning mapping of .

[0021] Preferably, the pre-trained collision distance inference learning model in S100 detects the collision distance between the dual-arm robot body and the obstacle, including:

[0022] S130: The task operation space of the dual-arm robot Split into Cube subspace unit , , set any set of dual-arm robot joint angles Input to the pre-trained collision distance inference learning model, output Collision distance between the dual-arm robot and the subspace unit , ;

[0023] S140: Using a depth camera to detect obstacles in the dual-arm robot's task space right The subspace unit occupies the information matrix ,in, for dimensional matrix, the corresponding element value of the occupied subspace unit in the matrix is ​​1, otherwise it is 0. Assume that there are subspace units are occupied;

[0024] S150: For the above Collision distance between the dual-arm robot and the subspace unit , If a subspace unit is found to be not occupied by an obstacle, the corresponding collision distance is eliminated. , and finally get The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance , ;

[0025] S160: From The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance Select the minimum distance value as the distance between the dual-arm robot and the obstacle The collision distance between them.

[0026] Preferably, the pre-trained collision distance reasoning learning model includes four parts. The first part is to calculate the joint angle of the dual-arm robot. As the initial input, the first part includes a fully connected layer, an activation function ReLU, and a DropOut layer, where yes dimensional matrix, the output of the first part will be used as the input of the second part;

[0027] The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part.

[0028] The third part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the third part will serve as the input of the fourth part.

[0029] The fourth part includes the fully connected layer, and the output features are Collision distance between the dual-arm robot and the subspace unit , .

[0030] Preferably, S200 includes:

[0031] S210: Define the starting joint configuration of the dual-arm robot as , the target joint configuration is , initialize a path search tree , and join in ;

[0032] S220: In the joint space of the dual-arm robot In-batch random sampling Group joint configuration points , for the above Group configuration points, find the path tree Each sampling point at the middle distance Recent Path nodes ;

[0033] S230: Generate a random probability ,if If the probability is greater than the preset threshold, the random sampling Group joint configuration points as of Candidate expansion nodes , otherwise, based on the artificial potential field method Batch Generation candidate expansion nodes.

[0034] Preferably, in S230, the artificial potential field method is Batch Generation candidate expansion nodes, specifically:

[0035] S231: Targeting middle Path nodes , at each node radius Randomly generated exploration points , , a total of exploration points, batch them into the dual-arm robot collision distance learning framework, and filter out the exploration points that collide around each exploration point , , ,in, is the number of exploration points where collisions occur;

[0036] S232: Based on the artificial potential field method Determine the expansion direction and target joint configuration in the dual-arm robot joint space right Gravity ,obstacle In the joint space repulsion = , the combined force is , , are the attraction coefficient and repulsion coefficient respectively;

[0037] S233: For nodes , its corresponding candidate expansion node for , repeat the above process times, a total of Candidate expansion nodes .

[0038] Preferably, S300 includes:

[0039] S310: Find the path tree Each sampling point at the middle distance Recent Path nodes , will obtain candidate expansion nodes and corresponding in path nodes, forming Candidate extension edges , ,…, ;

[0040] S320: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , a total of interpolation points , batch input L interpolation points into the dual-arm robot collision distance learning framework, and obtain Collision estimation results of interpolation points;

[0041] S330: Select the last non-collision interpolation point in each candidate extension edge interpolation point as a new path extension node, and obtain a total of A new extension node is added to the path search tree. ,in, ;

[0042] S340: When the following conditions are met, the dual-arm robot path planning process ends, otherwise the above sampling and expansion process continues to be repeated: 1) The dual-arm robot path planning time reaches the upper limit; 2) The latest expansion node satisfy , which indicates that a feasible dual-arm robot path has been successfully found , , preset by the user.

[0043] Preferably, S400 includes:

[0044] Obtained dual-arm robot path Send it to the dual-arm robot for execution. During the movement of the dual-arm robot, the depth camera is used to Detecting the task operation space within the operation space Obstacle location information and update the obstacle right The subspace unit occupies the information matrix , while the dual-arm robot collision distance learning framework detects the dual-arm robot path in real time Whether there is a collision with an obstacle;

[0045] If a collision occurs, the collision local path , execute the path planning process of S200 and S300 again, and perform real-time motion path repair to ensure that the dual-arm robot can adaptively adjust the motion path when the environmental obstacles change.

[0046] A dual-arm robot motion planning system for complex manipulation tasks, including a dual-arm robot collision distance learning framework building module, a candidate expansion node generation module, a path planning module, and an adaptive motion path adjustment module;

[0047] A dual-arm robot collision distance learning framework building module is used to build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. The module detects the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles, and obtains the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles.

[0048] The candidate expansion node generation module is used to initialize the path search tree and set the start and end joint configurations. Joint configuration points are generated through batch random sampling. At each iteration, a strategy is selected based on the size of the generated random probability and the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as the candidate expansion node; otherwise, the candidate expansion node is generated based on the artificial potential field method;

[0049] The path planning module is used to match each candidate expansion node with the nearest neighbor tree node to form an expansion edge. Through equidistant interpolation discretization processing, the interpolation points are input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each expansion edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends;

[0050] The adaptive motion path adjustment module is used to send the acquired dual-arm robot path to the dual-arm robot for execution. During the movement process, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurs to replan and perform real-time motion path repair to ensure that the dual-arm robot adaptively adjusts the motion path when the environmental obstacles change.

[0051] The above-mentioned dual-arm robot motion planning method and system for complex operation tasks, by constructing a collision distance learning framework for high-degree-of-freedom dual-arm robots, enables the dual-arm robots to have the ability to handle dynamic obstacles while maintaining the high-precision self-collision detection accuracy of the dual-arm robots. Secondly, a batch parallel expansion mechanism based on the artificial potential field method is designed to realize efficient path search of the dual-arm robots in complex environments, maintaining the advantages of random sampling and the guidance of the artificial potential field method. In the path search process, the dual-arm robot collision distance learning framework is used for batch collision detection, thereby realizing parallel path exploration of the dual-arm robots and further improving the efficiency of motion planning; by real-time monitoring and updating of environmental obstacle change information, the dual-arm robot's collision distance learning framework is used to quickly re-plan feasible paths, ensuring that the dual-arm robots can adapt to complex operation tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Flowchart of a dual-arm robot motion planning method for complex operation tasks in one embodiment of the present invention;

[0053] Figure 2 Schematic diagram of a collision distance learning framework for a high-degree-of-freedom dual-arm robot in one embodiment of the present invention. DETAILED DESCRIPTION

[0054] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings.

[0055] In one embodiment, Figure 1 As shown, a dual-arm robot motion planning method for complex operation tasks includes the following steps:

[0056] S100: Build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. Detect the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and an obstacle. Obtain the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and the obstacle.

[0057] S200: Initialize the path search tree and set the start and end joint configurations. Generate joint configuration points through batch random sampling. In each iteration, select a strategy based on the generated random probability and the value of the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as a candidate expansion node; otherwise, a candidate expansion node is generated based on the artificial potential field method.

[0058] S300: Match the nearest neighbor tree node for each candidate extension node to form an extension edge. Through equidistant interpolation discretization processing, the interpolation point is input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each extension edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends;

[0059] S400: The acquired dual-arm robot path is sent to the dual-arm robot for execution. During the movement, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurred to replan and perform real-time motion path repair to ensure that when the environmental obstacles change, the dual-arm robot adaptively adjusts the motion path.

[0060] Specifically, the dual-arm robot In terms of hardware structure, the left robotic arm and right robotic arm composition, accordingly, its joint space From the left arm joint space and right arm joint space Joint composition: ,in , , all by The present invention will directly in the joint space of the dual-arm robot Compared with the Cartesian space motion planning method, its advantage is that it avoids complex inverse kinematics solution.

[0061] The present invention decouples dual-arm robot collision detection into dual-arm self-collision detection and environmental obstacle collision detection, and quantifies it into collision distance detection. The dual-arm robot must meet two conditions for collision-free operation:

[0062] 1) ;

[0063] 2) ;

[0064] in Indicates the left robotic arm and right robotic arm The collision distance between Represents the dual-arm robot body and obstacles The collision distance between them.

[0065] Based on the above information, the present invention designs a collision distance learning framework for a dual-arm robot. Specifically, it comprises two sub-learning models: 1) a pairwise collision distance learning model for high-precision self-collision detection of the dual-arm robot; and 2) a collision distance inference learning model for collision detection between the dual-arm robot and environmental obstacles. This collision distance learning framework maintains the high accuracy of the dual-arm robot's self-collision detection while also enabling it to handle dynamic obstacles.

[0066] In one embodiment, Figure 2 As shown, the pre-trained paired collision distance learning model in S100 detects the collision distance between the left and right robotic arms, including:

[0067] S110: Set any set of dual-arm robot joint angles Input to the pairwise collision distance learning model, output Paired collision distance between different links of the left and right arms ,in, , Represents the left and right arms Connecting rods, , are all integers;

[0068] S120: From The minimum distance is selected from the paired collision distances between the different links of the left and right arms of the group as the left robotic arm and right robotic arm The collision distance between:

[0069] .

[0070] In one embodiment, Figure 2 As shown in the figure, the pre-trained paired collision distance learning model in S100 consists of five parts. The first part is to configure the dual-arm robot joints As the initial input, and use the position encoding function Encode the joint angles of the dual-arm robot and increase the frequency of the input values, where yes dimensional matrix, the output of the first part will be used as the input of the second part;

[0071] The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part.

[0072] The third part includes a fully connected layer, an activation function Sigmoid, and a DropOut layer. The output of the third part will serve as the input of the fourth part.

[0073] The fourth part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the fourth part will serve as the input of the fifth part.

[0074] The fifth part includes the fully connected layer, and the output features are Paired collision distance between different links of the left and right arms The learning mapping of .

[0075] In one embodiment, Figure 2 As shown, the pre-trained collision distance inference learning model in S100 detects the collision distance between the dual-arm robot body and the obstacle, including:

[0076] S130: In order to enable the dual-arm robot to have the ability to handle collision detection of unseen or dynamic obstacles, the task operation space of the dual-arm robot is (Note that the task operation space The size is much smaller than the reachable space of the entire dual-arm robot ) is divided into Cube subspace unit , , set any set of dual-arm robot joint angles Input to the pre-trained collision distance inference learning model, output Collision distance between the dual-arm robot and the subspace unit , ; where the side length of each cube subspace unit is centimeter, ;

[0077] S140: Using a depth camera to detect obstacles in the dual-arm robot's task space right The subspace unit occupies the information matrix ,in, for dimensional matrix, the corresponding element value of the occupied subspace unit in the matrix is ​​1, otherwise it is 0. Assume that there are subspace units are occupied;

[0078] S150: For the above Collision distance between the dual-arm robot and the subspace unit , If a subspace unit is found to be not occupied by an obstacle, the corresponding collision distance is eliminated. , and finally get The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance , ;

[0079] S160: From The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance Select the minimum distance value as the distance between the dual-arm robot and the obstacle The collision distance between them.

[0080] In one embodiment, Figure 2 As shown in the figure, the pre-trained collision distance reasoning learning model consists of four parts. The first part is to calculate the joint angle of the dual-arm robot. As the initial input, the first part includes a fully connected layer, an activation function ReLU, and a DropOut layer, where yes dimensional matrix, the output of the first part will be used as the input of the second part;

[0081] The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part.

[0082] The third part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the third part will serve as the input of the fourth part.

[0083] The fourth part includes the fully connected layer, and the output features are Collision distance between the dual-arm robot and the subspace unit , .

[0084] Furthermore, for any set of dual-arm robot joint configurations , when it is input into the collision distance learning framework composed of the above two learning models, if and only if 1) ;2) When both conditions are met, the joint configuration of the dual-arm robot Only when the joint configuration is non-collision-free, it is considered as a collision-free joint configuration; otherwise, it is considered as a collision-free joint configuration. Note that since the collision distance learning framework consists of two neural network models, it has batch data processing capabilities and can process collision detection tasks for multiple joint configurations in parallel at one time.

[0085] Based on the aforementioned collision distance learning framework, a batch parallel expansion mechanism based on the artificial potential field method is designed to achieve efficient path search for dual-arm robots in complex environments.

[0086] In one embodiment, S200 includes:

[0087] S210: Define the starting joint configuration of the dual-arm robot as , the target joint configuration is , initialize a path search tree , and join in ;

[0088] S220: In the joint space of the dual-arm robot In-batch random sampling Group joint configuration points , for the above Group configuration points, find the path tree Each sampling point at the middle distance Recent Path nodes ;

[0089] S230: Generate a random probability ,if If the probability is greater than the preset threshold, the random sampling Group joint configuration points as of Candidate expansion nodes , otherwise, based on the artificial potential field method Batch Generation candidate expansion nodes.

[0090] In one embodiment, S230 is based on the artificial potential field method. Batch Generation candidate expansion nodes, specifically:

[0091] S231: Targeting middle Path nodes , at each node radius Randomly generated exploration points , , a total of exploration points, batch them into the dual-arm robot collision distance learning framework, and filter out the exploration points that collide around each exploration point , , ,in, is the number of exploration points where collisions occur;

[0092] S232: Based on the artificial potential field method Determine the expansion direction and target joint configuration in the dual-arm robot joint space right Gravity ,obstacle In the joint space repulsion = , the combined force is , , are the attraction coefficient and repulsion coefficient respectively;

[0093] S233: For nodes , its corresponding candidate expansion node for , repeat the above process times, a total of Candidate expansion nodes .

[0094] In one embodiment, S300 includes:

[0095] S310: The obtained candidate expansion nodes and corresponding in path nodes, forming Candidate extension edges , ,…, ;

[0096] S320: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , a total of interpolation points , batch input L interpolation points into the dual-arm robot collision distance learning framework, and obtain Collision estimation results of interpolation points;

[0097] S330: Select the last non-collision interpolation point in each candidate extension edge interpolation point as a new path extension node, and obtain a total of A new extension node is added to the path search tree. ,in, ;

[0098] S340: When the following conditions are met, the dual-arm robot path planning process ends, otherwise the above sampling and expansion process continues to be repeated: 1) The dual-arm robot path planning time reaches the upper limit; 2) The latest expansion node satisfy , which indicates that a feasible dual-arm robot path has been successfully found , , preset by the user.

[0099] Furthermore, a path replanning module based on the collision distance learning framework is constructed, and the dual-arm robot can adaptively adjust its motion path when environmental obstacles change.

[0100] In one embodiment, S400 includes:

[0101] Obtained dual-arm robot path Send it to the dual-arm robot for execution. During the movement of the dual-arm robot, the depth camera is used to (generally Less than 1 second) Detect task operation space within the operation space Obstacle location information and update the obstacle right The subspace unit occupies the information matrix , while the dual-arm robot collision distance learning framework detects the dual-arm robot path in real time Whether there is a collision with an obstacle;

[0102] If a collision occurs, the collision local path , execute the path planning process of S200 and S300 again, and perform real-time motion path repair to ensure that the dual-arm robot can adaptively adjust the motion path when the environmental obstacles change.

[0103] The above dual-arm robot motion planning method for complex operation tasks has the following beneficial effects:

[0104] (1) The present invention innovatively constructs a collision distance learning framework for high-degree-of-freedom dual-arm robots, decoupling dual-arm robot collision detection into dual-arm self-collision detection and environmental obstacle detection. Specifically, it includes two sub-learning models: a paired collision distance learning model is used to achieve high-precision self-collision detection of the dual-arm robot. By further decomposing the left and right arm collision detection into collision distance detection between different links, the learning difficulty is reduced, solving the problem that the existing learning-based collision detection method is difficult to accurately infer the highly non-convex dual-arm self-collision state; secondly, a collision distance inference learning model is used to achieve collision detection between the dual-arm robot and environmental obstacles. By segmenting the dual-arm robot's task operation space and obtaining the obstacle occupancy matrix, the dual-arm robot can detect unseen or dynamic obstacles. The above-mentioned collision distance learning framework for high-degree-of-freedom dual-arm robots enables the dual-arm robot to handle dynamic obstacles while maintaining its high-precision self-collision detection accuracy.

[0105] (2) Compared with the existing methods that lack parallel expansion and guided expansion, this invention designs a batch parallel expansion mechanism based on the artificial potential field method on the basis of the dual-arm robot collision distance learning framework, realizing efficient path search for the dual-arm robot in complex environments, maintaining the advantages of random sampling and the guided nature of the artificial potential field method. In addition, the dual-arm robot collision distance learning framework is used to perform batch collision detection during the path search process, thereby realizing parallel path exploration for the dual-arm robot and further improving the efficiency of motion planning.

[0106] (3) In the part of the batch parallel expansion mechanism based on the artificial potential field method, the present invention constructs a repulsive field for obstacles by generating exploration points around the path nodes and screening the exploration points where collisions occur, thus solving the problem that obstacles cannot be effectively represented in the joint configuration space of the dual-arm robot.

[0107] (4) The present invention provides a path replanning module based on a collision distance learning framework. By monitoring and updating the information of environmental obstacle changes in real time, the dual-arm robot uses the collision distance learning framework to quickly replan feasible paths, ensuring that the dual-arm robot can adapt to complex operation tasks.

[0108] In one embodiment, a dual-arm robot motion planning system for complex operation tasks is also provided, including a dual-arm robot collision distance learning framework building module, a candidate expansion node generation module, a path planning module, and an adaptive motion path adjustment module;

[0109] A dual-arm robot collision distance learning framework building module is used to build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. The module detects the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles, and obtains the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles.

[0110] The candidate expansion node generation module is used to initialize the path search tree and set the start and end joint configurations. Joint configuration points are generated through batch random sampling. At each iteration, a strategy is selected based on the size of the generated random probability and the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as the candidate expansion node; otherwise, the candidate expansion node is generated based on the artificial potential field method;

[0111] The path planning module is used to match each candidate expansion node with the nearest neighbor tree node to form an expansion edge. Through equidistant interpolation discretization processing, the interpolation points are input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each expansion edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends;

[0112] The adaptive motion path adjustment module is used to send the acquired dual-arm robot path to the dual-arm robot for execution. During the movement process, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurs to replan and perform real-time motion path repair to ensure that the dual-arm robot adaptively adjusts the motion path when the environmental obstacles change.

[0113] Regarding the specific limitations of the dual-arm robot motion planning system for complex operation tasks, please refer to the limitations of the dual-arm robot motion planning method for complex operation tasks mentioned above, which will not be repeated here. The various modules in the above-mentioned dual-arm robot motion planning system for complex operation tasks can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0114] The above is a detailed introduction to the dual-arm robot motion planning method and system for complex operation tasks provided by the present invention. This article uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, the present invention can also be improved and modified, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A dual-arm robot motion planning method for complex manipulation tasks, characterized by: The method comprises the following steps: S100: Build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. Detect the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and an obstacle. Obtain the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms, and the collision distance between the dual-arm robot body and the obstacle. S200: Initialize the path search tree and set the start and end joint configurations. Generate joint configuration points through batch random sampling. In each iteration, select a strategy based on the generated random probability and the value of the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as a candidate expansion node; otherwise, a candidate expansion node is generated based on the artificial potential field method. S300: Match the nearest neighbor tree node for each candidate extension node to form an extension edge. Through equidistant interpolation discretization processing, the interpolation point is input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each extension edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends; S400: The acquired dual-arm robot path is sent to the dual-arm robot for execution. During the movement, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurred to replan and perform real-time motion path repair to ensure that when the environmental obstacles change, the dual-arm robot adaptively adjusts the motion path.

2. The method according to claim 1, characterized in that The pre-trained paired collision distance learning model in S100 detects the collision distance between the left and right robotic arms, including: S110: Set any set of dual-arm robot joint angles Input to the pairwise collision distance learning model, output Paired collision distance between different links of the left and right arms ,in, , Represents the left and right arms Connecting rods, , are all integers; S120: From The minimum distance is selected from the paired collision distances between the different links of the left and right arms of the group as the left robotic arm and right robotic arm The collision distance between: 。 3. The method according to claim 2, characterized in that The pre-trained paired collision distance learning model in S100 consists of five parts. The first part is to configure the dual-arm robot joints As the initial input, and use the position encoding function Encode the joint angles of the dual-arm robot and increase the frequency of the input values, where yes dimensional matrix, the output of the first part will be used as the input of the second part; The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part. The third part includes a fully connected layer, an activation function Sigmoid, and a DropOut layer. The output of the third part will serve as the input of the fourth part. The fourth part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the fourth part will serve as the input of the fifth part. The fifth part includes the fully connected layer, and the output features are Paired collision distance between different links of the left and right arms The learning mapping of .

4. The method according to claim 3, characterized in that The pre-trained collision distance inference learning model in S100 detects the collision distance between the dual-arm robot body and obstacles, including: S130: The task operation space of the dual-arm robot Split into Cube subspace unit , , set any set of dual-arm robot joint angles Input to the pre-trained collision distance inference learning model, output Collision distance between the dual-arm robot and the subspace unit , ; S140: Using a depth camera to detect obstacles in the dual-arm robot's task space right The subspace unit occupies the information matrix ,in, for dimensional matrix, the corresponding element value of the occupied subspace unit in the matrix is ​​1, otherwise it is 0. Assume that there are subspace units are occupied; S150: For the above Collision distance between the dual-arm robot and the subspace unit , If a subspace unit is found to be not occupied by an obstacle, the corresponding collision distance is eliminated. , and finally get The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance , ; S160: From The subspace unit occupied by the group of obstacles With dual-arm robot Collision distance Select the minimum distance value as the distance between the dual-arm robot and the obstacle The collision distance between them.

5. The method according to claim 4, characterized in that The pre-trained collision distance inference learning model consists of four parts. The first part is to calculate the joint angles of the dual-arm robot. As the initial input, the first part includes a fully connected layer, an activation function ReLU, and a DropOut layer, where yes dimensional matrix, the output of the first part will be used as the input of the second part; The second part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the second part will serve as the input of the third part. The third part includes a fully connected layer, an activation function ReLU, and a DropOut layer. The output of the third part will serve as the input of the fourth part. The fourth part includes the fully connected layer, and the output features are Collision distance between the dual-arm robot and the subspace unit , .

6. The method according to claim 5, characterized in that S200 includes: S210: Define the starting joint configuration of the dual-arm robot as , the target joint configuration is , initialize a path search tree , and join in ; S220: In the joint space of the dual-arm robot In-batch random sampling Group joint configuration points , for the above Group configuration points, find the path tree Each sampling point at the middle distance Recent Path nodes ; S230: Generate a random probability ,if If the probability is greater than the preset threshold, the random sampling Group joint configuration points as of Candidate expansion nodes , otherwise, based on the artificial potential field method Batch Generation candidate expansion nodes.

7. The method according to claim 6, characterized in that In S230, the artificial potential field method is used to Batch Generation candidate expansion nodes, specifically: S231: Targeting middle Path nodes , at each node radius Randomly generated exploration points , , a total of exploration points, batch them into the dual-arm robot collision distance learning framework, and filter out the exploration points that collide around each exploration point , , ,in, is the number of exploration points where collisions occur; S232: Based on the artificial potential field method Determine the expansion direction and target joint configuration in the dual-arm robot joint space right Gravity ,obstacle In the joint space repulsion = , the combined force is , , are the attraction coefficient and repulsion coefficient respectively; S233: For nodes , its corresponding candidate expansion node for , repeat the above process times, a total of Candidate expansion nodes .

8. The method according to claim 7, characterized in that S300 includes: S310: Find the path tree Each sampling point at the middle distance Recent Path nodes , will obtain candidate expansion nodes and corresponding in path nodes, forming Candidate extension edges , ,…, ; S320: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , a total of interpolation points , batch input L interpolation points into the dual-arm robot collision distance learning framework, and obtain Collision estimation results of interpolation points; S330: Select the last non-collision interpolation point in each candidate extension edge interpolation point as a new path extension node, and obtain a total of A new extension node is added to the path search tree. ,in, ; S340: When the following conditions are met, the dual-arm robot path planning process ends, otherwise the above sampling and expansion process continues to be repeated: 1) The dual-arm robot path planning time reaches the upper limit; 2) The latest expansion node satisfy , which indicates that a feasible dual-arm robot path has been successfully found , , preset by the user.

9. The method according to claim 8, characterized in that S400 includes: Obtained dual-arm robot path Send it to the dual-arm robot for execution. During the movement of the dual-arm robot, the depth camera is used to Detecting the task operation space within the operation space Obstacle location information and update the obstacle right The subspace unit occupies the information matrix , while the dual-arm robot collision distance learning framework detects the dual-arm robot path in real time Whether there is a collision with an obstacle; If a collision occurs, the collision local path , execute the path planning process of S200 and S300 again, and perform real-time motion path repair to ensure that the dual-arm robot can adaptively adjust the motion path when the environmental obstacles change.

10. A dual-arm robot motion planning system for complex manipulation tasks, characterized by: It includes a dual-arm robot collision distance learning framework building module, a candidate expansion node generation module, a path planning module, and an adaptive motion path adjustment module; A dual-arm robot collision distance learning framework building module is used to build a dual-arm robot collision distance learning framework, including a pre-trained paired collision distance learning model and a pre-trained collision distance inference learning model. The module detects the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles, and obtains the dual-arm robot system collision detection results based on the collision distance between the left and right robotic arms and the collision distance between the dual-arm robot body and obstacles. The candidate expansion node generation module is used to initialize the path search tree and set the start and end joint configurations. Joint configuration points are generated through batch random sampling. At each iteration, a strategy is selected based on the size of the generated random probability and the preset probability threshold: if the probability exceeds the threshold, the joint configuration point is directly used as the candidate expansion node; otherwise, the candidate expansion node is generated based on the artificial potential field method; The path planning module is used to match each candidate expansion node with the nearest neighbor tree node to form an expansion edge. Through equidistant interpolation discretization processing, the interpolation points are input into the dual-arm robot collision distance learning framework for batch safety verification. The last collision-free point at the end of each expansion edge is selected as a new node to be added to the search tree. When the preset end condition is met, the dual-arm robot path planning process ends; The adaptive motion path adjustment module is used to send the acquired dual-arm robot path to the dual-arm robot for execution. During the movement process, the dual-arm robot collision distance learning framework is used to detect in real time whether the dual-arm robot path collides with obstacles. If a collision occurs, S200 and S300 are executed on the local path where the collision occurs to replan and perform real-time motion path repair to ensure that the dual-arm robot adaptively adjusts the motion path when the environmental obstacles change.

Images