Efficient and safe motion planning method and system for space manipulators in aerospace scenarios

Abstract

The present invention discloses a method and system for efficient and safe motion planning of a space manipulator for aerospace scenarios. The method adopts multiple depth cameras to perceive multi-view point cloud information, and fuses the complete occupancy Boolean matrix of obstacles in the space reachable by the space manipulator. Based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the space manipulator is constructed to realize efficient collision detection of the manipulator in a small and confined space. The path search tree is initialized, and the nearest neighbor nodes are obtained by random sampling. The nearest neighbor nodes are parallel adaptively expanded based on the target guidance strategy to obtain new extended nodes, which greatly shortens the space manipulator path extension time. For the new extended nodes of the path search tree, batch target direct connection operations based on the efficient collision detection model of the space manipulator are performed to generate multiple feasible space manipulator paths, and the shortest path is selected and sent to the space manipulator for execution, further reducing the motion planning time.

Description

Technical Field

[0001] The present invention belongs to the technical field of aerospace manipulator motion planning, and in particular relates to an efficient and safe motion planning method and system for aerospace manipulators. Background Art

[0002] With the continuous development of missions such as on-orbit spacecraft servicing, space station maintenance, and deep space exploration, the operating environment is becoming increasingly complex. Efficient and safe path planning for multi-degree-of-freedom manipulators in confined spaces presents unprecedented challenges. On the one hand, traditional path planning methods rely on pre-programmed fixed paths, making them inflexible in addressing the uncertainties introduced by dynamic obstacles, structural changes, and unstructured environments. On the other hand, existing methods generally rely on high-precision environmental modeling and complex physical simulations, resulting in significant computational overhead and insufficient response speed, making it difficult to meet the dual requirements of efficiency and real-time performance in space applications. Against this backdrop, breakthroughs in real-time obstacle avoidance and path generation for dynamic environments are urgently needed to enhance the autonomous decision-making and flexible operation capabilities of space manipulators in complex mission scenarios such as debris removal, extravehicular equipment repair, and on-orbit assembly and maintenance. Achieving efficient and autonomous path planning for space manipulators in unstructured environments will not only effectively reduce the frequency of ground intervention and improve operational safety and success rates, but will also play a key role in supporting the long-term operation of future space stations and on-orbit support of deep space probes. This has important theoretical significance and significant engineering application value.

[0003] Existing motion planning technologies are unable to meet the requirements for efficient operation and rapid response of space manipulators in complex aerospace scenarios. Therefore, in light of the aforementioned engineering background, this paper proposes a method and system for efficient and safe motion planning of space manipulators in aerospace scenarios. Summary of the Invention

[0004] In response to the above technical problems, the present invention provides a space robot arm efficient and safe motion planning method and system for aerospace scenarios.

[0005] The technical solution adopted by the present invention to solve the technical problem is:

[0006] An efficient and safe motion planning method for a space robot arm for aerospace scenarios, the method comprising the following steps:

[0007] S100: Uses a multi-purpose depth camera to perceive multi-view point cloud information of aerospace scenes, and fuses the multi-view point cloud information to obtain the complete occupancy Boolean matrix of obstacles within the reach of the space robot arm;

[0008] S200: Based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the spatial manipulator is constructed. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the spatial manipulator is input into the pre-trained gap field neural network learning model to obtain a gap value matrix from the spatial manipulator to each voxel. This is then input into the environment interaction component to output a collision information matrix, and the collision result is obtained based on the collision information matrix.

[0009] S300: Initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes;

[0010] S400: Path search tree The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new expansion node is added to the path search tree;

[0011] S500: Path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

[0012] Preferably, S100 includes:

[0013] S110: Pre-arranged around the space robot arm in the aerospace scene depth cameras with different perspectives, denoted as , The number is set by the user, and it is necessary to ensure that there is no visual obstruction within the operating range of the spatial manipulator;

[0014] S120: Expand the reachable space of the space manipulator Split into Each voxel represents a cube space with a side length of ; For any depth camera , , the obtained obstacle point cloud information is downsampled to the voxel grid, and a dimension of Boolean matrix of , for any Boolean matrix , if a sampling point exists in a voxel, then the Boolean matrix The corresponding element value is set to 1, otherwise it is set to 0; Depth cameras with different viewing angles Repeat the above steps to obtain the information of the obstacle point cloud. Boolean matrices The specific element value of

[0015] S130: In order to fuse obstacle information from multiple perspectives, the Boolean matrices from different depth cameras are merged. Boolean matrices Perform a bitwise OR operation, that is, for the same voxel , , if in any matrix In the matrix, the element value of the voxel is 1, then in the final fused matrix, the corresponding element value is also 1, which can be formally expressed as:

[0016] ;

[0017] in, is the Boolean matrix of the obstacles after final fusion, represents the index of the voxel, and Represents a logical OR operation.

[0018] Preferably, the structure of the gap field neural network learning model in S200 includes five parts. The first part adopts the arbitrary joint angle of the space manipulator. As input features, the structure includes a fully connected layer, an activation function PRelu and a DropOut layer. The output of the first part will be used as the input of the second part;

[0019] The structure of the second part is exactly the same as the first part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the second part will serve as the input of the third part;

[0020] The structure of the third part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the third part serves as the input of the fourth part.

[0021] The structure of the fourth part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the fourth part will be used as the input of the fifth part;

[0022] The fifth part includes the fully connected layer, and the output features are the space robot to Voxel gap value matrix , where the fully connected layers from the second to the fourth part include 1024 neurons, and the fully connected layer of the fifth part includes neurons, gap value matrix The dimension is matrix.

[0023] Preferably, in S200, the gap value matrix is ​​input to the environment interaction component, and a collision information matrix is ​​outputted, and a collision result is obtained based on the collision information matrix, including:

[0024] The environment interaction component converts the output of the gap field neural network learning model into a spatial manipulator. Voxel gap value matrix Converted into a Boolean matrix, that is, if If the value of an element in is greater than 0, it is set to 1; if it is less than 0, it is set to 1; then Obstacle occupancy Boolean matrix after fusion with multiple views Bitwise "OR" operation is performed to obtain a dimension of Collision information matrix ;

[0025] For one or more sets of spatial manipulator joint angles, when they are input into the efficient collision detection model of the spatial manipulator, if the corresponding output collision information matrix If there is an element with a value of "1", it indicates that the spatial manipulator collides with the obstacle under this joint angle configuration; otherwise, it is considered as no collision.

[0026] Preferably, S400 includes:

[0027] S410: For any sampling point The corresponding nearest neighbor node , , the expansion direction is the bias of the random expansion direction and the target expansion direction, the expansion direction = ,in , The new candidate node is ,in , represents the weight coefficient of the target direction bias, represents the target joint configuration;

[0028] S420: For Nearest neighbor nodes , all perform the above operations synchronously to find their corresponding candidate expansion nodes ; then candidate expansion nodes and The nearest neighbor nodes are formed Candidate extension edges ;

[0029] S430: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , thus obtaining a total of interpolation points ,Will interpolation points One-time input of efficient collision detection model for space manipulators to obtain Collision detection results of interpolation points;

[0030] S440: Select the last collision-free interpolation point among the interpolation points starting from the nearest neighbor node of each candidate extension edge as the new path extension node , get a total of A new expansion node is added to the path search tree .

[0031] Preferably, S500 includes:

[0032] S510: Extend the new path node Configured with target joints respectively Separately formed Directly connected candidate edge of the entry ;

[0033] S520: Perform equidistant interpolation on each target directly connected candidate edge, with an interpolation step size of , thus obtaining a total of interpolation points ;

[0034] S530: The interpolation points are input into the efficient collision detection model of the space manipulator at one time to obtain The collision detection results of the interpolation points; if there is no collision at all interpolation points of a target direct candidate edge, it means that the current new path extension node is successfully configured with the target joint. Connect, now a feasible motion path is found , where the batch target direct connection operation process will generate multiple feasible paths, named , is the number of feasible paths, from A shortest path is selected and sent to the spatial manipulator for execution; if a feasible motion path is not successfully found, S400 and S500 are iteratively executed until a feasible motion path is found.

[0035] Preferably, it also includes:

[0036] S600: Targeting the Path Search Tree of A new expansion node is created, and batch rewiring operations based on the efficient collision detection model of the spatial manipulator are performed to further optimize the path quality. The structural reconstruction of the path search tree is completed, and the next expansion and rewiring operations are performed based on the reconstructed path search tree.

[0037] Preferably, S600 includes:

[0038] S610: For each new expansion node , taking it as the center, set the search radius to ,in is the interpolation step size, Indicates the current node number in the path tree; searches for neighboring nodes within its radius and selects the nearest node. nodes as candidate rewiring targets;

[0039] S620: For each pair of new nodes and their neighbor nodes ( ), , perform equidistant interpolation operation to obtain a path interpolation sequence, and the interpolation step size is still ,If the distance between a pair of nodes is not enough to complete the interpolation between the two points, then skip the pair of nodes;

[0040] S630: After all legal interpolation sequences are uniformly spliced, they are input into the efficient collision detection model of the spatial manipulator for batch collision detection. The collision detection results of all interpolation points are obtained. Subsequently, the collision results are grouped according to the start and end indexes of the interpolation sequences to determine whether there are collision points in each interpolation path. If there is no collision in the entire path, it is considered a valid rerouting path.

[0041] S640: For each collision-free path, calculate its rewiring cost , ,in, Represents the current path length cost of the node in the path tree; if Less than the current path length cost of the neighboring node , then update the parent node of the neighbor node to the new node , complete the structural reconstruction of the path tree, and perform the next expansion and rewiring operations based on the reconstructed path search tree.

[0042] An efficient and safe motion planning system for space manipulators for aerospace scenarios, including a complete occupancy matrix acquisition module, a collision result output module, a nearest neighbor node determination module, an adaptive expansion module, and a path generation module;

[0043] The complete occupancy matrix acquisition module is used to use a multi-viewpoint point cloud information perception with a multi-depth camera, and fuse it to obtain the complete occupancy Boolean matrix of obstacles in the space reachable by the space manipulator;

[0044] The collision result output module is used to build an efficient collision detection model for the space manipulator based on the obstacle occupancy Boolean matrix after multi-view fusion. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the space manipulator is input into the pre-trained gap field neural network learning model to obtain the gap value matrix from the space manipulator to the voxel, and then input into the environment interaction component to output the collision information matrix. The collision result is obtained based on the collision information matrix.

[0045] The nearest neighbor node determination module is used to initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes;

[0046] Adaptive extension module for path search trees The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new expansion node is added to the path search tree;

[0047] Path generation module for path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

[0048] The above-mentioned efficient and safe motion planning method and system for space manipulators for aerospace scenarios adopts multiple depth cameras to perceive multi-view point cloud information, and fuses them to obtain the complete occupancy Boolean matrix of obstacles in the space reachable by the space manipulator; based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the space manipulator is constructed to realize efficient collision detection of the manipulator in a narrow and confined space; the path search tree is initialized, and the nearest neighbor nodes are obtained by random sampling, and the nearest neighbor nodes are parallel adaptively expanded based on the target guidance strategy to obtain new extended nodes, which greatly shortens the path extension time of the space manipulator; for the new extended nodes of the path search tree, batch target direct connection operations based on the efficient collision detection model of the space manipulator are performed to generate multiple feasible space manipulator paths, and the shortest path is selected and sent to the space manipulator for execution, which further reduces the motion planning time and realizes efficient and safe motion planning of the space manipulator. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1This is a flow chart of a method for efficient and safe motion planning of a space manipulator for aerospace scenarios in one embodiment of the present invention;

[0050] Figure 2 Schematic diagram of an efficient collision detection model for a space manipulator in one embodiment of the present invention;

[0051] Figure 3 Schematic diagram of batch target direct connection operation based on the efficient collision detection model of the space manipulator in one embodiment of the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] In one embodiment, Figure 1 As shown, a space robot arm efficient and safe motion planning method for aerospace scenarios includes the following steps:

[0054] S100: Uses a multi-purpose depth camera to perceive multi-view point cloud information of aerospace scenes, and fuses the multi-view point cloud information to obtain the complete occupancy Boolean matrix of obstacles within the reach of the space robot arm;

[0055] S200: Based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the spatial manipulator is constructed. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the spatial manipulator is input into the pre-trained gap field neural network learning model to obtain a gap value matrix from the spatial manipulator to each voxel. This is then input into the environment interaction component to output a collision information matrix, and the collision result is obtained based on the collision information matrix.

[0056] S300: Initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes;

[0057] S400: Path search tree The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new expansion node is added to the path search tree;

[0058] S500: Path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

[0059] Specifically, for the following typical aerospace scenarios: a multi-degree-of-freedom space manipulator (degrees of freedom , ), several obstacles, including but not limited to aerospace electrical connectors, cables, and truss joints. Because objects in small, confined spaces are easily obscured, a single depth camera struggles to capture all obstacle information. Therefore, this invention utilizes multiple depth cameras and constructs an efficient collision detection model based on multi-viewpoint point cloud information perception to achieve efficient collision detection for robotic arms in small, confined spaces.

[0060] In one embodiment, S100 includes:

[0061] S110: Pre-arranged around the space robot arm in the aerospace scene depth cameras with different perspectives, denoted as , The number is set by the user, and it is necessary to ensure that there is no visual obstruction within the operating range of the spatial manipulator;

[0062] S120: Expand the reachable space of the space manipulator Split into Each voxel represents a cube space with a side length of ; For any depth camera , , the obtained obstacle point cloud information is downsampled to the voxel grid (that is, the point set falling in each voxel is detected, and a sampling point is taken in each voxel to replace the original point set), and a dimension of Boolean matrix of (a matrix whose elements take only values ​​0 or 1), for any Boolean matrix , if a sampling point exists in a voxel, then the Boolean matrix The corresponding element value is set to 1, otherwise it is set to 0; Depth cameras with different viewing angles Repeat the above steps to obtain the information of the obstacle point cloud. Boolean matrices The specific element value of

[0063] S130: In order to fuse obstacle information from multiple perspectives, the Boolean matrices from different depth cameras are merged. Boolean matrices Perform a bitwise OR operation, that is, for the same voxel , , if in any matrix In the matrix, the element value of the voxel is 1, then in the final fused matrix, the corresponding element value is also 1, which can be formally expressed as:

[0064] ;

[0065] in, is the Boolean matrix of the obstacles after final fusion, represents the index of the voxel, and Represents a logical OR operation.

[0066] Specifically, in this way, the obstacle information from all depth camera perspectives is integrated into a unified Boolean matrix Finally, if a voxel is detected as occupied (i.e., there is an obstacle) at any viewing angle, the voxel is marked as 1 in the fused matrix, otherwise it is marked as 0.

[0067] In one embodiment, Figure 2 As shown in the figure, the structure of the gap field neural network learning model in S200 consists of five parts. The first part adopts the arbitrary joint angle of the space manipulator. As input features, the structure includes a fully connected layer, an activation function PRelu and a DropOut layer. The output of the first part will be used as the input of the second part;

[0068] The structure of the second part is exactly the same as the first part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the second part will serve as the input of the third part;

[0069] The structure of the third part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the third part serves as the input of the fourth part.

[0070] The structure of the fourth part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the fourth part will be used as the input of the fifth part;

[0071] The fifth part includes the fully connected layer, and the output features are the space robot to Voxel gap value matrix , where the fully connected layers from the second to the fourth part include 1024 neurons, and the fully connected layer of the fifth part includes neurons, gap value matrix The dimension is matrix.

[0072] In one embodiment, in S200 , the gap value matrix is ​​input to the environment interaction component, and a collision information matrix is ​​outputted. The collision result is obtained based on the collision information matrix, including:

[0073] The environment interaction component converts the output of the gap field neural network learning model into a spatial manipulator. Voxel gap value matrix Converted into a Boolean matrix, that is, if If the value of an element in is greater than 0, it is set to 1; if it is less than 0, it is set to 1; then Obstacle occupancy Boolean matrix after fusion with multiple views Bitwise "OR" operation is performed to obtain a dimension of Collision information matrix ;

[0074] For one or more sets of spatial manipulator joint angles, when they are input into the efficient collision detection model of the spatial manipulator, if the corresponding output collision information matrix If there is an element with a value of "1", it indicates that the spatial manipulator collides with the obstacle under this joint angle configuration; otherwise, it is considered as no collision.

[0075] In one embodiment, S400 includes:

[0076] S410: For any sampling point The corresponding nearest neighbor node , , the expansion direction is the bias of the random expansion direction and the target expansion direction, the expansion direction = ,in , The new candidate node is ,in , represents the weight coefficient of the target direction bias, represents the target joint configuration;

[0077] S420: For Nearest neighbor nodes , all perform the above operations synchronously to find their corresponding candidate expansion nodes ; then candidate expansion nodes and The nearest neighbor nodes are formed Candidate extension edges ;

[0078] S430: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , thus obtaining a total of interpolation points ,Will interpolation points One-time input of efficient collision detection model for space manipulators to obtain Collision detection results of interpolation points;

[0079] S440: Select the last collision-free interpolation point among the interpolation points starting from the nearest neighbor node of each candidate extension edge as the new path extension node , get a total of A new expansion node is added to the path search tree .

[0080] In one embodiment, Figure 3 As shown, S500 includes:

[0081] S510: Extend the new path node Configured with target joints respectively Separately formed Directly connected candidate edge of the entry ;

[0082] S520: Perform equidistant interpolation on each target directly connected candidate edge, with an interpolation step size of , thus obtaining a total of interpolation points ;

[0083] S530: The interpolation points are input into the efficient collision detection model of the space manipulator at one time to obtain The collision detection results of the interpolation points; if there is no collision at all interpolation points of a target direct candidate edge, it means that the current new path extension node is successfully configured with the target joint. Connect, now a feasible motion path is found , where the batch target direct connection operation process will generate multiple feasible paths, named , is the number of feasible paths, from A shortest path is selected and sent to the spatial manipulator for execution; if a feasible motion path is not successfully found, S400 and S500 are iteratively executed until a feasible motion path is found.

[0084] In one embodiment, it further includes:

[0085] S600: Targeting the Path Search Tree of A new expansion node is created, and batch rewiring operations based on the efficient collision detection model of the spatial manipulator are performed to further optimize the path quality. The structural reconstruction of the path search tree is completed, and the next expansion and rewiring operations are performed based on the reconstructed path search tree.

[0086] In one embodiment, S600 includes:

[0087] S610: For each new expansion node , taking it as the center, set the search radius to ,in is the interpolation step size, Indicates the current node number in the path tree; searches for neighboring nodes within its radius and selects the nearest node. Nodes ( Set by the user, usually ) as candidate rewiring targets;

[0088] S620: For each pair of new nodes and their neighbor nodes ( ), , perform equidistant interpolation operation to obtain a path interpolation sequence, and the interpolation step size is still ,If the distance between a pair of nodes is not enough to complete the interpolation between the two points, then skip the pair of nodes;

[0089] S630: After all legal interpolation sequences are uniformly spliced, they are input into the efficient collision detection model of the spatial manipulator for batch collision detection. The collision detection results of all interpolation points are obtained. Subsequently, the collision results are grouped according to the start and end indexes of the interpolation sequences to determine whether there are collision points in each interpolation path. If there is no collision in the entire path, it is considered a valid rerouting path.

[0090] S640: For each collision-free path, calculate its rewiring cost , ,in, Represents the current path length cost of the node in the path tree; if Less than the current path length cost of the neighboring node , then update the parent node of the neighbor node to the new node , complete the structural reconstruction of the path tree, and perform the next expansion and rewiring operations based on the reconstructed path search tree.

[0091] The present invention aims to solve the problem that existing motion planning technologies are difficult to meet the requirements of efficient operation and rapid response of space manipulators in complex aerospace scenes, and designs an efficient and safe motion planning method and system for space manipulators in aerospace scenes. The main contents of this method are as follows: 1. Construct an efficient collision detection model based on multi-view point cloud information perception to achieve efficient collision detection of manipulators in small and confined spaces: first, multiple depth cameras are used to perceive multi-view point cloud information, and the complete occupancy matrix of obstacles in the space that can be reached by the space manipulator is obtained by fusion; secondly, based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the space manipulator is constructed. 2. Design a parallel adaptive expansion and rewiring mechanism for the joint space of the space manipulator, and construct a fast response motion planning module based on the efficient collision detection model: first, for the above path search tree of The nearest neighbor nodes are expanded in parallel based on the target guidance strategy; secondly, the path search tree of A new extension node is created to perform batch rewiring operations based on the efficient collision detection model of the space manipulator to further optimize the path quality; finally, the path search tree of A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

[0092] The beneficial effects of the present invention are as follows:

[0093] (1) The present invention uses multiple depth cameras to perceive multi-view point cloud information of the space manipulator, and the obstacle information under the perspective of all depth cameras is integrated into a unified Boolean matrix Finally, if a voxel is detected as occupied (i.e., an obstacle exists) from any viewpoint, it is marked as 1 in the fused matrix; otherwise, it is marked as 0. This method effectively combines obstacle information from multiple depth camera views through a simple bitwise OR operation, ensuring that the occupancy information of each voxel fully reflects the obstacle position from multiple viewpoints.

[0094] (2) The parallel adaptive expansion strategy based on the target guidance strategy proposed in this invention guides the expansion direction toward the target joint configuration for parallel expansion to a certain extent, avoiding the waste of computing resources caused by random expansion. At the same time, an efficient collision detection model of the space manipulator is used for parallel collision detection during the expansion process, which greatly shortens the path expansion time of the space manipulator.

[0095] (3) After the parallel adaptive path expansion, the present invention designs a batch rewiring strategy based on the efficient collision detection model of the space manipulator to further optimize the path quality. Compared with the traditional method of rewiring operation, which shortens the path length but significantly increases the time consumption, the batch rewiring strategy introduces the efficient collision detection model of the space manipulator for parallel collision detection, shortening the path length while maintaining the advantage of low time consumption.

[0096] (4) The batch target direct connection strategy designed by the present invention based on the efficient collision detection model of the space manipulator helps the space manipulator to quickly find a feasible path, further reducing the motion planning time.

[0097] In one embodiment, an efficient and safe motion planning system for a space manipulator for aerospace scenarios is also provided, comprising a complete occupancy matrix acquisition module, a collision result output module, a nearest neighbor node determination module, an adaptive expansion module, and a path generation module;

[0098] The complete occupancy matrix acquisition module is used to use a multi-viewpoint point cloud information perception with a multi-depth camera, and fuse it to obtain the complete occupancy Boolean matrix of obstacles in the space reachable by the space manipulator;

[0099] The collision result output module is used to build an efficient collision detection model for the space manipulator based on the obstacle occupancy Boolean matrix after multi-view fusion. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the space manipulator is input into the pre-trained gap field neural network learning model to obtain the gap value matrix from the space manipulator to the voxel, and then input into the environment interaction component to output the collision information matrix. The collision result is obtained based on the collision information matrix.

[0100] The nearest neighbor node determination module is used to initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes;

[0101] Adaptive extension module for path search trees The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new expansion node is added to the path search tree;

[0102] Path generation module for path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

[0103] Regarding the specific limitations of the efficient and safe motion planning system for space manipulators for aerospace scenarios, please refer to the limitations of the efficient and safe motion planning method for space manipulators for aerospace scenarios above, which will not be repeated here. Each module in the above-mentioned efficient and safe motion planning system for space manipulators for aerospace scenarios can be implemented in whole or in part through 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 of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0104] The above is a detailed introduction to the efficient and safe motion planning method and system for space manipulators for aerospace scenarios 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. An efficient and safe motion planning method for aerospace robotic arms, characterized by: The method comprises the following steps: S100: Uses a multi-purpose depth camera to perceive multi-view point cloud information of aerospace scenes, and fuses the multi-view point cloud information to obtain the complete occupancy Boolean matrix of obstacles within the reach of the space robot arm; S200: Based on the obstacle occupancy Boolean matrix after multi-view fusion, an efficient collision detection model for the spatial manipulator is constructed. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the spatial manipulator is input into the pre-trained gap field neural network learning model to obtain a gap value matrix from the spatial manipulator to each voxel, and then input into the environment interaction component to output a collision information matrix. A collision result is obtained based on the collision information matrix. In S200, the gap value matrix is ​​input into the environment interaction component to output a collision information matrix. A collision result is obtained based on the collision information matrix, including: The environment interaction component converts the output of the gap field neural network learning model into a spatial manipulator. Voxel gap value matrix Converted into a Boolean matrix, that is, if If the value of an element in is greater than 0, it is set to 0, and if it is less than 0, it is set to 1; then Obstacle occupancy Boolean matrix after fusion with multiple views Bitwise "OR" operation is performed to obtain a dimension of Collision information matrix ; For one or more sets of spatial manipulator joint angles, when they are input into the efficient collision detection model of the spatial manipulator, if the corresponding output collision information matrix If there is an element with a value of "1", it indicates that the spatial manipulator collides with the obstacle under this joint angle configuration; otherwise, it is considered as no collision; S300: Initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes; S400: Path search tree The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new extension node is added to the path search tree; S400 includes: S410: For any sampling point The corresponding nearest neighbor node , , the expansion direction is the bias of the random expansion direction and the target expansion direction, the expansion direction = ,in , The new candidate node is ,in , represents the weight coefficient of the target direction bias, represents the target joint configuration; S420: For Nearest neighbor nodes , all perform the above operations synchronously to find their corresponding candidate expansion nodes ; then candidate expansion nodes and The nearest neighbor nodes are formed Candidate extension edges ; S430: Perform equidistant interpolation on each candidate extension edge, with an interpolation step size of , thus obtaining a total of interpolation points ,Will interpolation points One-time input of efficient collision detection model for space manipulators to obtain Collision detection results of interpolation points; S440: Select the last collision-free interpolation point among the interpolation points starting from the nearest neighbor node of each candidate extension edge as the new path extension node , get a total of A new expansion node is added to the path search tree ; S500: Path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

2. The method according to claim 1, characterized in that S100 includes: S110: Pre-arranged around the space robot arm in the aerospace scene depth cameras with different perspectives, denoted as , The number is set by the user, and it is necessary to ensure that there is no visual obstruction within the operating range of the spatial manipulator; S120: Expand the reachable space of the space manipulator Split into Each voxel represents a cube space with a side length of ; For any depth camera , , the obtained obstacle point cloud information is downsampled to the voxel grid, and a dimension of Boolean matrix of , for any Boolean matrix , if a sampling point exists in a voxel, then the Boolean matrix The corresponding element value is set to 1, otherwise it is set to 0; Depth cameras with different viewing angles Repeat the above steps to obtain the information of the obstacle point cloud. Boolean matrices The specific element value of S130: In order to fuse obstacle information from multiple perspectives, the Boolean matrices from different depth cameras are merged. Boolean matrices Perform bitwise "OR" operation, that is, for the same voxel , , if in any matrix In the matrix, the element value of the voxel is 1, then in the final fused matrix, the corresponding element value is also 1, which can be formally expressed as: ； in, is the Boolean matrix of the obstacles after final fusion, represents the index of the voxel, and Represents a logical "OR" operation.

3. The method according to claim 2, characterized in that The structure of the gap field neural network learning model in S200 consists of five parts. The first part adopts the arbitrary joint angle of the space manipulator. As input features, the structure includes a fully connected layer, an activation function PRelu and a DropOut layer. The output of the first part will be used as the input of the second part; The structure of the second part is exactly the same as the first part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the second part will serve as the input of the third part; The structure of the third part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the third part serves as the input of the fourth part. The structure of the fourth part is exactly the same as that of the second part, including the fully connected layer, the activation function PRelu and the DropOut layer. The output of the fourth part will be used as the input of the fifth part; The fifth part includes the fully connected layer, and the output features are the space robot to Voxel gap value matrix , where the fully connected layers from the second to the fourth part include 1024 neurons, and the fully connected layer of the fifth part includes neurons, gap value matrix The dimension is matrix.

4. The method according to claim 3, characterized in that S500 includes: S510: Extend the new path node Configured with target joints respectively Separately formed Directly connected candidate edge of the entry ; S520: Perform equidistant interpolation on each target directly connected candidate edge, with an interpolation step size of , thus obtaining a total of interpolation points ; S530: The interpolation points are input into the efficient collision detection model of the space manipulator at one time to obtain The collision detection results of the interpolation points; if there is no collision at all interpolation points of a target direct candidate edge, it means that the current new path extension node is successfully configured with the target joint. Connect, now a feasible motion path is found , where the batch target direct connection operation process will generate multiple feasible paths, named , is the number of feasible paths, from A shortest path is selected and sent to the spatial manipulator for execution; if a feasible motion path is not successfully found, S400 and S500 are iteratively executed until a feasible motion path is found.

5. The method according to claim 4, characterized in that Also includes: S600: Targeting the Path Search Tree of A new expansion node is created, and batch rewiring operations based on the efficient collision detection model of the spatial manipulator are performed to further optimize the path quality. The structural reconstruction of the path search tree is completed, and the next expansion and rewiring operations are performed based on the reconstructed path search tree.

6. The method according to claim 5, characterized in that S600 includes: S610: For each new expansion node , taking it as the center, set the search radius to ,in is the interpolation step size, Indicates the current node number in the path tree; searches for neighboring nodes within its radius and selects the nearest node. nodes as candidate rewiring targets; S620: For each pair of new nodes and their neighbor nodes ( ), , perform equidistant interpolation operation to obtain a path interpolation sequence, and the interpolation step size is still ,If the distance between a pair of nodes is not enough to complete the interpolation between the two points, then skip the pair of nodes; S630: After all legal interpolation sequences are uniformly spliced, they are input into the efficient collision detection model of the spatial manipulator for batch collision detection. The collision detection results of all interpolation points are obtained. Subsequently, the collision results are grouped according to the start and end indexes of the interpolation sequences to determine whether there are collision points in each interpolation path. If there is no collision in the entire path, it is considered a valid rerouting path. S640: For each collision-free path, calculate its rewiring cost , ,in, Represents the current path length cost of the node in the path tree; if Less than the current path length cost of the neighboring node , then update the parent node of the neighbor node to the new node , complete the structural reconstruction of the path tree, and perform the next expansion and rewiring operations based on the reconstructed path search tree.

7. An efficient and safe motion planning system for aerospace-oriented space manipulators based on the method according to any one of claims 1 to 6, characterized in that: It includes a complete occupancy matrix acquisition module, a collision result output module, a nearest neighbor node determination module, an adaptive expansion module, and a path generation module; The complete occupancy matrix acquisition module is used to use a multi-viewpoint point cloud information perception with a multi-depth camera, and fuse it to obtain the complete occupancy Boolean matrix of obstacles in the space reachable by the space manipulator; The collision result output module is used to build an efficient collision detection model for the space manipulator based on the obstacle occupancy Boolean matrix after multi-view fusion. The model includes a pre-trained gap field neural network learning model and an environment interaction component. Any joint angle of the space manipulator is input into the pre-trained gap field neural network learning model to obtain the gap value matrix from the space manipulator to the voxel, and then input into the environment interaction component to output the collision information matrix. The collision result is obtained based on the collision information matrix. The nearest neighbor node determination module is used to initialize the path search tree of the space manipulator, set the starting joint configuration and target joint configuration of the space manipulator, use the starting joint configuration as the initial node of the path search tree, and randomly sample the joint space of the space manipulator at one time. Group joint configuration points and find the closest distance to each sampling point in the path search tree path nodes, and get nearest neighbor nodes; Adaptive extension module for path search trees The nearest neighbor nodes are expanded in parallel based on the target-guided strategy to obtain A new expansion node is added to the path search tree; Path generation module for path search tree A new extension node is created to perform batch target direct connection operations based on the efficient collision detection model of the space manipulator, generate multiple feasible space manipulator paths, and select the shortest path to send to the space manipulator for execution.

Images