A two-stage opportunity decision view planning method for automobile hub detection

Abstract

The application discloses a two-stage opportunity decision viewpoint planning method for automobile wheel hub detection and belongs to the technical field of industrial detection robots. The method takes a wheel hub CAD model as input, divides candidate viewpoints into three categories of forced viewpoints, backbone viewpoints and opportunity viewpoints in an offline stage, and constructs a candidate viewpoint space suitable for online planning through key hole positioning, surface parameterization sampling and other methods. In the online stage, a reference viewpoint with global guidance is first determined in the forced viewpoint set and the backbone viewpoint set, then the opportunity viewpoint set is evaluated according to coverage benefits and motion costs, if there is a better opportunity viewpoint, the current reference viewpoint is dynamically replaced and updated, and finally the coverage state and the remaining task are updated in real time according to the execution feedback. The application can guarantee reliable observation of key features, improve complex surface coverage and planning efficiency, reduce redundant motion, and enhance the adaptability to occlusion and execution deviation.

Description

Technical Field

[0001] This invention belongs to the field of industrial inspection robot technology, specifically relating to a two-stage opportunistic decision-making viewpoint planning method for automobile wheel hub inspection. Background Technology

[0002] As a critical safety component, the surface quality of automotive wheel hubs directly impacts driving safety. During the production process, comprehensive inspection of the wheel hub surface is necessary to identify defects such as cracks, porosity, and scratches. Traditional manual inspection is inefficient and highly subjective; therefore, automated inspection based on machine vision has become the mainstream solution.

[0003] In machine vision inspection systems, viewpoint planning determines the positions and angles from which the inspection robot acquires images, directly affecting the inspection coverage and efficiency. As a complex curved surface part, the wheel hub has the following characteristics: (1) complex geometry, including key features such as bolt holes and valve holes; (2) large surface curvature changes, resulting in self-occlusion; (3) different areas have different inspection importance, and key features must be completely omitted.

[0004] Existing viewpoint planning methods mainly suffer from the following problems:

[0005] (1) Fixed viewpoint method: Pre-setting a fixed number and position of viewpoints cannot adapt to the geometric features of different wheel hubs, which may lead to the omission of key features or viewpoint redundancy.

[0006] (2) Geometric coverage-based methods: aim to maximize surface coverage, but do not distinguish the importance difference between key features and ordinary areas, and cannot guarantee the complete detection of key features.

[0007] (3) Optimization algorithm-based method: The optimal viewpoint configuration is solved by global optimization, but the computational complexity is high and it is difficult to meet the needs of online dynamic adjustment.

[0008] (4) Lack of online feedback mechanism: The viewpoint sequence planned offline cannot be dynamically adjusted according to the actual execution situation. When there are situations such as occlusion or positioning error, it cannot be compensated in time.

[0009] Therefore, a viewpoint planning method is needed that can ensure zero omission of key features, efficiently cover complex surfaces, and support online dynamic adjustment. Summary of the Invention

[0010] Purpose of the invention

[0011] The purpose of this invention is to provide a two-stage opportunistic decision-making viewpoint planning method for automobile wheel hub detection, which solves the problem in the prior art of achieving both zero omission of key features and efficient coverage of complex surfaces.

[0012] Technical solution

[0013] A two-stage opportunistic decision-making viewpoint planning method for automobile wheel hub inspection, characterized by the following steps:

[0014] Step 1, Offline Candidate Viewpoint Space Construction: Using the CAD model of the car wheel hub as input, and considering the task differences in efficiency between complete observation of key features and ordinary surface coverage, the candidate viewpoint space is modeled. This involves sequentially performing key hole localization, surface parametric sampling, candidate viewpoint generation and accessibility verification, and backbone viewpoint selection, thus constructing a forced viewpoint set. Backbone viewpoint set and Opportunity Viewpoints The candidate viewpoint space is formed;

[0015] Step 2, Online Planning: Based on the robot's current state, the covered area, and the remaining task set, determine a reference viewpoint within the action space comprised of the remaining forced viewpoint set and the remaining backbone viewpoint set; plan around the reference viewpoint using the opportunity viewpoint set... A local opportunity window is constructed, and a coarse screening is performed based on the proxy motion cost and a precise verification is performed based on the actual path cost to obtain candidate viewpoints within the opportunity window. For candidate viewpoints within the opportunity window, a utility ratio is constructed based on their marginal coverage gain and motion cost. When a candidate opportunity viewpoint simultaneously meets the coverage gain improvement condition and the utility ratio improvement condition relative to the benchmark viewpoint, the candidate opportunity viewpoint is executed; otherwise, the benchmark viewpoint is executed. After the selected viewpoint is executed, the robot state, coverage state, and remaining task set are updated, and the next round of planning is triggered until all forced viewpoints are completed and the surface coverage reaches the preset target.

[0016] Furthermore, the offline candidate viewpoint space construction in step 1 includes:

[0017] a) Based on the observation integrity requirements and coverage efficiency requirements of the detection area, perform task modeling for candidate viewpoints;

[0018] b) The center positions of the wheel hub bolt holes, center holes, and valve holes are determined using a critical hole positioning process, and a forced viewpoint set is generated. ;

[0019] c) A cylindrical coordinate-based surface parameterization method is used to parametrically sample the hub surface to obtain the parametric coordinates of the surface vertices, and coarse and fine sampling meshes are established. The coarse sampling mesh is used to generate candidate viewpoints for the surface, and the fine sampling mesh is used to generate the set of sampling points for coverage evaluation. ;

[0020] d) Generate initial surface candidate viewpoints at the effective parameter positions corresponding to the coarse sampling grid, and perform inverse kinematics solution and collision detection on the initial surface candidate viewpoints to obtain effective surface viewpoints;

[0021] e) Select a backbone viewpoint set from the effective surface viewpoints. The remaining valid surface viewpoints that were not selected constitute the opportunity viewpoint set. .

[0022] Furthermore, the task modeling is as follows: the viewpoints corresponding to the key feature regions that need to be observed in their entirety are defined as a forced viewpoint set. The viewpoints used to maintain the overall coverage structure of ordinary curved surfaces are defined as the backbone viewpoint set. Viewpoints used to supplement local along-path coverage during online execution are defined as opportunity viewpoint sets. .

[0023] Furthermore, the key hole location process includes: utilizing the property that the surface normal vector of the hole wall is approximately orthogonal to the hub spindle, selecting a candidate vertex set from the hub mesh; projecting the candidate vertex set onto a plane perpendicular to the hub spindle, and using DBSCAN clustering to distinguish the vertex clusters corresponding to different holes; for each vertex cluster corresponding to a hole, calculating the geometric center in the projection plane, and correcting the hole center by combining the extreme position in the spindle direction, so as to determine the hole center position.

[0024] Furthermore, generate a forced viewpoint set. At that time, the observation distance was determined based on the pinhole camera imaging model. Let the aperture be The camera's horizontal field of view is The redundancy coefficient is Then the equivalent aperture The observation distance satisfies And observation distance Each hole generates a central viewpoint and two left and right offset viewpoints.

[0025] Furthermore, the surface parameterization method based on cylindrical coordinates includes: projecting the vertices of the hub surface onto the XZ plane, and calculating the azimuth angle parameters of the vertices relative to the geometric center of the hub. ;

[0026] Combining the radial and axial information of the vertices Map it to the hub section profile curve The axial coordinates are defined by the arc length on the cross-sectional profile. Therefore, all mesh vertices have corresponding parametric coordinates. Based on this, a two-layer sampling mesh was constructed, in which the coarse mesh generates candidate surface viewpoints, forming a set of surface viewpoints. Fine grids generate dense sampling points, forming a coverage evaluation set. It is used to quantitatively evaluate the coverage capability of a viewpoint.

[0027] Furthermore, the candidate viewpoint generation and backbone screening include: selecting sampling points at the corresponding effective parameter positions on the coarse sampling grid, and generating initial surface candidate viewpoints using local normal vector regularization and normal offset;

[0028] The initial candidate surface viewpoints are subjected to inverse kinematics solution and collision detection. Candidate viewpoints with collision-free inverse kinematics solutions are retained to form an effective set of surface viewpoints.

[0029] The upper limit of the number of backbone viewpoints is determined based on the total area of ​​the surface under test, the field of view coverage area of ​​a single viewpoint, and the field of view overlap coefficient. Backbone viewpoints are then selected sequentially from the effective surface viewpoint set according to the marginal coverage gain until the upper limit is reached or the maximum marginal coverage gain is lower than a preset threshold. Obtain the backbone viewpoint set The remaining valid surface viewpoints that were not selected constitute the opportunity viewpoint set. .

[0030] Furthermore, the online planning in step 2 includes:

[0031] a) Determining the reference viewpoint: Based on the robot's current state, the covered area, and the remaining forced viewpoints and remaining backbone viewpoints, determine the reference viewpoint that has a global guiding role;

[0032] b) Baseline viewpoint adjustment: Around the said baseline viewpoint, in the opportunity viewpoint set A local opportunity window is constructed, and the decision on whether to replace the baseline viewpoint with an opportunity viewpoint is made based on the coverage benefits and motion costs.

[0033] c) Viewpoint path generation: After selecting a viewpoint, update the robot state, coverage state, and remaining task set, and trigger the next round of planning to dynamically generate the robotic arm viewpoint execution sequence.

[0034] Furthermore, the determination of the reference viewpoint includes: modeling the online decision-making process as a Markov decision process, and using Monte Carlo tree search to determine the reference viewpoint. The state includes at least the robot's current joint angle, the set of sampled points on the covered surface, the set of remaining forced viewpoints, and the set of remaining backbone viewpoints. The action space consists of the set of remaining forced viewpoints and the set of remaining backbone viewpoints. A reward function is constructed that simultaneously considers the cost of new coverage, proxy motion, and the priority of forced viewpoints. The reference viewpoint at the current moment is determined using Monte Carlo tree search through a selection, expansion, simulation, and backhaul process.

[0035] Furthermore, the baseline viewpoint adjustment and viewpoint path generation include: first adjusting the opportunity viewpoint set based on the agent motion cost. Perform a coarse screening and retain those that meet the requirements. The opportunity viewpoints are then precisely verified based on the actual path cost output by the motion planner to obtain candidate viewpoints within the opportunity window.

[0036] Opportunity window relaxation parameters are satisfied ,in This is the baseline relaxation value. For the current coverage rate, yes Task adjustment factor.

[0037] This is the coverage adjustment factor, where, Coverage rate at the current moment; For target coverage; For sensitivity coefficient, control The strength of the response to coverage gaps; To regulate the upper bound of the factor, to prevent Too large; To perform the positive part operation, For the remaining Number of tasks. hour, , hour, .

[0038] Furthermore, regarding candidate viewpoints within the opportunity window Based on its marginal coverage gain Motion cost from the current state to the candidate viewpoint Construct utility ratio And select the opportunity viewpoint with the highest utility ratio. ;when , At that time, the opportunity viewpoint Otherwise, execute the baseline viewpoint. After execution, update the robot joint configuration and the set of covered surface sampling points. When the viewpoint actually executed belongs to the remaining forced viewpoint set or the remaining backbone viewpoint set, remove the viewpoint from the corresponding remaining task set until all forced viewpoints are completed and the surface coverage reaches the preset target.

[0039] The beneficial effects of this invention are:

[0040] (1). By classifying candidate viewpoints into three categories—mandatory viewpoints, backbone viewpoints, and opportunistic viewpoints—a unified representation of the requirements for complete observation of key features and the requirements for coverage efficiency of ordinary curved surfaces is achieved, which is conducive to improving the pertinence and structure of viewpoint planning.

[0041] (2). By combining key hole location, cylindrical coordinate parameterization, candidate viewpoint accessibility verification and backbone selection in an offline construction method, the online search space can be compressed while ensuring geometric adaptability, thereby improving planning efficiency.

[0042] (3). By combining the determination and adjustment of the reference viewpoint in the online planning mechanism, it is possible to explore local along-the-way coverage opportunities while maintaining the overall path structure, thereby reducing redundant pose switching and ineffective movement.

[0043] (4). By implementing a feedback-driven closed-loop update mechanism, subsequent decisions can be dynamically adjusted based on the real-time coverage status and remaining tasks, thereby improving the system's adaptability to positioning errors, partial occlusion, and execution deviations.

[0044] (5). This invention takes into account key feature observation constraints, complex surface coverage quality and online planning efficiency, and is suitable for complex surface industrial inspection scenarios such as automobile wheel hubs, and has good engineering deployability. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the overall process framework of the method of the present invention.

[0046] Figure 2 This is a diagram illustrating the overall framework of offline candidate viewpoint space construction and online planning for this invention.

[0047] Figure 3 This is a schematic diagram of the surface parameterization method based on cylindrical coordinates according to the present invention;

[0048] Figure 4 This is a schematic diagram illustrating the candidate viewpoint generation and reachability verification of the present invention.

[0049] Figure 5 This is a flowchart of the online decision-making process of the present invention. Detailed Implementation

[0050] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0051] Example

[0052] This embodiment provides a specific implementation of a two-stage opportunity decision-making viewpoint planning method for automobile wheel hub inspection.

[0053] like Figures 1 to 5 As shown, the method of this invention includes two stages: offline candidate viewpoint space construction and online planning. The offline stage uses a wheel hub CAD model as input and sequentially completes task modeling, key hole localization, surface parametric sampling, candidate viewpoint generation and accessibility verification, and backbone viewpoint selection, resulting in a forced viewpoint set. Backbone viewpoint set and Opportunity Viewpoints The candidate viewpoint space is formed; in the online phase, based on the robot's current state, the covered area, and the remaining task set, the baseline viewpoint is determined, the baseline viewpoint is adjusted, and the closed-loop execution is updated, ultimately generating the viewpoint execution path.

[0054] The offline phase is detailed as follows:

[0055] Step 1: Task Modeling

[0056] Different regions of the wheel hub exhibit significant differences in detection priority, completeness requirements, and access strategies. Therefore, this embodiment categorizes candidate viewpoints based on the rigidity of task constraints. Viewpoints corresponding to key feature regions requiring complete observation are defined as the mandatory viewpoint set. The viewpoints used to maintain the overall coverage structure of ordinary curved surfaces are defined as the backbone viewpoint set. Viewpoints used to supplement local along-path coverage during online execution are defined as opportunity viewpoint sets. By employing the modeling approach described above, key feature observation constraints and ordinary surface coverage tasks are integrated into a single viewpoint planning framework.

[0057] Step 2: Locating key holes

[0058] To generate a forced viewpoint set First, the key holes in the wheel hub are located. This embodiment adopts a positioning process of "normal filtering - projection clustering - axial correction". Specifically, the candidate vertex set is filtered from the wheel hub mesh by utilizing the property that the surface normal vector of the hole wall is approximately orthogonal to the main axis of the wheel hub. Then, the candidate vertices are projected onto a plane perpendicular to the main axis of the wheel hub, and DBSCAN is used to cluster the projected points to distinguish the vertex clusters corresponding to different holes. Finally, for the vertex cluster corresponding to each hole, the geometric center is calculated in the projection plane, and the hole center is corrected by combining the extreme position of the main axis direction, thereby reducing the positioning deviation caused by the chamfer structure.

[0059] Step 3: Surface parametric sampling

[0060] To address the issue of large curvature variations in the wheel hub surface and the potential for uneven distribution due to direct uniform sampling in Cartesian space, this embodiment employs a cylindrical coordinate-based surface parameterization method for parametric sampling of the wheel hub surface. Let the geometric center of the wheel hub be... For surface vertices First, project it onto the XZ plane and calculate the azimuth parameters relative to the geometric center: .

[0061] Then combine the radial and axial information of the vertex. Map it to the hub section profile curve The axial coordinates are defined by the arc length on the cross-sectional profile. Therefore, all mesh vertices have corresponding parametric coordinates. Based on this, a two-layer sampling mesh was constructed, in which the coarse mesh generates candidate surface viewpoints, forming a set of surface viewpoints. Fine grids generate dense sampling points, forming a coverage evaluation set. It is used to quantitatively evaluate the coverage capability of a viewpoint.

[0062] Step 4: Candidate viewpoint generation and reachability verification

[0063] After obtaining the location of the key hole and the surface parameterization results, different types of candidate viewpoints can be generated. For the key hole region, the observation distance is determined based on the pinhole camera imaging model. Let the aperture be... The camera's horizontal field of view is The redundancy coefficient is Then the equivalent aperture The observation distance satisfies: and will Limited to Within the range. For each hole, a central viewpoint and two left and right offset viewpoints are generated, forming the corresponding forced viewpoint.

[0064] For ordinary curved surface regions, sampling points are selected on the effective parameter grid, and initial candidate viewpoints are generated using local normal vector regularization and normal offset. Candidate viewpoints generated by geometric rules may not be directly executable; therefore, inverse kinematics solving and collision detection are required. Candidate viewpoints with collision-free inverse kinematics solutions are retained, forming a set of effective surface viewpoints.

[0065] Step 5: Core Viewpoint Selection

[0066] After obtaining the effective set of surface viewpoints, to reduce online search overhead, this embodiment selects a small number of backbone viewpoints as the main structure for ordinary surface coverage. First, the upper limit of the number of backbone viewpoints is determined based on the total area of ​​the surface under test, the single-viewpoint field of view coverage area, and the field of view overlap coefficient. Then, a greedy selection strategy driven by marginal coverage gain is adopted to select backbone viewpoints from the effective surface viewpoints.

[0067] From the point set, the candidate viewpoints with the largest new coverage capabilities are selected and added to the backbone viewpoint set in turn. When the maximum quantity is reached or the maximum marginal coverage gain is lower than a preset threshold. When the selection stops, the remaining valid surface viewpoints that were not selected constitute the opportunity viewpoint set. .

[0068] The online phase is detailed as follows:

[0069] Step 6: Determine the reference viewpoint

[0070] To obtain the next execution target with global guidance, this embodiment models the online decision-making process as a Markov decision process and uses Monte Carlo tree search to determine the baseline viewpoint. The state includes at least the robot's current joint angle, the set of sampled points on the covered surface, the set of remaining forced viewpoints, and the set of remaining backbone viewpoints. The action space is defined by the set of remaining forced viewpoints. and the remaining backbone viewpoint set Composition. The reward function simultaneously considers new coverage, proxy motion costs, and forced viewpoint priority. As one implementation method, it can take the following form:

[0071] ;

[0072] in, Indicates the execution of an action The resulting increase in coverage Indicates the cost of proxy movement. For indicator functions, This is used to enhance the forced viewpoint priority. A Monte Carlo tree search is employed, involving selection, expansion, simulation, and backpropagation processes, to determine the reference viewpoint at the current moment. .

[0073] Step 7: Adjusting the reference viewpoint

[0074] In obtaining the reference viewpoint Subsequently, this embodiment focuses on the opportunity viewpoint set around the benchmark viewpoint. Constructing a local opportunity window. First, a coarse screening is performed based on the agent's motion cost, retaining those that meet the requirements. The opportunity viewpoints are then precisely validated based on the actual path cost output by the motion planner to obtain candidate viewpoints within the opportunity window. The opportunity window relaxation parameters are adaptive.

[0075] ;

[0076] in, As a coverage adjustment factor, As the baseline relaxation value, Coverage rate at the current moment; For target coverage; Sensitivity coefficient; This is the upper bound of the regulation factor; To perform the positive part operation, For the remaining Number of tasks. When hour, ; hour, .

[0077] For candidate viewpoints within the opportunity window The utility ratio is constructed based on its marginal coverage gain and actual motion cost:

[0078] ;

[0079] Select the opportunity perspective that maximizes utility and with reference viewpoint Comparison. When , At that time, the opportunity viewpoint Otherwise, execute the baseline viewpoint. .

[0080] Step 8: Viewpoint Path Generation

[0081] After selecting a viewpoint, the robot joint configuration, the set of covered surface sampling points, and the remaining task set are updated to dynamically generate the robot arm's viewpoint execution sequence. If the actually executed viewpoint belongs to the remaining forced viewpoint set or the remaining backbone viewpoint set, the viewpoint is removed from the corresponding remaining task set; if the executed viewpoint is an opportunistic viewpoint, the coverage status is updated but the remaining backbone task set is not changed. Then, the next round of online planning is triggered, and the process is repeated until all forced viewpoints are completed and the surface coverage reaches the preset target.

[0082] In this embodiment, the offline candidate viewpoint space construction is used to compress the online search space and provide structured priors; online planning is used to explore local along-the-path coverage opportunities while ensuring the observation constraints of key features, thereby taking into account coverage quality, movement cost and planning efficiency.

Claims

1. A two-stage opportunistic decision-making viewpoint planning method for automobile wheel hub inspection, characterized in that, Includes the following steps: Step 1, Offline Candidate Viewpoint Space Construction: Using the CAD model of the car wheel hub as input, and considering the task differences in efficiency between complete observation of key features and ordinary surface coverage, the candidate viewpoint space is modeled. This involves sequentially performing key hole localization, surface parametric sampling, candidate viewpoint generation and accessibility verification, and backbone viewpoint selection, thus constructing a forced viewpoint set. Backbone viewpoint set and Opportunity Viewpoints The candidate viewpoint space is formed; Step 2, Online Planning: Based on the robot's current state, the covered area, and the remaining task set, determine the reference viewpoint within the action space composed of the remaining forced viewpoint set and the remaining backbone viewpoint set; Around the baseline viewpoint in the opportunity viewpoint set A local opportunity window is constructed, and a coarse screening is performed based on the proxy motion cost and a precise verification is performed based on the actual path cost to obtain candidate viewpoints within the opportunity window; For candidate viewpoints within the opportunity window, a utility ratio is constructed based on their marginal coverage gain and motion cost. When a candidate opportunity viewpoint simultaneously satisfies the coverage gain improvement condition and the utility ratio improvement condition relative to the benchmark viewpoint, the candidate opportunity viewpoint is executed; otherwise, the benchmark viewpoint is executed. After executing the selected viewpoint, the robot state, coverage state, and remaining task set are updated, and the next round of planning is triggered until all forced viewpoints are completed and the surface coverage reaches the preset target.

2. The method according to claim 1, characterized in that, Step 1, the offline candidate viewpoint space construction, includes: a) Based on the observation integrity requirements and coverage efficiency requirements of the detection area, perform task modeling for candidate viewpoints; b) The center positions of the wheel hub bolt holes, center holes, and valve holes are determined using a critical hole positioning process, and a forced viewpoint set is generated. ; c) A cylindrical coordinate-based surface parameterization method is used to parametrically sample the hub surface to obtain the parametric coordinates of the surface vertices, and coarse and fine sampling meshes are established. The coarse sampling mesh is used to generate candidate viewpoints for the surface, and the fine sampling mesh is used to generate the set of sampling points for coverage evaluation. ; d) Generate initial surface candidate viewpoints at the effective parameter positions corresponding to the coarse sampling grid, and perform inverse kinematics solution and collision detection on the initial surface candidate viewpoints to obtain effective surface viewpoints; e) Select a backbone viewpoint set from the effective surface viewpoints. The remaining valid surface viewpoints that were not selected constitute the opportunity viewpoint set. .

3. The method according to claim 2, characterized in that, In step a), the task is modeled as follows: The viewpoints corresponding to the key feature regions that require all observations are defined as the forced viewpoint set. ; The viewpoints used to maintain the overall coverage structure of ordinary curved surfaces are defined as the backbone viewpoint set. ; Viewpoints used to supplement local along-the-path coverage during online execution are defined as opportunity viewpoint sets. .

4. The method according to claim 2, characterized in that, The critical hole positioning process described in step b) includes: By utilizing the property that the surface normal vector of the hole wall is approximately orthogonal to the hub's main axis, a set of candidate vertices is selected from the hub mesh; The candidate vertex set is projected onto a plane perpendicular to the hub's main axis, and DBSCAN clustering is used to distinguish the vertex clusters corresponding to different holes. For each vertex cluster corresponding to a hole, the geometric center is calculated in the projection plane, and the hole center is corrected by combining the extreme position of the principal axis direction to determine the hole center position.

5. The method according to claim 2, characterized in that, In step b), a forced viewpoint set is generated. At that time, the observation distance was determined based on the pinhole camera imaging model. Let the aperture be The camera's horizontal field of view is The redundancy coefficient is Then the equivalent aperture The observation distance satisfies And observation distance Each hole generates a central viewpoint and two left and right offset viewpoints.

6. The method according to claim 2, characterized in that, The surface parameterization method based on cylindrical coordinates described in step c) includes: Project the vertex of the wheel hub surface onto the XZ plane and calculate the azimuth angle parameter of the vertex relative to the geometric center of the wheel hub. ; Combining the radial and axial information of the vertices Map it to the hub section profile curve The axial coordinates are defined by the arc length on the cross-sectional profile. Therefore, all mesh vertices have corresponding parametric coordinates. ; Based on this, a two-layer sampling mesh was constructed, in which the coarse mesh generates surface viewpoint candidates, forming a set of surface viewpoints. Fine grids generate dense sampling points, forming a coverage evaluation set. It is used to quantitatively evaluate the coverage capability of a viewpoint.

7. The method according to claim 2, characterized in that, Steps d) and e) include: Sampling points are selected at the effective parameter locations corresponding to the coarse sampling grid, and initial surface candidate viewpoints are generated by local normal vector regularization and normal offset. The initial candidate surface viewpoints are subjected to inverse kinematics solution and collision detection. Candidate viewpoints with collision-free inverse kinematics solutions are retained to form an effective set of surface viewpoints. The upper limit of the number of backbone viewpoints is determined based on the total area of ​​the surface under test, the field of view coverage area of ​​a single viewpoint, and the field of view overlap coefficient. Backbone viewpoints are then selected sequentially from the effective surface viewpoint set according to the marginal coverage gain until the upper limit is reached or the maximum marginal coverage gain is lower than a preset threshold. Obtain the backbone viewpoint set The remaining valid surface viewpoints that were not selected constitute the opportunity viewpoint set. .

8. The method according to claim 1, characterized in that, Step 2, the online planning, includes: a) Determining the reference viewpoint: Based on the robot's current state, the covered area, and the remaining forced viewpoints and remaining backbone viewpoints, determine the reference viewpoint that has a global guiding role; b) Baseline viewpoint adjustment: Around the said baseline viewpoint, in the opportunity viewpoint set A local opportunity window is constructed, and the decision on whether to replace the baseline viewpoint with an opportunity viewpoint is made based on the coverage benefits and motion costs. c) Viewpoint path generation: After selecting a viewpoint, update the robot state, coverage state, and remaining task set, and trigger the next round of planning to dynamically generate the robotic arm viewpoint execution sequence.

9. The method according to claim 8, characterized in that, The determination of the reference viewpoint in step a) includes: The online decision-making process is modeled as a Markov decision process, and Monte Carlo tree search is used to determine the baseline viewpoint. The state includes at least the robot's current joint angle, the set of sampled points on the covered surface, the set of remaining forced viewpoints, and the set of remaining backbone viewpoints. The action space is defined by the set of remaining forced viewpoints. and the remaining backbone viewpoint set Composition. A reward function is constructed that simultaneously considers new coverage, proxy motion costs, and forced viewpoint priority. The reward function is: ； in, Indicates the execution of an action The resulting increase in coverage Indicates the cost of proxy movement. For indicator functions, This is used to enhance the forced viewpoint priority. A Monte Carlo tree search is employed, involving selection, expansion, simulation, and backpropagation processes, to determine the reference viewpoint at the current moment. .

10. The method according to claim 8, characterized in that, The reference viewpoint adjustment and viewpoint path generation described in steps b) and c) include: In obtaining the reference viewpoint Subsequently, around this baseline viewpoint, in the opportunity viewpoint set Constructing a local opportunity window. First, a coarse screening is performed based on the agent's motion cost, retaining those that meet the requirements. The opportunity viewpoints are then precisely validated based on the actual path cost output by the motion planner to obtain candidate viewpoints within the opportunity window. The opportunity window relaxation parameters are adaptive. ； in, As a coverage adjustment factor, As the baseline relaxation value, Coverage rate at the current moment; For target coverage; Sensitivity coefficient; This is the upper bound of the regulation factor; To perform the positive part operation, For the remaining Number of tasks. When hour, ; hour, ; For candidate viewpoints within the opportunity window Based on its marginal coverage gain Motion cost from the current state to the candidate viewpoint Utility ratio of construction: And select the opportunity viewpoint with the highest utility ratio. ; when , At that time, the opportunity viewpoint Otherwise, execute the baseline viewpoint. ; After execution, update the robot joint configuration and the set of covered surface sampling points. When the viewpoint actually executed belongs to the remaining forced viewpoint set or the remaining backbone viewpoint set, remove the viewpoint from the corresponding remaining task set until all forced viewpoints are completed and the surface coverage reaches the preset target.

Images