Training method, spraying method and system for intelligent spraying model of curved surface

Abstract

This invention provides a method for training a curved surface intelligent spraying model, a spraying method, and a system thereof, relating to the field of robot path planning technology. The curved surface intelligent spraying method includes: acquiring point cloud data of an object using a depth camera; performing voxel filtering, statistical filtering, and RANSAC plane segmentation preprocessing on the point cloud data; calculating the centroid and PCA principal component orientation of the preprocessed point cloud; rasterizing the space containing the point cloud according to a preset grid size and number of rows and columns to generate a grid map; using a DQN model for path planning; and having the robot perform spraying operations according to the planned path. This invention also provides a method for training a curved surface intelligent spraying model and a curved surface intelligent spraying system to achieve optimal path coverage of all sprayable areas, avoiding unnecessary movement.

Description

Technical Field

[0001] This invention belongs to the field of robot path planning technology, and specifically relates to a training method, spraying method and system for intelligent surface spraying model. Background Technology

[0002] In the field of industrial automation, robot path planning is a critical issue, especially in scenarios requiring comprehensive surface coverage operations, such as painting and inspection. Traditional path planning methods are typically based on pre-defined rules or models, making them difficult to adapt to complex and ever-changing environments and task requirements.

[0003] Currently, most painting robots use teach-in painting, which involves painting according to a trajectory pre-set by technicians. This method is effective based on the experience of previous painters, but it requires a lot of programming work in the early stages, the trajectory optimization is not obvious, the efficiency is low, and the paint waste is relatively large.

[0004] With the development of deep reinforcement learning technology, it has shown great potential in solving complex decision-making problems. However, research on applying deep reinforcement learning to point cloud processing and path planning is still in its early stages, and effective systems and methods are currently lacking.

[0005] Therefore, in this technical field, when faced with unknown objects, existing technologies have not yet achieved an effective path for intelligent spraying, thus failing to achieve comprehensive coverage of the object's surface. Summary of the Invention

[0006] Therefore, the technical problem to be solved by the present invention is to provide a training method, a spraying method and system for a curved surface intelligent spraying model, which is based on a reinforcement learning strategy to cover all sprayable areas with the optimal path and avoid invalid movement.

[0007] In a first aspect, the present invention provides a method for training a curved surface intelligent spraying model, comprising the following steps:

[0008] Step S11, collect workpiece point cloud data:

[0009] Point cloud data of the workpiece is acquired using a depth camera;

[0010] Step S12, Point cloud data preprocessing and rasterization:

[0011] The collected workpiece point cloud data is preprocessed to remove noise points and outliers; the preprocessed point cloud data is then rasterized to generate a raster map.

[0012] Step S13, train the DQN (Deep Q-Network) model:

[0013] Define the environment state as a raster map;

[0014] The robot's motion space includes moving up, down, left, and right, as well as jumping;

[0015] The reward mechanism encourages robots to cover all painting areas and reduce repeated visits;

[0016] A multi-layer fully connected neural network is used as the DQN. The network collects experience replay buffers by interacting with the environment and uses the experience replays to train the network and learn the optimal path planning strategy.

[0017] Furthermore, step S12 specifically includes:

[0018] Step S121: Perform voxel filtering and statistical filtering on the collected workpiece point cloud data in sequence;

[0019] Step S122: Plane segmentation is performed using RANSAC (Random Sample Consensus);

[0020] Step S123: Calculate the centroid of the point cloud and perform feature extraction using PCA (Principal Component Analysis);

[0021] Step S124: Based on the PCA results, project the point cloud data onto a two-dimensional plane;

[0022] Step S125: Divide the grid according to the preset grid size and number of rows and columns;

[0023] Step S126: Construct a raster map based on the point cloud density within the raster.

[0024] Furthermore, in step S13, the action space A is:

[0025] A = {a0, a1, a2, a3, a4, a5}

[0026] Where a0 moves up, a1 moves down, a2 moves left, a3 moves right, a4 jumps to the top of the next column, and a5 jumps to the bottom of the next column;

[0027] The reward is set as follows:

[0028]

[0029] in,

[0030] r new Rewards for unvisited areas;

[0031] r direction For unvisited areas, vertical movement is rewarded, while horizontal movement is penalized.

[0032] r jump Penalty for performing a jump action;

[0033] r repeat Penalties apply to repeated visits to already covered areas or to remaining in the same spot.

[0034] r boundary There is no additional reward or penalty for moving outside the boundary.

[0035] Furthermore, the reward settings are as follows:

[0036] r new =15, r repeat =-20,

[0037]

[0038]

[0039] Secondly, the present invention provides a method for intelligent spraying of curved surfaces, comprising the following steps:

[0040] Step S21: Collect point cloud data of the object using a depth camera;

[0041] Step S22: Perform voxel filtering, statistical filtering, and RANSAC plane segmentation preprocessing on the point cloud data in sequence.

[0042] Calculate the centroid and PCA principal component orientation of the preprocessed point cloud, and rasterize the space where the point cloud is located according to the preset grid size and number of rows and columns to generate a raster map;

[0043] Step S23: Perform path planning using the DQN model trained by any of the training methods described in the first aspect above;

[0044] In step S24, the robot performs the spraying operation according to the planned path.

[0045] Furthermore, step S24 specifically includes:

[0046] Acquire point cloud data of the grid; calculate the position of the center point coordinates of the grid in the coordinate system of the robot arm base; determine the spraying posture by calculating the point cloud normal vector of the grid through PCA, and form a pose sequence.

[0047] The pose sequence is transformed to generate joint space instructions for the robotic arm, and the robot performs the painting operation according to the joint space instructions.

[0048] Furthermore, the transformation sequence for generating the robotic arm joint space instructions includes the transformation matrix T from the camera coordinate system to the base coordinate system. camera_to_base Calculated using the following formula:

[0049] T camera_to_base =T end_effect0r ×T camera_to_end_effector

[0050] Among them, T end_effector For the pose of the robotic arm's end effector, T camera_to_end_effector The pose of the camera relative to the end effector;

[0051] The coordinates p of the center point of the raster point cloud in the base coordinate system base for:

[0052]

[0053] Where (x,y,z) are the coordinates of the center point of the raster point cloud in the camera coordinate system.

[0054] Furthermore, the instructions for generating the robotic arm joint space by transforming the pose sequence also include:

[0055] 1) Determine the target's Z-axis direction:

[0056]

[0057] Where normal is the point cloud normal vector in the raster, and targetZ is the Z-axis of the target;

[0058] If targetZ > 0, then reverse the direction:

[0059] targetZ = -targetZ

[0060] 2) Project the X-axis of the current robotic arm posture onto a plane perpendicular to the Z-axis of the target. The calculation formula is as follows:

[0061] newX=currentX-(currentX·targetZ)targetZ

[0062] Where newX is the projected X-axis and currentX is the X-axis of the current robotic arm posture;

[0063] 3) Calculate the orthogonal Y-axis

[0064] Calculate the orthogonal Y-axis using the cross product method:

[0065] newY = targetZ × newX

[0066] Normalize newY:

[0067]

[0068] 4) Construct a new rotation matrix

[0069] Combine the calculated newX, newY, and targetZ into a new rotation matrix:

[0070] Rnew=(newX,newY,newZ)

[0071] 5) Convert to ZYZ Euler angles

[0072] The new rotation matrix is ​​converted into ZYZ Euler angles to control the robot arm's posture:

[0073] α,β,γ=rotationMatrixToZYZ(Rnew)

[0074] Where α, β, and γ are ZYZ Euler angles, respectively.

[0075] Thirdly, the present invention also provides a curved surface intelligent spraying system, comprising:

[0076] The point cloud data acquisition module acquires point cloud data of objects using a depth camera;

[0077] The point cloud data preprocessing and rasterization module performs voxel filtering, statistical filtering, and RANSAC plane segmentation preprocessing on the point cloud data in sequence.

[0078] Calculate the centroid and PCA principal component orientation of the preprocessed point cloud, and rasterize the space where the point cloud is located according to the preset grid size and number of rows and columns to generate a raster map;

[0079] The path planning module uses a pre-trained DQN model for path planning.

[0080] In the spraying module, the robot performs spraying operations according to the planned path.

[0081] Furthermore, the system is based on the ROS framework, and the modules communicate through topic publishing and subscription; robot control communicates with the Modbus server via the Modbus TCP protocol using the IP address and port.

[0082] Beneficial effects:

[0083] The surface intelligent spraying model training method provided by this invention is based on reinforcement learning strategy. The data acquisition, data processing and training process steps are clear and highly operable. When the trained model is performing surface intelligent spraying, it can solve an effective path when facing unknown objects, making the spraying more intelligent.

[0084] The reward mechanism encourages the robot to cover all paintable areas and reduce repeated visits. The motion design adds the ability to jump to the top and bottom of the next column of paintable points to adapt to complex grid map structures, improving the flexibility and efficiency of path planning.

[0085] The intelligent spraying method for curved surfaces provided by this invention requires no programming teaching and no pre-made point cloud templates. The model trained by the intelligent spraying model training method for curved surfaces achieves intelligent spraying.

[0086] The intelligent spraying system for curved surfaces provided by this invention has the same technical effects as described above. Attached Figure Description

[0087] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0088] Figure 1 This is a flowchart of Embodiment 1 of the present invention;

[0089] Figure 2 This is a flowchart of Embodiment 2 of the present invention;

[0090] Figure 3 Point cloud data of an object was collected in Embodiment 2 of the present invention;

[0091] Figure 4 This is a grid map of the object in Embodiment 2 of the present invention;

[0092] Figure 5 This is the spraying path of the object in Embodiment 2 of the present invention. Detailed Implementation

[0093] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. The principles and features of the present invention are described below with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other. The embodiments given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0094] Example 1

[0095] like Figure 1 As shown, this embodiment is a training method for a curved surface intelligent spraying model based on a reinforcement learning strategy, including:

[0096] Hardware environment setup: In the industrial spraying robot system, 3D sensors (such as machine vision cameras) are used to collect workpiece point cloud data. The industrial robot is equipped with a spraying end effector, and communication and control between modules are realized through the ROS (Robot Operating System) framework.

[0097] Software system initialization: Start the point cloud data processing node (pcd_deal_node), subscribe to point cloud topics published by the 3D sensor, load the deep reinforcement learning path planning service (DQNpathplan_node), and initialize the DQN model parameters. Start the robotic arm control client and establish a communication connection with the robotic arm controller.

[0098] Point cloud data processing and rasterization: 3D sensors acquire workpiece point cloud data and publish it to the ROS topic. `pcd_deal_node` receives the point cloud data and performs preprocessing operations such as voxel filtering, statistical filtering, and RANSAC plane segmentation. It then calculates the centroid and PCA principal component orientation of the preprocessed point cloud, rasterizes it according to the preset grid size and number of rows and columns, generates a raster map, and publishes it to the raster topic.

[0099] Path planning request and response: The DQNpathplan_node listens to the grid topic and receives grid map data. It constructs a GridEnv (grid environment simulator) environment, using the grid map as the initial state, and calls the trained DQN model for path planning. The DQN model outputs the optimal action sequence based on the current state, maps it to a grid index sequence as the planned path, and publishes it to the path topic.

[0100] Point cloud data processing specifically includes:

[0101] 1) Perform voxel filtering and statistical filtering on the point cloud data in sequence to remove noise points and outliers;

[0102] 2) The background plane is separated using the Random Sample Consensus (RANSAC) algorithm, and the workpiece point cloud data is extracted;

[0103] 3) Feature extraction, calculation of centroids and principal component analysis (PCA) of point clouds to provide a basis for subsequent rasterization processing;

[0104] 4) Based on the PCA results, project the point cloud data onto a two-dimensional plane;

[0105] 5) Divide the grid according to the preset grid size and number of rows and columns;

[0106] 6) Analyze the point cloud density within the raster and construct a raster map.

[0107] Deep reinforcement learning path planning specifically includes:

[0108] The grid environment simulator (GridEnv) is used to treat the grid map as the environment state. The robot's motion space is defined, including up, down, left and right movement and jumping. A reward mechanism is set to encourage the robot to cover all paintable areas and reduce repeated visits.

[0109] DQN training is performed using a multi-layer fully connected neural network. The input is the grid state, and the output is the action value. The network collects experience replay buffers by interacting with the environment and uses the experience replays to train the network and learn the optimal path planning strategy.

[0110] The deep reinforcement learning path planning incorporates novel action and reward mechanisms, specifically including:

[0111] Action design: In addition to the traditional up, down, left, and right movement actions, jumping actions have been added, including jumping to the top of the next column and jumping to the bottom of the next column to adapt to the complex grid map structure and improve the flexibility and efficiency of path planning.

[0112] For example, the action space is:

[0113] A = {a0, a1, a2, a3, a4, a5}

[0114] Where a0 moves upward, a1 moves downward, a2 moves left, a3 moves right, a4 jumps to the top of the next column of paintable points, and a5 jumps to the bottom of the next column of paintable points.

[0115] The reward mechanism includes: a reward for covering unvisited areas, an additional reward for vertical movement, and a penalty for repeatedly visiting covered areas and for lateral movement, guiding the robot to cover all paintable areas along the optimal path while avoiding ineffective movement.

[0116] The reward is set as follows:

[0117]

[0118] Where, r new =15, r repeat =-20, r boundary =No additional reward or penalty (position remains unchanged)

[0119]

[0120] Example 2

[0121] This embodiment presents a method for intelligent spraying of curved surfaces based on a reinforcement learning strategy. This method uses the model trained in Embodiment 1 to perform intelligent spraying on curved surfaces, such as... Figures 2-5 As shown, the method includes the following steps:

[0122] Step S21: Collect point cloud data of the object using a depth camera;

[0123] Step S22: Perform voxel filtering, statistical filtering, and RANSAC plane segmentation preprocessing on the point cloud data in sequence.

[0124] Calculate the centroid and PCA principal component orientation of the preprocessed point cloud, and rasterize the space where the point cloud is located according to the preset grid size and number of rows and columns to generate a raster map;

[0125] Step S23: Use the trained DQN model to perform path planning;

[0126] In step S24, the robot performs the spraying operation according to the planned path.

[0127] In step S24, the robotic arm control and spraying execution includes: pcd_deal_node receiving the planned path and obtaining the corresponding grid center point cloud data based on the path index. The pose of the grid's point cloud center point coordinates in the robotic arm base coordinate system is calculated, and the spraying posture is determined by combining the PCA normal vector. The pose sequence is converted into robotic arm joint space commands, and the robotic arm control service is invoked to drive the robot to perform the spraying operation according to the planned path.

[0128] Robotic arm control requires a transformation from the camera coordinate system to the robotic arm base coordinate system.

[0129] The pose of the robotic arm's end effector (end_effector_pose) and the camera's pose relative to the end effector (camera_to_end_effector_pose) are known. We need to transform the point cloud data in the camera coordinate system to the robotic arm's base coordinate system.

[0130] Assume: The pose of the robotic arm's end effector is denoted as Tend_effector, represented by a homogeneous transformation matrix. The pose of the camera relative to the end effector is denoted as Tcamera_to_end_effector, also represented by a homogeneous transformation matrix.

[0131] Transformation matrix T from camera to base camera_to_base Calculated using the following formula:

[0132] T camera_to_base =T end_effector ×T camera_to_end_effector

[0133] Among them, T end_effector For the pose of the robotic arm's end effector, T camera_to_end_effector The pose of the camera relative to the end effector;

[0134] The coordinates of the center point of the point cloud in the base coordinate system are p base for:

[0135]

[0136] Where (x,y,z) are the coordinates of the center point of the point cloud in the camera coordinate system.

[0137] Then, the normal vector is adjusted:

[0138] The commands for generating the robotic arm joint space include:

[0139] 1) Determine the target's Z-axis direction:

[0140] The Z-axis direction of the target is determined by the normal vector of the point cloud, ensuring that it faces downwards:

[0141]

[0142] If targetZ > 0, then reverse the direction:

[0143] targetZ = -targetZ

[0144] 2) Project the perpendicular plane from the current X-axis to the target Z-axis:

[0145] Project the X-axis of the current attitude onto a plane perpendicular to the Z-axis of the target. The calculation formula is as follows:

[0146] newX=currentX-(currentX·targetZ)targetZ

[0147] Where newX is the projected X-axis and currentX is the X-axis of the current robotic arm posture;

[0148] 3) Addressing degradation:

[0149] If the projected X-axis is close to a zero vector (i.e., |newX| < 1e-6), then use the Y-axis of the current pose for projection instead:

[0150] newX=currentY-(currentY·targetZ)targetZ

[0151] If it still degenerates, then project using the global X-axis:

[0152] newX=globalX-(globalX·targetZ)targetZ

[0153] Finally, normalize newX:

[0154]

[0155] 4) Calculate the orthogonal Y-axis:

[0156] Calculate the orthogonal Y-axis using the cross product method: newY = targetZ × newX

[0157] And normalize newY:

[0158] 5) Construct a new rotation matrix:

[0159] Combine the calculated newX, newY, and targetZ into a new rotation matrix:

[0160] Rnew=(newX,newY,newZ)

[0161] 6) Convert to ZYZ Euler angles:

[0162] The new rotation matrix is ​​converted into ZYZ Euler angles to control the robot arm's posture:

[0163] α,β,γ=rotationMatrixToZYZ(Rnew)

[0164] α, β, and γ are ZYZ Euler angles, respectively.

[0165] After adjusting the above posture, the following objectives are achieved:

[0166] The Z-axis is parallel and downward: Through the above adjustment of the normal vector and construction of the rotation matrix, it is ensured that the Z-axis of the end of the robotic arm is parallel to the normal vector and downward.

[0167] Reduce X and Y axis rotation: When constructing the rotation matrix, keep the X and Y axis directions consistent with the base coordinate system as much as possible, thereby reducing the rotation angle of the robotic arm's end effector on the X and Y axes. This prevents the camera from changing direction as the robotic arm rotates, ensuring the camera's stability during the painting process.

[0168] The converted pose data and Euler angles are then populated into the robot arm control command message:

[0169] data=[xbase,ybase,zbase,α,β,γ]

[0170] By calling the ROS service, control commands are sent to the robotic arm controller, driving the robotic arm to perform the spraying operation according to the planned path and posture.

[0171] Example 3

[0172] The point cloud data acquisition module acquires point cloud data of objects using a depth camera;

[0173] The point cloud data preprocessing and rasterization module performs voxel filtering, statistical filtering, and RANSAC plane segmentation preprocessing on the point cloud data in sequence.

[0174] Calculate the centroid and PCA principal component orientation of the preprocessed point cloud, and rasterize the space where the point cloud is located according to the preset grid size and number of rows and columns to generate a raster map;

[0175] The path planning module uses a pre-trained DQN model for path planning.

[0176] In the spraying module, the robot performs spraying operations according to the planned path.

[0177] The system is based on the ROS framework, and the modules communicate through topic publishing and subscription; robot control communicates with the Modbus server via the Modbus TCP protocol and the IP address and port.

[0178] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for curved surface intelligent spraying, characterized in that, Comprising the following steps: Step S21, collecting point cloud data of the object by a depth camera; Step S22, sequentially performing voxel filtering, statistical filtering and RANSAC plane segmentation preprocessing on the point cloud data; Calculate the centroid and PCA principal component direction of the preprocessed point cloud, and divide the space where the point cloud is located into a grid according to the preset grid size and row and column number to generate a grid map; Step S23, using the trained DQN model for path planning; Step S24, the robot performs spraying operation according to the planned path; Step S24 specifically comprises: Obtaining the point cloud data of the grid; calculating the position of the point cloud center point coordinate of the grid in the base coordinate system of the robot arm; determining the spraying posture by calculating the point cloud normal vector of the grid through PCA to form a pose sequence; Converting the pose sequence to generate a robot arm joint space instruction, and the robot performs spraying operation according to the joint space instruction; The converted pose sequence generates a robot joint space instruction, which includes a transformation matrix from a camera coordinate system to a base coordinate system This is calculated by the following formula: ； wherein, is a pose of the end effector of the robot arm, is a pose of the camera relative to the end effector; The coordinate of the center point of the point cloud of the grid in the base coordinate system Is: ； wherein, is the coordinate of the center point of the point cloud of the grid in the camera coordinate system; Converting the pose sequence to generate a robot arm joint space instruction further comprises: 1) determining the target Z-axis direction: ； wherein, is the point cloud normal vector in the grid, is the target Z-axis; If then reverse direction: ； 2) projecting the X-axis of the current robot arm posture to a plane perpendicular to the target Z-axis, and the calculation formula is as follows: ； wherein, is the X axis of the projection, is the X axis of the current robot pose; 3) calculating the orthogonal Y-axis: Calculate the orthogonal Y-axis by cross multiplication: ； right Normalize: ； 4) constructing a new rotation matrix: The calculated , and are combined into a new rotation matrix: ； 5) converting to ZYZ Euler angles: Convert the new rotation matrix to ZYZ Euler angles for controlling the posture of the robot arm: ； wherein respectively are ZYZ Euler angles.

2. The method of claim 1, wherein, In the step S23, the training process of the DQN model comprises: Step S231, collecting workpiece point cloud data: Collecting point cloud data of the workpiece by a depth camera; Step S232, point cloud data preprocessing and gridding: Preprocessing the collected workpiece point cloud data to remove noise points and outliers; gridding the preprocessed point cloud data to generate a grid map; Step S233, training the DQN model: Defining the environment state as a grid map; The action space of the robot includes up, down, left, right movement and jumping; The reward mechanism is to encourage the robot to cover all spraying areas and reduce repeated visits; A multi-layer fully connected neural network is used as the DQN, experience replay buffer is collected through interaction with the environment, the network is trained using experience replay, and the optimal path planning strategy is learned.

3. The method of claim 2, wherein, Step S232 specifically comprises: Step S2321, sequentially performing voxel filtering and statistical filtering on the collected workpiece point cloud data; Step S2322, performing plane segmentation by RANSAC; Step S2323, calculating the centroid of the point cloud and performing feature extraction by PCA; Step S2324, projecting the point cloud data to a two-dimensional plane according to the PCA result; Step S2325, gridding according to the preset grid size and row and column number; Step S2326, constructing a grid map according to the point cloud density in the grid.

4. The method of claim 2, wherein, In step S233, the action space A is: ； wherein, is to move up, is to move down, is to move left, is to move right, is to jump to the top of the next column, is to jump to the bottom of the next column; The reward R is set as: ； Wherein, : reward for unvisited areas; : the moving direction to the unvisited area is rewarded for longitudinal movement and penalized for lateral movement; : penalize for performing a jump action; : penalty for repeated access to covered areas or staying in place; : No extra reward or penalty for moving outside the bounds.

5. The method of claim 4, wherein, The reward setting specifically comprises: =15， =-20， ， 。 6. A curved smart spraying system employing the method of any one of claims 1-5, wherein, Comprising: A point cloud data collection module collects point cloud data of an object by a depth camera; A point cloud data preprocessing and gridding module sequentially performs voxel filtering, statistical filtering and RANSAC plane segmentation preprocessing on the point cloud data; The centroid and the PCA principal component direction of the pre-processed point cloud are calculated, and the space where the point cloud is located is rasterized and divided according to a preset grid size and row and column number, so as to generate a grid map; The path planning module uses the trained DQN model to plan a path; The spraying operation module is used for spraying operation by the robot according to the planned path.

7. The spray system of claim 6, wherein, The system is based on the ROS framework, and each module communicates through topic publishing and subscription; the robot control communicates with the Modbus server with an IP address and a port through the Modbus TCP protocol.

Images