Rebar binding robot planning method based on visual semantic constraint

Abstract

The planning method of reinforcement binding mechanical arm based on visual semantic constraint, the algorithm first constructs the visual perception system composed of high-precision six-degree-of-freedom mechanical arm, end binding executor, visual sensor and industrial camera. Then, the steel semantic understanding and key point detection module is developed, the multi-modal deep vision network is fused with the RGB image and the depth map information, the high-precision semantic segmentation of the reinforcement grid and the binding node key point detection are realized through data preprocessing, feature fusion and other steps, and the binding point coordinate in the world coordinate system is output. Then, the dynamic task decomposition and reinforcement learning decision mechanism is designed, the global binding task is converted into local action sequence, and the path planning and binding sequence are optimized through reinforcement learning. Finally, according to the optimal action sequence output by the reinforcement learning, the adaptive path planning is generated and the precise binding operation is executed. The application effectively improves the automation level of reinforcement binding, and solves the deficiencies of the prior art in visual perception, task planning and path control.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to a planning method for a rebar tying robotic arm based on visual semantic constraints. Background Technology

[0002] With the rapid development of the construction industry, reinforced concrete structures are widely used due to their excellent mechanical properties and economic efficiency. However, traditional rebar tying work mainly relies on manual labor, which is not only labor-intensive and inefficient, but also makes it difficult to guarantee consistent quality, especially when dealing with complex geometries and large-scale construction projects. Therefore, developing a system capable of automating rebar tying is of paramount importance.

[0003] In recent years, advancements in robotics and computer vision have offered a solution to these problems. By combining a high-precision six-DOF robotic arm with an advanced visual perception system, automated identification and precise manipulation of rebar mesh can be achieved. While some research has explored the use of robots for rebar tying, most solutions exhibit poor adaptability to different rebar mesh specifications and cannot flexibly adjust tying strategies to cope with changing working environments.

[0004] Furthermore, existing rebar tying robots often lack effective dynamic task decomposition mechanisms and path planning algorithms, resulting in low efficiency and a high risk of collisions when performing tying tasks. To improve the intelligence level and work efficiency of rebar tying robots, it is necessary to develop a rebar tying robotic arm system that integrates advanced technologies such as visual perception, semantic understanding, key point detection, dynamic task decomposition, and reinforcement learning decision-making.

[0005] This patent proposes a rebar tying robotic arm and its control method based on a multimodal deep vision network. This system not only achieves high-precision semantic segmentation of the rebar mesh and detection of key points in the tying nodes, but also optimizes the tying sequence and path planning through a dynamic task decomposition mechanism and reinforcement learning algorithms, thereby significantly improving the automation level and execution efficiency of rebar tying operations. This invention is of great significance for promoting the intelligent development of the construction industry. Summary of the Invention

[0006] To address the above problems, this invention proposes a planning method for a rebar tying robotic arm based on visual semantic constraints. The specific steps are as follows:

[0007] Step 1: Constructing the visual perception system for the rebar tying robotic arm

[0008] To meet the needs of automated horizontal rebar tying operations, a high-precision six-degree-of-freedom robotic arm, an end effector for tying, and a vision sensor are integrated, with industrial cameras deployed above and to the side of the robotic arm's working area.

[0009] Step 2: Develop a module for semantic understanding and key point detection of reinforcing bars.

[0010] In the development of the rebar semantic understanding and key point detection module, RGB image and depth map information are integrated to achieve high-precision semantic segmentation of rebar mesh and key point detection of binding nodes, and automatically output the coordinates of binding points.

[0011] Step 3: Design a dynamic task decomposition and reinforcement learning decision-making mechanism

[0012] Based on the binding point coordinates output in step 2, a dynamic task decomposition mechanism is introduced to automatically transform the global binding task into a local action sequence; the path planning and binding sequence of the robotic arm are optimized through reinforcement learning algorithm to achieve adaptive operation of steel mesh of different specifications.

[0013] Step 4: Integrating Adaptive Path Planning

[0014] The optimal motion sequence output by reinforcement learning is used to generate the real-time motion trajectory of the robotic arm, ensuring that the path planning avoids obstacles and optimizes the binding sequence; the real-time motion controller is used to execute the local motion sequence to control the end effector of the robotic arm to complete the precise binding operation.

[0015] The present invention provides a planning method for a rebar tying robotic arm based on visual semantic constraints. The technical advantages of the present invention are as follows:

[0016] 1. This invention significantly reduces the time and effort required for manual rebar tying by automating the process. Utilizing a dynamic task decomposition mechanism and reinforcement learning algorithms to optimize the tying sequence and path planning effectively reduces unnecessary movement and adjustment time, thereby improving the efficiency of the entire tying process.

[0017] 2. This invention combines a high-precision six-degree-of-freedom robotic arm with an advanced vision perception system, enabling high-precision semantic segmentation and key point detection of rebar mesh. This ensures that each binding point can be accurately identified and bound, improving the consistency and reliability of binding quality.

[0018] 3. This invention promotes the development of the construction industry towards intelligentization, laying the foundation for building a more intelligent and efficient construction process in the future. With continuous technological advancements and improvements, its application scope is expected to expand further, including but not limited to the construction of large-scale infrastructure projects such as bridges and tunnels. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention;

[0020] Figure 2 This is a diagram of the reinforcement learning model of the present invention. Detailed Implementation

[0021] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0022] This invention proposes a planning method for a rebar tying robotic arm based on visual semantic constraints. By fusing RGB images and depth maps, it achieves high-precision semantic segmentation of the rebar mesh and detection of tying nodes. Combining dynamic task decomposition and deep reinforcement learning, it optimizes the tying sequence and path planning. Ultimately, it drives a six-degree-of-freedom robotic arm to complete adaptive, high-precision, and high-efficiency automatic tying operations, significantly improving the level of intelligent construction. The invention flowchart is shown below. Figure 1 As shown, the steps of the present invention will be described in detail below.

[0023] To address the above problems, this invention proposes a planning method for a rebar tying robotic arm based on visual semantic constraints. The specific steps are as follows:

[0024] Step 1: Constructing the visual perception system for the rebar tying robotic arm

[0025] To address the need for automated horizontal rebar tying operations, a high-precision six-degree-of-freedom robotic arm, an end effector for tying, and a vision sensor are integrated, with industrial cameras deployed above and to the side of the robotic arm's working area.

[0026] Step 2: Develop a module for semantic understanding and key point detection of reinforcing bars.

[0027] In the development of the rebar semantic understanding and key point detection module, RGB image and depth map information are integrated to achieve high-precision semantic segmentation of rebar mesh and detection of key points of binding nodes, and automatically output the coordinates of binding points.

[0028] Based on the RGB images and depth maps acquired by the visual perception system in step 1, a multimodal depth vision network is used to achieve semantic segmentation that distinguishes rebar from the background and high-precision detection of binding target points. Finally, the coordinates of the binding points in the world coordinate system that the robotic arm can recognize are output.

[0029] Input data: Synchronized RGB images captured by the industrial-grade camera in step 1. With depth map .

[0030] Step 2.1: Data Preprocessing and Alignment

[0031] Step 2.1.1 Image Normalization

[0032] right To normalize pixel values, use the following formula:

[0033]

[0034] in, The normalized RGB image, Image pixel coordinates, The global mean of the RGB image. The global standard deviation of the RGB image. For RGB images The pixel value.

[0035] The following formula will be used to... Normalize to the [0,1] interval:

[0036]

[0037] in, This is the normalized interval depth map. , These are the minimum and maximum depth values ​​for a single depth map, respectively. For depth images in The pixel value.

[0038] Step 2.1.2 Obtain camera calibration parameters

[0039] Due to differences in camera deployment locations, the depth map pixel coordinates are mapped to the RGB image coordinate system using an extrinsic parameter matrix for unification. The formula is as follows:

[0040]

[0041] in, These are the pixel coordinates of the depth map. This represents the depth value corresponding to that coordinate. For the pixel coordinates of the RGB image, the intrinsic parameters for camera calibration are obtained using the above formula. and external references , Let be a rotation matrix. It is a translation vector.

[0042] Step 2.2: Deployment and Feature Fusion of Multimodal Deep Vision Networks

[0043] It adopts a dual-branch feature extraction and cross-modal fusion architecture, loads a pre-trained network model, and fuses RGB and depth map features to provide high-dimensional feature support for semantic segmentation and key point detection.

[0044] The feature extraction branch consists of an RGB branch and a depth map branch. The RGB branch uses ResNet50 as the backbone, and the depth map branch uses DenseNet121 as the backbone. The last fully connected layer is removed from both branches, retaining the convolutional feature output. The RGB image and depth image are taken as inputs respectively, and the output is an RGB feature map. and depth feature map Finally, a channel concatenation and attention-weighted fusion strategy is used to fuse the features, as shown in the formula:

[0045]

[0046] in, For channel-dimensional splicing, for convolution.

[0047] Step 2.3: Semantic segmentation of reinforcing bars and detection of tying nodes

[0048] From fusion characteristics The steel reinforcement area is segmented out, and the coordinates of the binding node in the image coordinate system are located.

[0049] Step 2.3.1 Semantic segmentation of reinforcing bars

[0050] The semantic segmentation head structure for rebar uses a U-Net decoder to fuse features. As input, the output is a segmentation probability map. The loss function is jointly optimized using Dice loss and binding entropy loss, and the formula is as follows:

[0051]

[0052] in, For binding entropy loss; The Dice loss is calculated using the following formula:

[0053]

[0054] in, The supervised annotation data is used to train the rebar semantic segmentation model. The annotation is the segmentation mask, with the rebar region set to 1 and the background set to 0. Avoid denominators of 0; A threshold of 0.7 is used to generate a binary mask. The specific formula for the binding entropy loss is as follows:

[0055]

[0056] Step 2.3.2 Key Point Detection of Binding Nodes

[0057] The lap joint is defined as the intersection of horizontal and vertical reinforcing bars, with the intersection point being more than 5mm from the edge of the reinforcing bar. The following method is used for inspection:

[0058] Step 2.3.2.1 Determine the direction of the reinforcing bars using Hough line detection, based on the segmentation probability map. Extracting horizontal reinforcement masks and vertical reinforcement mask ;

[0059] Step 2.3.2.2 Calculate the intersection mask Filter connected components with an area ≥ 25 pixels ;

[0060] Step 2.3.2.3 Calculate the centroid coordinates for each connected component. , which is the coordinate of the binding node in the image coordinate system;

[0061] Step 2.4: Coordinate transformation and output of binding points

[0062] The coordinates of the binding nodes in the image coordinate system are converted into world coordinates that the robotic arm can recognize.

[0063] Step 2.4.1 Coordinate Transformation Formula

[0064] Internal parameters based on camera calibration and external references Image coordinates Convert to world coordinates The formula is:

[0065]

[0066] in, This represents the depth value of the tethered node in the camera coordinate system. The inverse of the intrinsic parameter matrix. Let be a rotation matrix. The translation vector is obtained from step 2.1.2; These are the homogeneous pixel coordinates of the image; These are the homogeneous coordinates of the camera.

[0067] The world coordinate system takes the center of the robotic arm base as the origin, the X-axis along the length of the horizontal reinforcing bars, the Y-axis along the length of the vertical reinforcing bars, and the Z-axis perpendicular to the ground and upwards.

[0068] Step 3: Design a dynamic task decomposition and reinforcement learning decision-making mechanism

[0069] Based on the binding point coordinates output in step 2, a dynamic task decomposition mechanism is introduced to automatically transform the global binding task into a sequence of local actions. A reinforcement learning algorithm is used to optimize the path planning and binding sequence of the robotic arm, enabling adaptive operation on steel meshes of different specifications. The reinforcement learning model diagram is shown below. Figure 2 As shown.

[0070] Step 3.1: Implementation of Dynamic Task Decomposition Mechanism

[0071] The coordinates of the binding points output in step 2 Combine the coordinates to obtain the set S of the binding point coordinates. Defined as ,Will Defined as the safety coordinates of the binding point , To determine the offset of the safety coordinates on the z-axis, the binding task of traversing all points in set S is transformed into a sequence of local actions that the robotic arm can execute, clearly defining the goal and constraints of each action.

[0072] Step 3.1.1 Defining the Action Space

[0073] Define four basic types of movements to form a movement sequence. :

[0074] Movement From current position Reach a safe location near the target point The parameters are Coordinates, constraint: motion speed ≤ 0.5m / s.

[0075] Positioning Action :from Descend to the lashing point The parameters are Coordinates, constraints: positioning error ≤ ±2mm, descent speed ≤ 0.2m / s.

[0076] Bandaging action : Control the end effector to complete the binding operation. The parameters are binding force and number of binding turns. Constraint: execution time ≥ 0.5s.

[0077] Reset action :from return The parameters are Coordinates, constraint: upward velocity ≤ 0.3 m / s.

[0078] Step 3.1.2 Task Decomposition Logic

[0079] Based on point sequence The task is decomposed into M local task units using a greedy strategy. , , Let M be the coordinates of the binding point, and each Complete workflow for a single location:

[0080]

[0081] in, This represents the movement action of the k-th point. Included parameters , Represents the safe coordinates of the binding point for the movement of the k-th point; This indicates the positioning action at the k-th point. Included parameters , This represents the coordinates of the binding point for the k-th positioning action; This indicates the binding action at the k-th point. Included binding strength Number of binding loops parameter; This indicates the reset action at the k-th point. Included parameters , This represents the safe coordinates of the binding point for the k-th point reset action. The global task is decomposed into a task sequence. .

[0082] Step 3.2: Construction of Reinforcement Learning Decision Model

[0083] Design a reinforcement learning model to optimize the execution order and action parameters of task sequence T, and minimize the total motion time and energy consumption.

[0084] Step 3.2.1 State Space Definition

[0085] Define a high-dimensional continuous state space, which includes:

[0086] Current status of the robotic arm: real-time coordinates of the end effector, joint angles, and end effector speed.

[0087] Task progress status: Number of completed binding points, and collection of incomplete binding points.

[0088] Environmental constraints: current energy consumption, remaining operating time.

[0089] Step 3.2.2 Motion Space Optimization

[0090] Based on the basic motion space A, the motion parameters are transformed into continuously optimizable variables:

[0091] Movement :optimization The Z-direction offset and the velocity of motion.

[0092] Positioning Action Optimize descent speed and positioning dwell time.

[0093] Bandaging action : Optimize the binding strength and the number of binding loops.

[0094] Step 3.2.3 Reward Function Design

[0095] A multi-objective weighted reward function is adopted to balance motion time, energy consumption, and operational accuracy. The formula is as follows:

[0096]

[0097] Among them, completion rewards 10.0 points are awarded for completing the binding of a single point; an additional 50.0 points are awarded for completing all points; time penalty applies. For a single task unit Actual execution time With optimal time The difference, The preset time is 1.5 seconds. Energy consumption penalty To calculate energy consumption based on the dynamics model of the robotic arm, , For the torque of the j-th joint, Let be the angular velocity of the j-th joint. Error penalty. Positioning error , These are the actual positioning coordinates. .

[0098] Step 3.2.4 Model Selection and Network Structure

[0099] The Deep Deterministic Policy Gradient Algorithm is selected. Based on the trained Deep Deterministic Policy Gradient Model, the current state vector is input and the optimal action sequence is output.

[0100] Step 4: Integrating Adaptive Path Planning

[0101] The optimal motion sequence output by reinforcement learning is used to generate the real-time motion trajectory of the robotic arm, ensuring that the path planning avoids obstacles and optimizes the binding sequence; the real-time motion controller is used to execute the local motion sequence to control the end effector of the robotic arm to complete the precise binding operation.

[0102] Step 4: Adaptive Path Planning Generation

[0103] Based on the action sequence in step 3, and combined with the kinematic constraints of the robotic arm and information about environmental obstacles, a smooth, efficient, and collision-free joint spatial trajectory is generated.

[0104] Step 4.1 Calculation of forward and inverse kinematics

[0105] Based on the DH parameter table, the position and attitude of the end effector in the world coordinate system are calculated using the homogeneous transformation matrix. The formula is as follows:

[0106]

[0107] The DH parameter table is obtained from the physical structure design data of the robotic arm and actual measurement calibration, and the homogeneous transformation matrix of each link is also included. Defined as:

[0108]

[0109] In the formula, Let be the joint angle to be solved. The length of the link. The link twist angle, This refers to joint displacement.

[0110] The positioning coordinates issued for each step 3 action sequence The joint angles are solved using a numerical iteration method. The iterative formula is:

[0111]

[0112] in The Jacobian matrix is ​​calculated in real time based on the DH parameters; + indicates the pseudo-inverse matrix. This represents the end position of the k-th iteration, and the iteration termination condition is... .

[0113] Step 4.2 Trajectory Interpolation and Smoothing

[0114] Cubic spline interpolation is used to generate the angle-time trajectory of each joint, ensuring that the joint motion velocity and acceleration are continuous without jumps. The interpolated trajectory is then subjected to a low-pass filter with a cutoff frequency of 5Hz to suppress high-frequency jitter.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention. A planning method for a rebar tying robotic arm based on visual semantic constraints.

Claims

1. A planning method for a rebar tying robotic arm based on visual semantic constraints, comprising the following specific steps, characterized in that: Step 1: Construct a visual perception system for the rebar tying robotic arm; To meet the needs of automated horizontal rebar tying operations, a high-precision six-degree-of-freedom robotic arm, an end effector for tying, and a vision sensor are integrated, with industrial cameras deployed above and to the side of the robotic arm's working area. Step 2: Develop a module for semantic understanding and key point detection of reinforcing bars; In the development of the rebar semantic understanding and key point detection module, RGB image and depth map information are integrated to achieve high-precision semantic segmentation of rebar mesh and key point detection of binding nodes, and automatically output the coordinates of binding points. The development of the rebar semantic understanding and key point detection module in step 2 can be represented as follows: Based on the RGB images and depth maps acquired by the visual perception system in step 1, a multimodal deep vision network is used to achieve semantic segmentation that distinguishes rebar from the background and high-precision detection of binding target points. Finally, the coordinates of the binding points in the world coordinate system that the robotic arm can recognize are output. Input data: Synchronized RGB images captured by the industrial-grade camera in step 1. With depth map ; Step 2.1: Data preprocessing and alignment; Step 2.1.1 Image normalization; right To normalize pixel values, use the following formula: ； in, The normalized RGB image, Image pixel coordinates, The global mean of the RGB image. The global standard deviation of the RGB image. For RGB images Pixel values; The following formula will be used to... Normalize to the [0,1] interval: ； in, This is the normalized interval depth map. , These are the minimum and maximum depth values ​​for a single depth map, respectively. For depth images in Pixel values; Step 2.1.2 Obtain camera calibration parameters; Due to differences in camera deployment locations, the depth map pixel coordinates are mapped to the RGB image coordinate system using an extrinsic parameter matrix for unification. The formula is as follows: ； in, These are the pixel coordinates of the depth map. This represents the depth value corresponding to that coordinate. For the pixel coordinates of the RGB image, the intrinsic parameters for camera calibration are obtained using the above formula. and external references , For rotation matrix, It is a translation vector; Step 2.2: Deployment and feature fusion of multimodal deep vision networks; It adopts a dual-branch feature extraction and cross-modal fusion architecture, loads a pre-trained network model, and fuses RGB and depth map features to provide high-dimensional feature support for semantic segmentation and key point detection; The feature extraction branch consists of an RGB branch and a depth map branch. The RGB branch uses ResNet50 as the backbone, and the depth map branch uses DenseNet121 as the backbone. The last fully connected layer is removed from both branches, retaining the convolutional feature output. The RGB image and depth image are taken as inputs respectively, and the output is an RGB feature map. and depth feature map Finally, a channel concatenation and attention-weighted fusion strategy is used to fuse the features, as shown in the formula: ； in, For channel-dimensional splicing, for convolution; Step 2.3: Semantic segmentation of reinforcing bars and detection of tying nodes; From fusion characteristics The steel reinforcement area is segmented out, and the coordinates of the binding node in the image coordinate system are located. Step 2.3.1 Semantic segmentation of reinforcing bars; The semantic segmentation head structure for rebar uses a U-Net decoder to fuse features. As input, the output is a segmentation probability map. The loss function is jointly optimized using Dice loss and binding entropy loss, and the formula is as follows: ； in, For binding entropy loss; The Dice loss is calculated using the following formula: ； in, The supervised annotation data is used to train the rebar semantic segmentation model. The annotation is the segmentation mask, with the rebar region set to 1 and the background set to 0. Avoid denominators of 0; Generate a binary mask by setting a threshold of 0.7; The specific formula for the binding entropy loss is as follows: ； Step 2.3.2 Key point detection of binding nodes; The lap joint is defined as the intersection of horizontal and vertical reinforcing bars, with the intersection point being more than 5mm from the edge of the reinforcing bar. The following method is used for inspection: Step 2.3.2.1 Determine the direction of the reinforcing bars using Hough line detection, based on the segmentation probability map. Extracting horizontal reinforcement masks and vertical reinforcement mask ; Step 2.3.2.2 Calculate the intersection mask Filter connected components with an area ≥ 25 pixels ; Step 2.3.2.3 Calculate the centroid coordinates for each connected component. , which is the coordinate of the binding node in the image coordinate system; Step 2.4: Coordinate transformation and output of binding points; Convert the binding node coordinates in the image coordinate system to world coordinate system coordinates that the robotic arm can recognize; Step 2.4.1 Coordinate transformation formula; Internal parameters based on camera calibration and external references Image coordinates Convert to world coordinates The formula is: ； in, This represents the depth value of the tethered node in the camera coordinate system. The inverse of the intrinsic parameter matrix. For rotation matrix, The translation vector is obtained from step 2.1.2; These are the homogeneous pixel coordinates of the image; Use the camera's homogeneous coordinates; The world coordinate system takes the center of the robotic arm base as the origin, the X-axis along the length of the horizontal reinforcing bar, the Y-axis along the length of the vertical reinforcing bar, and the Z-axis perpendicular to the ground and upward. Step 3: Design a dynamic task decomposition and reinforcement learning decision-making mechanism; Based on the binding point coordinates output in step 2, a dynamic task decomposition mechanism is introduced to automatically transform the global binding task into a local action sequence; the path planning and binding sequence of the robotic arm are optimized through reinforcement learning algorithm to achieve adaptive operation of steel mesh of different specifications. Step 4: Integrate adaptive path planning; The optimal motion sequence output by reinforcement learning is used to generate the real-time motion trajectory of the robotic arm, ensuring that the path planning avoids obstacles and optimizes the binding sequence; the real-time motion controller is used to execute the local motion sequence to control the end effector of the robotic arm to complete the precise binding operation.

2. The rebar tying robotic arm planning method based on visual semantic constraints according to claim 1, characterized in that: The dynamic task decomposition and reinforcement learning decision-making mechanism designed in step 3 can be represented as follows: Step 3.1: Implementation of dynamic task decomposition mechanism; The coordinates of the binding points output in step 2 Combine the coordinates to obtain the set S of the binding point locations; Will Defined as ,Will Defined as the safety coordinates of the binding point , To determine the offset of the safety coordinates on the z-axis, the binding task of traversing all points in set S is transformed into a sequence of local actions that the robotic arm can execute, and the goal and constraints of each action are clearly defined. Step 3.1.1 Define the action space; Define four basic types of movements to form a movement sequence. : Movement From current position Reach a safe location near the target point The parameters are Coordinates, constraint: motion speed ≤ 0.5m / s; Positioning Action :from Descend to the lashing point The parameters are Coordinates, constraints: positioning error ≤ ±2mm, descent speed ≤ 0.2m / s; Bandaging action : Control the end effector to complete the binding operation, with parameters including binding force and number of binding turns, and constraint: execution time ≥ 0.5s; Reset action :from return The parameters are Coordinates, constraint: upward velocity ≤ 0.3 m / s; Step 3.1.2 Task decomposition logic; Based on point sequence The task is decomposed into M local task units using a greedy strategy. , , Let M be the coordinates of the binding point, and each Complete workflow for a single location: ； in, This represents the movement action of the k-th point. Included parameters , Represents the safe coordinates of the binding point for the movement of the k-th point; This indicates the positioning action at the k-th point. Included parameters , This represents the coordinates of the binding point for the k-th positioning action; This indicates the binding action at the k-th point. Included binding strength Number of binding loops parameter; This indicates the reset action at the k-th point. Included parameters , This represents the safe coordinates of the binding point for the k-th point reset action; the global task is decomposed into a task sequence. ; Step 3.2: Construction of the reinforcement learning decision model; Design a reinforcement learning model to optimize the execution order and action parameters of task sequence T, and minimize the total motion time and energy consumption; Step 3.2.1 State space definition; Define a high-dimensional continuous state space, which includes: Current status of the robotic arm: real-time coordinates of the end effector, joint angles, and end effector speed; Task progress status: Number of completed binding points, collection of incomplete binding points; Environmental constraints: current energy consumption, remaining operating time; Step 3.2.2 Motion space optimization; Based on the basic motion space A, the motion parameters are transformed into continuously optimizable variables: Movement :optimization Z-direction offset and velocity; Positioning Action Optimize descent speed and positioning dwell time; Bandaging action : Optimize the binding strength and number of binding loops; Step 3.2.3 Reward function design; A multi-objective weighted reward function is adopted to balance motion time, energy consumption, and operational accuracy. The formula is as follows: ； Among them, completion rewards 10.0 points are awarded for completing the binding of a single point; an additional 50.0 points are awarded for completing all points; time penalty applies. For a single task unit Actual execution time With optimal time The difference, The preset time is 1.5 seconds. Energy consumption penalty To calculate energy consumption based on the dynamics model of the robotic arm, , For the torque of the j-th joint, Let ω be the angular velocity of the j-th joint; error penalty Positioning error , These are the actual positioning coordinates. ; Step 3.2.4 Model selection and network structure; The Deep Deterministic Policy Gradient Algorithm is selected. Based on the trained Deep Deterministic Policy Gradient Model, the current state vector is input and the optimal action sequence is output.

3. The rebar tying robotic arm planning method based on visual semantic constraints according to claim 1, characterized in that: Step 4, integrating adaptive path planning, can be represented as follows: Step 4: Adaptive path planning generation; Based on the action sequence in step 3, and combined with the kinematic constraints of the robotic arm and environmental obstacle information, a smooth, efficient, and collision-free joint space trajectory is generated. Step 4.1 Calculation of forward and inverse kinematics; Based on the DH parameter table, the position and attitude of the end effector in the world coordinate system are calculated using the homogeneous transformation matrix. The formula is as follows: ； The DH parameter table is obtained from the physical structure design data of the robotic arm and actual measurement calibration, and the homogeneous transformation matrix of each link is also included. Defined as: ； In the formula, Let be the joint angle to be solved. The length of the link. The link twist angle, Joint displacement; The positioning coordinates issued for each step 3 action sequence The joint angles are solved using a numerical iteration method. The iterative formula is: ； in The Jacobian matrix is ​​calculated in real time based on the DH parameters; + indicates the pseudo-inverse matrix. This represents the end position of the k-th iteration, and the iteration termination condition is... ; Step 4.2 Trajectory interpolation and smoothing; Cubic spline interpolation is used to generate the angle-time trajectory of each joint, ensuring that the joint motion velocity and acceleration are continuous without jumps. The interpolated trajectory is then subjected to a low-pass filter with a cutoff frequency of 5Hz to suppress high-frequency jitter.

Images