An autonomous motion planning method and system for robotic arms performing plug-in / plug-out operations in confined spaces of power distribution cabinets
By using a motion strategy network that integrates observation features from multi-view images and robot body state, the collision and motion accuracy issues of the robotic arm in the plug-in and unplugging operations of the power distribution cabinet are solved, achieving high precision and stability in autonomous motion planning.
Patent Information
- Application Number
- CN202611585471.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
In existing technologies, during the plug-in and plug-out operations of power distribution cabinets, the robotic arm is prone to collisions with the cabinet or electrical components during its movement. The plug-in and plug-out actions lack precision and cannot respond to obstructions and status changes in a timely manner, resulting in plug jamming or damage.
By constructing fused observation features using multi-view images and robot body state, future action blocks are predicted through an action policy network, and the action policy network is trained using latent variables and exponential decay loss to generate weighted fused robotic arm execution actions, thereby achieving autonomous motion planning.
It improves the local visual representation capability within the power distribution cabinet, the short-term motion accuracy during the key stages of insertion and removal, and the closed-loop correction capability, enhances the prediction accuracy and stability of the motion strategy network, and improves the insertion and removal operation effect of the robotic arm.
Smart Images

Figure CN123060733A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robotic arm motion planning technology, and in particular to a method and system for autonomous motion planning of robotic arms for plug-in / plug-out operations in confined spaces of power distribution cabinets. Background Technology
[0002] Distribution cabinets typically house circuit breakers, relays, terminal blocks, wiring harnesses, and various pluggable functional components. During component inspection, replacement, and testing, a robotic arm must sequentially complete the following steps: target component identification, approach to confined spaces, gripper alignment, component insertion or removal, and safe withdrawal. Due to the narrow space, dense components, and significant obstructions within the cabinet, the robotic arm is prone to collisions with the cabinet, adjacent electrical components, or wiring harnesses during its movement. Furthermore, component insertion and removal have specific end-position, posture, and insertion / removal directions; even minor deviations can cause component jamming, insertion failure, or component damage.
[0003] Currently, the main control methods for robotic arms used in power distribution cabinet plug-in / plug-out operations include manual teaching trajectory reproduction, trajectory planning based on precise geometric models, and control methods based on visual servoing or imitation learning. Manual teaching trajectory reproduction methods are less adaptable to changes in plug-in position and cabinet layout; trajectory planning methods based on geometric models rely on precise models of the cabinet, plug-in, slots, and obstacles, resulting in high modeling costs under conditions of occlusion, installation errors, and wiring harness disturbances; and conventional visual servoing methods require designing control rules for each plug-in individually.
[0004] Basic action block imitation learning methods can generate future continuous actions based on current observations, but when directly applied to the plug-in and plug-out tasks of power distribution cabinets, there are still three problems: First, ordinary convolutional visual features are insufficient in expressing the local spatial relationships between plug-in edges, slot contours, end effectors, and nearby obstacles; Second, equal-weighted training of all time steps within an action block cannot highlight the alignment, contact, and plug-in / plug-out actions that are about to be executed; Third, during online execution, predictions of old action blocks continue to participate in fusion, which may cause response lag when the plug-in position, end effector state, or occlusion relationship changes. Summary of the Invention
[0005] This invention provides a method and system for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of power distribution cabinets, in order to solve the technical problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an autonomous motion planning method for a robotic arm performing plug-in / plug-out operations in confined spaces of a power distribution cabinet, comprising the following steps: S1. Acquire multi-view images of the power distribution cabinet plug-in / plug-out operation environment and the robot's body state. Construct a robot body state vector based on the robot's body state. Acquire expert teaching actions for plug-in insertion and plug-in removal; S2, for multi-view images and robot body state vectors The expert demonstration actions are preprocessed, and then expert action blocks are constructed. ; S3. Based on the preprocessed multi-view images, generate visual features corresponding to the camera viewpoints. Then, input the visual features from multiple camera viewpoints into the action policy network to obtain fused observation features. ; S4, Action Policy Network based on latent variables and by fusing observation features Predict future action blocks based on conditions Then based on future action blocks Calculate the overall training loss Based on overall training loss The action policy network is dynamically adjusted until the training of the action policy network is completed; S5. Use the trained motion policy network to perform autonomous motion planning for the robotic arm, generating multiple candidate actions; perform weighted fusion of all candidate actions to obtain the current action to be executed. ; S6. Change the currently executing action The command is sent to the robotic arm controller and the robotic arm is controlled to execute the command. After execution, the command returns to S5 and enters the next control cycle until the robotic arm's autonomous motion planning task is completed.
[0007] Furthermore, step S1 specifically includes the following steps: S11. Install several environmental cameras at designated locations on the cabinet to obtain images of the overall layout of the power distribution cabinet and the positional relationship of the robotic arm relative to the cabinet. S12. Install several wrist cameras at the end of the robotic arm to acquire local images of the plug handle, slot edge, gripper and adjacent devices; S13. Synchronously acquire the robot's body state, including the joint position, joint velocity, end effector pose, and gripper state of the robotic arm, and construct the robot's body state vector based on the robot's body state. Then, using the images obtained in S11 and S12, as well as the robot's body state vector... Build t Complete set of observations at time ; ; in, express t Time of the first M Images from a single camera perspective; MThis represents the sum of environmental cameras and wrist cameras; S14. Obtain expert teaching actions for plug-in insertion and plug-in removal.
[0008] Furthermore, step S2 specifically includes the following steps: S21. Organize multi-view images and robot body state vectors according to a unified timestamp. Align with the actions demonstrated by experts; S22. Solve for the aligned robot body state vectors respectively. The mean and standard deviation of the expert-taught actions are used to align the robot's body state vector. The mean and standard deviation of the aligned robot body state vector Standardization is performed; the expert demonstration actions are standardized using the mean and standard deviation; the formula used for standardization is: ; in, Represents the standardized robot body state vector Or the first step in the expert demonstration. One variable; This indicates a small positive number that prevents the denominator from being zero. This represents the mean; Indicates standard deviation; At this point, the multi-view images and robot body state vectors have been completed. And the preprocessing of expert demonstration movements; S23, at the current time t Starting from the standardized expert demonstration movements, a length of [length missing] is extracted. T Future continuous actions constitute expert action blocks .
[0009] Furthermore, step S3 specifically includes the following steps: S31, the preprocessed first m Images from a camera perspective The image patch feature sequence is resampled to a resolution divisible by the image patch size of the visual encoder DINOv2 and input into the visual encoder DINOv2. ; S32, Image patch feature sequence Reconstruct a two-dimensional feature grid based on the original positions of the image patches. ; S33. Using 1×1 convolution to transform the two-dimensional feature grid The feature dimension C is mapped to the hidden layer dimension d of the policy network. , Obtain the visual features of the m-th camera viewpoint ; S34, repeat S31 to S33 until you get M Visual features from a single camera perspective; S35, will M Visual features from each camera viewpoint are concatenated in viewpoint order and then compared with the preprocessed robot body state vector. coding features A common input action policy network is used to form fused observation features. .
[0010] Furthermore, the two-dimensional feature mesh The expression is: ; in, Indicates the first m Image block feature sequences output from a camera viewpoint image by the visual encoder DINOv2; Represents a sequence of image patch features The two-dimensional feature mesh is restored according to the spatial arrangement of image patches in the original image; Reshape indicates the restoration of the two-dimensional mesh according to the original spatial order of the image patches. Represents the set of real numbers; B Indicates batch size; C The feature dimension representing the features of each image patch; N Indicates the number of image patches, and ; This represents the height of the two-dimensional feature grid, i.e., the number of rows in the two-dimensional feature grid; This represents the width of the two-dimensional feature grid, i.e., the number of columns in the image patch grid; m Indicates the camera viewpoint number. m=1,2,…,M ; The visual features The expression is: ; in, Represents a 1×1 convolution; The fusion observation features The expression is: ; Among them, Transformer represents the feature fusion module based on the Transformer model in the action policy network, which is used to encode and fuse the stitched multi-camera view visual features and robot body state encoding features. The encoder represents the robot's body state. Indicates the concatenation of feature sequences; Indicates the firstM Visual features from a single camera perspective.
[0011] Furthermore, step S4 specifically includes the following steps: S41, Latent variable encoder based on expert action block and their corresponding fusion observation features Estimate the posterior distribution of latent variables In the prior distribution Under the constraints, using the posterior distribution Obtaining latent variables ; latent variables Combined observation features The common input action block decoder generates future action blocks; S42, Based on the posterior distribution Prior distribution Calculate divergence loss ; S43, Targeting Future Actions For each future action within the time frame, calculate the normalized time distance separately; S44. Calculate the unnormalized exponential decay weight of each future action based on the normalized time distance of each future action, and further obtain the normalized weight. S45. Calculate the exponentially decaying action reconstruction loss based on the normalized weights of each future action. ; S46. Divergence loss Reconstruction loss with exponential decay action Combined, the overall training loss is obtained. Determine the overall training loss If the loss is within the set range, output the trained action policy network and pass it to S5; otherwise, use the overall training loss. Dynamically adjust the action policy network; S47. Return to S3 and continue training the action policy network.
[0012] Furthermore, the future action block The expression is: ; in, Indicates that the trainable parameters are Action block decoder in action policy network, This represents the trainable parameters of the action block decoder. The action policy network is based on time. t The first observation predicted T -1 future action; The divergence loss The expression is: ; in, Represents the relative entropy divergence operation, used to measure the posterior distribution of latent variables. With prior distribution The differences between them; Represents the trainable parameters of the latent variable encoder; S43 i The expression for the normalized time distance of a future action is: ; in, Indicates the first i Normalized time distance of a future action; S44 i The expression for the unnormalized exponentially decaying weight of each future action is: ; in, Indicates the first i The unnormalized exponentially decaying weight of each future action, Represents the natural exponential function; This refers to the weight decay coefficient during action block training; S44 i The expression for the normalized weights of a future action is: ; in, Indicates the first i Normalized weights for future actions, Indicates the summation index, = 0, 1, 2, ..., T -1; Indicates the first Unnormalized exponentially decaying weights of future actions; normalized weights The sum of the normalized weights of all future actions is 1 under different values of the action block training weight decay coefficient λ, as expressed by: ; Exponential decay action reconstructs loss The expression is: ; in, The action policy network is based on time. t The first observation predicted i A future action; Indicates expert action block The Middle t+i A series of future actions; Represents predicted future actions and expert action blocks. The middle corresponds to the absolute error of future continuous actions in each control dimension.
[0013] Furthermore, step S5 specifically includes the following steps: S51. Use the trained motion policy network to perform autonomous motion planning for the robotic arm and generate multiple candidate actions. S52. Then, for each candidate action, construct unnormalized temporal proximity weights, where the first... The expression for the unnormalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Unnormalized temporal proximity weights for each candidate action Indicates the time proximity integration coefficient; Indicates the first One candidate action; j =0,1,…, K -1; K Indicates the total number of candidate actions; S53. Based on the unnormalized temporal proximity weights of each candidate action, construct the corresponding normalized temporal proximity weights, where the i-th... The expression for the normalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Normalized time proximity weights for each candidate action; n Indicates the summation index, n = 0,1,2,…, K -1; Indicates the first n The unnormalized exponentially decaying weights of each candidate action; the sum of the normalized time proximity weights of all candidate actions is 1, expressed as: ; S54. Perform weighted fusion on all candidate actions to obtain the current action to be executed and sent to the robotic arm controller. The expression is as follows: ; in, Indicates the first One candidate action.
[0014] Furthermore, step S6 specifically includes the following steps: S61. Change the currently executing action The data is sent to the robotic arm controller, which controls the robotic arm to perform corresponding joint movements, end-effector pose adjustments, and gripper control. S62. After execution, proceed to the next control cycle and re-acquire the current complete observation set. Then, it returns to S5 to form a closed-loop motion planning process of continuous perception, action block prediction, candidate action fusion and action execution, until the autonomous motion planning task of the robotic arm is completed.
[0015] In another aspect, the present invention provides an autonomous motion planning system for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, which is configured to execute an autonomous motion planning method for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet.
[0016] The beneficial effects of this invention are: This invention discloses an autonomous motion planning method for robotic arms in confined space plug-in / plug-out operations in power distribution cabinets. While retaining the ability to generate candidate actions, it improves the local visual representation ability within the power distribution cabinet, the short-term motion accuracy during key plug-in / plug-out stages, the ability to perform closed-loop correction based on the latest observations, and the closed-loop stability. It also further improves the prediction accuracy of the motion strategy network. Attached Figure Description
[0017] Figure 1 This is a flowchart of the autonomous motion planning method for the robotic arm in this invention. Detailed Implementation
[0018] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many other different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.
[0019] This invention aims to address the problems of inaccurate spatial relationships between plug-in components, slots, and adjacent obstacles due to densely packed and obstructed components inside power distribution cabinets; insufficient short-term motion accuracy during critical insertion and removal phases; and the inability of the robotic arm to promptly correct its course based on the latest cabinet status due to the influence of historical motion prediction. To solve these problems, this invention provides an autonomous motion planning method for robotic arms performing plug-in / removal operations in confined spaces of power distribution cabinets, as detailed below: Reference Figure 1 This application provides an autonomous motion planning method for a robotic arm performing plug-in / plug-out operations in confined spaces of a power distribution cabinet, comprising the following steps: S1. Set up the plug-in modules to be operated, corresponding slots, cabinet boundaries, adjacent electrical components and wiring harnesses in the power distribution cabinet, and configure the robotic arm and end effector. Collect multi-view images of the plug-in / plug-out operation environment in the power distribution cabinet and the robot's body state, and construct the robot's body state vector based on the robot's body state. Acquire expert teaching actions for plug-in insertion and plug-in removal; S2, for multi-view images and robot body state vectors The expert demonstration actions are preprocessed, and then expert action blocks are constructed. ; S3. Based on the preprocessed multi-view images, generate visual features corresponding to the camera viewpoints. Then, input the visual features from multiple camera viewpoints into the action policy network to obtain fused observation features. ; S4, Action Policy Network based on latent variables and by fusing observation features Predict future action blocks based on conditions Then based on future action blocks Calculate the overall training loss Based on overall training loss The action policy network is dynamically adjusted until the training of the action policy network is completed; S5. Use the trained motion policy network to perform autonomous motion planning for the robotic arm, generating multiple candidate actions; perform weighted fusion of all candidate actions to obtain the current action to be executed. ; S6. Change the currently executing action The command is sent to the robotic arm controller and the robotic arm is controlled to execute the command. After execution, the command returns to S5 and enters the next control cycle until the robotic arm's autonomous motion planning task is completed.
[0020] While retaining the ability to generate candidate actions, this invention improves the local visual representation capability within the distribution cabinet, the short-term action accuracy during the key stages of insertion and removal, the ability to perform closed-loop correction based on the latest observations, and the closed-loop stability. It also further improves the prediction accuracy of the action strategy network in plug alignment, initial contact, and movement along the insertion and removal direction.
[0021] In some embodiments, S1 specifically includes the following steps: S11. Install several environmental cameras at designated locations on the cabinet to obtain images of the overall layout of the power distribution cabinet and the positional relationship of the robotic arm relative to the cabinet. S12. Install several wrist cameras at the end of the robotic arm to acquire local images of the plug handle, slot edge, gripper and adjacent devices; S13. Synchronously acquire the robot's body state, including the joint position, joint velocity, end effector pose, and gripper state of the robotic arm, and construct the robot's body state vector based on the robot's body state. Then, using the images obtained in S11 and S12, as well as the robot's body state vector... Build t Complete set of observations at time ; ; in, express t Time of the first M Images from a single camera perspective; M This represents the sum of environmental cameras and wrist cameras; S14. Acquire expert teaching actions for plug-in insertion and plug-in removal. The expert teaching action trajectory is a collision-free trajectory or a trajectory that meets the preset safety clearance, enabling the motion strategy network to learn confined space motion and insertion / removal direction constraints from specific power distribution cabinet operation examples. Expert teaching actions include plug-in insertion teaching actions and plug-in removal teaching actions; The plug-in insertion teaching action includes plug-in recognition and clamping, safe posture adjustment outside the cabinet, entering the confined space, aligning the plug-in with the slot, advancing along the insertion direction, and end-effector withdrawal. The plug-in removal teaching action includes target plug-in identification, approach to confined space, gripper alignment and clamping, withdrawal in the removal direction, avoidance of adjacent devices, and plug-in transfer.
[0022] In some embodiments, S2 specifically includes the following steps: S21. Organize multi-view images and robot body state vectors according to a unified timestamp. Align with the actions demonstrated by experts; S22. Solve for the aligned robot body state vectors respectively. The mean and standard deviation of the expert-taught actions are used to align the robot's body state vector. The mean and standard deviation of the aligned robot body state vector Standardization is performed; the expert demonstration actions are standardized using the mean and standard deviation; the formula used for standardization is: ; in, Represents the standardized robot body state vector Or the first step in the expert demonstration. One variable; This indicates a small positive number that prevents the denominator from being zero. This invention utilizes the synchronized robot body state vector Standardizing the movements demonstrated by experts can ensure that physical quantities with different dimensions, such as joint position, joint velocity, end-effector posture, and gripper state, have similar numerical ranges, thus preventing a particular type of physical quantity from dominating the training of the movement strategy network due to its large numerical scale.
[0023] At this point, the multi-view images and robot body state vectors have been completed. And the preprocessing of expert demonstration movements; S23, at the current time t Starting from the standardized expert demonstration movements, a length of [length missing] is extracted. T Future continuous actions constitute expert action blocks The expression is as follows: ; in, express t+T Actions at time -1; Expert action block in this invention The front end of the action block directly corresponds to the upcoming plug-in approach, alignment, and contact actions, while the back end of the action block describes the subsequent plug-in, plug-out, and withdrawal trends. Action block segmentation enables the action policy network to predict a continuous motion at a time, reducing motion jitter and cumulative errors caused by independent prediction at each time step.
[0024] In some embodiments, S3 specifically includes the following steps: S31, the preprocessed first m Images from a camera perspective The image patch feature sequence is resampled to a resolution divisible by the image patch size of the visual encoder DINOv2 and input into the visual encoder DINOv2. ; S32. To preserve the two-dimensional spatial arrangement of plug-ins, slots, grippers, and adjacent devices in the image, the image block feature sequence is... Reconstruct a two-dimensional feature grid based on the original positions of the image patches. ; S33. Using 1×1 convolution to transform the two-dimensional feature grid Feature Dimensions C Mapped to the hidden layer dimension of the policy network d, Get the first m Visual features from each camera's perspective ; S34, repeat S31 to S33 until you get M Visual features from a single camera perspective; S35, will M Visual features from each camera viewpoint are concatenated in viewpoint order and then compared with the preprocessed robot body state vector. coding features A common input action policy network is used to form fused observation features. .
[0025] Fusion observation features It simultaneously includes global layout information within the power distribution cabinet, local geometric information of the robotic arm's end effector neighborhood, and the robotic arm's own motion state information. The visual encoder DINOv2 is built based on the Transformer model. It first divides the input image into multiple image blocks and converts each block into corresponding image block features. Then, through the self-attention mechanism within the visual encoder DINOv2, it calculates the correlation between different image block features and interacts with contextual information between different image blocks based on this correlation. This ensures that the output image block features not only contain information about the corresponding local image region but also contextual information related to other image regions. Therefore, the image block feature sequence output by the visual encoder DINOv2... It can comprehensively characterize the relationships between plug edges, slot contours, grippers, adjacent electrical components, and wire harnesses.
[0026] This invention extracts image patch feature sequences from multi-view images using the visual encoder DINOv2. and image patch feature sequences Restored to a two-dimensional feature mesh This enables the motion strategy network to simultaneously express the spatial relationships between plug edges, slot contours, grippers, and adjacent devices, thereby improving target perception and alignment capabilities in confined spaces, and enhancing target localization, confined space access, and end-point alignment capabilities in densely obscured distribution cabinets.
[0027] In some embodiments, the two-dimensional feature mesh The expression is: ; in, Indicates the first m Image block feature sequences output from a camera viewpoint image by the visual encoder DINOv2; Represents a sequence of image patch features The two-dimensional feature mesh is restored according to the spatial arrangement of image patches in the original image; Reshape indicates the restoration of the two-dimensional mesh according to the original spatial order of the image patches. Represents the set of real numbers; B Indicates batch size; C The feature dimension representing the features of each image patch; N Indicates the number of image patches, and ; This represents the height of the two-dimensional feature grid, i.e., the number of rows in the two-dimensional feature grid; This represents the width of the two-dimensional feature grid, i.e., the number of columns in the image patch grid; m Indicates the camera viewpoint number. m= 1 , 2 ,…,M ; The visual features The expression is: ; in, Represents a 1×1 convolution; The fusion observation features The expression is: ; Among them, Transformer represents the feature fusion module based on the Transformer model in the action policy network, which is used to encode and fuse the stitched multi-camera view visual features and robot body state encoding features. The encoder represents the robot's body state. Indicates the concatenation of feature sequences; Indicates the first M Visual features from a single camera perspective.
[0028] In some embodiments, S4 specifically includes the following steps: S41, Latent variable encoder based on expert action block and their corresponding fusion observation features Estimate the posterior distribution of latent variables In the prior distribution Under the constraints, using the posterior distribution Obtaining latent variables ; latent variables Combined observation features The common input action block decoder generates future action blocks; S42, Based on the posterior distribution Prior distribution Calculate divergence loss Divergence loss This is used to constrain the distribution of latent variables, preventing the action policy network from merely memorizing expert-taught actions, and enabling the action policy network to learn from prior distributions during the inference phase. Obtaining stable latent variables And generate future action blocks; S43, Targeting Future Actions For each future action within the time frame, calculate the normalized time distance separately; S44. Calculate the unnormalized exponential decay weight of each future action based on the normalized time distance of each future action, and further obtain the normalized weight. S45. Calculate the exponentially decaying action reconstruction loss based on the normalized weights of each future action. ; Exponential decay action reconstructs loss To improve the prediction accuracy of specific insertion and extraction actions, divergence loss Both are used to ensure the stability of the latent variable distribution, and together they enable the action policy network to maintain the ability to generate continuous actions and highlight key actions in the near future. S46. Divergence loss Reconstruction loss with exponential decay action Combined, the overall training loss is obtained. Determine the overall training loss If the loss is within the set range, the trained action policy network is obtained and entered into S5; otherwise, the overall training loss is used. Dynamically adjust the action policy network; S47. Return to S3 and continue training the action policy network.
[0029] Existing basic Action Chunking with Transformers (ACT) strategies apply equally weighted action reconstruction losses to each time step within an action block. However, in the closed-loop control of plug-in / plug-out in a distribution cabinet, actions closer to the current time are applied to the robotic arm earlier, directly impacting end-effector alignment, initial contact, and movement along the insertion / plugging direction. Longer-term actions require re-prediction based on new observations before actual execution. To address this issue, this invention employs exponentially decaying action reconstruction losses. Make future action blocks In this model, near-future actions are given larger normalization weights, while far-future actions are given smaller, non-zero normalization weights. Therefore, the action policy network can focus on learning near-future actions such as alignment, initial contact, and movement along the insertion / removal direction, while retaining the learning of long-term action trends. This improves the accuracy of action prediction and the stability of continuous action generation during the critical insertion / removal stages.
[0030] In some embodiments, the future action block The expression is: ; in, Indicates that the trainable parameters are Action block decoder in action policy network, This represents the trainable parameters of the action block decoder. The action policy network is based on time. t The first observation predicted T -1 future action; The divergence loss The expression is: ; in, Represents the relative entropy divergence operation, used to measure the posterior distribution of latent variables. With prior distribution The differences between them; Represents the trainable parameters of the latent variable encoder; The expression for the normalized time distance of the i-th future action in S43 is: ; in, Indicates the first i Normalized time distance of a future action; when i When = 0, normalized time distance =0, which means that the action is executed first; when i=T At -1, normalized time distance =1 indicates that the action is the furthest action in the action block; S44 i The expression for the unnormalized exponentially decaying weight of each future action is: ; in, Indicates the first i The unnormalized exponentially decaying weight of each future action, Represents the natural exponential function; This refers to the weight decay coefficient during action block training; S44 i The expression for the normalized weights of a future action is: ; in, Indicates the first i Normalized weights for future actions, Indicates the summation index, = 0, 1, 2, ..., T -1; Indicates the first Unnormalized exponentially decaying weights of future actions. Normalized time distance. The unnormalized exponential decay weight increases with the increase of the prediction time distance. and normalized weights It decreases as the predicted time interval increases; Normalized weights In action block training weight decay coefficient λ The sum of the normalized weights of all future actions under different values is 1, expressed as: ; Exponential decay action reconstructs loss The expression is: ; in, The action policy network is based on time. t The first observation predicted i A future action; Indicates expert action block The Middle t +i future consecutive actions; Represents predicted future actions and expert action blocks. The middle corresponds to the absolute error of future continuous actions in each control dimension.
[0031] because > >…> ,in Indicates the first T The normalized weights for future actions are set to -1, so the action policy network prioritizes reducing the prediction error of actions to be executed, thereby improving the short-term accuracy of plug-in approach, alignment, initial contact, and movement along the insertion / removal axis. The weights for long-term actions are not zero, allowing the action policy network to still learn the complete insertion / removal trajectory trend. Preferably, the action block training weight decay coefficient λ is 5.0; when the action block training weight decay coefficient λ is 0, the weights of each action are the same, and the action reconstruction loss decays exponentially. Degradation is used to reconstruct losses based on equal-weighted actions.
[0032] This invention improves the prediction accuracy of plug alignment, initial contact, and movement along the insertion / removal direction by making the normalized weight of near-future actions higher than that of far-future actions.
[0033] In some embodiments, S5 specifically includes the following steps: S51. The trained motion policy network is used for autonomous motion planning of the robotic arm. During continuous control, the trained motion policy network generates future action blocks in each control cycle and stores the most recently generated future action block in the action block queue. For the current moment, the corresponding action prediction is extracted from each future action block in the action block queue to form a motion prediction for the current moment. K There are 10 candidate actions, numbered from oldest to newest according to the generation time of the corresponding action block. j =0,1,…, K 1; S52. Then, for each candidate action, construct unnormalized temporal proximity weights, where the first... The expression for the unnormalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Unnormalized temporal proximity weights for each candidate action Represents the time proximity integration coefficient, preferably, the time proximity integration coefficient Take 0.01; Indicates the first One candidate action; j =0,1,…, K -1; K Indicates the total number of candidate actions; S53. Based on the unnormalized temporal proximity weights of each candidate action, construct the corresponding normalized temporal proximity weights, where the i-th... The expression for the normalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Normalized time proximity weights for each candidate action; n The index represents the summation subscript, n = 0, 1, 2, ..., K-1; Indicates the first n The unnormalized exponential decay weight of each candidate action. When the state inside the distribution cabinet is stable, the difference between each future action and the candidate action in the future action block is small, and this formula can maintain smooth action; when the plug position, end state or occlusion relationship changes, the action generated based on the latest observation has a higher weight, and this formula can correct the direction of the robot arm movement more quickly.
[0034] The sum of the normalized time proximity weights of all candidate actions is 1, expressed as: ; S54. Perform weighted fusion on all candidate actions to obtain the current action to be executed and sent to the robotic arm controller. The expression is as follows: ; in, Indicates the first One candidate action.
[0035] This invention is based on time proximity integration coefficients By constructing temporal proximity weights for candidate actions, higher fusion weights are obtained for candidate actions generated based on more recent observations within the distribution cabinet. This improves response capability under state changes and closed-loop correction capability while maintaining action smoothness. This invention does not change the basic paradigm of action block division in existing basic action block converter strategies, can reuse existing teaching data processing, training, and deployment processes, and can be extended to plug-in / plug-out operation scenarios for different types of distribution cabinets.
[0036] In some embodiments, S6 specifically includes the following steps: S61. Change the currently executing action The data is sent to the robotic arm controller, which controls the robotic arm to perform corresponding joint movements, end-effector pose adjustments, and gripper control. During insertion, the robotic arm needs to sequentially complete the following steps: target insertion identification and gripping, safe positioning adjustment outside the cabinet, entry into the confined space, insertion and slot alignment, stable advancement along the insertion axis, and end effector withdrawal. During removal, the robotic arm needs to sequentially complete the following steps: target insertion identification, approach to the confined space, gripper alignment and gripping, stable withdrawal along the removal axis, avoidance of adjacent devices, and insertion transfer. The completion status of insertion / removal can be determined based on changes in wrist image, gripper status, joint displacement, end effector pose changes, or feedback from the power distribution cabinet equipment. S62. After execution, proceed to the next control cycle and re-acquire the current complete observation set. Then, it returns to S5 to form a closed-loop motion planning process of continuous perception, action block prediction, candidate action fusion and action execution, until the autonomous motion planning task of the robotic arm is completed.
[0037] To verify the closed-loop execution effect of the technical solution adopted in this invention, a comparative test was conducted on the existing basic motion block converter strategy and the motion strategy using the autonomous motion planning method of the robotic arm described in this invention, under the same simulation platform, the same robot body, the same task interface, and the same closed-loop scoring mechanism. The test was conducted on the KDC embodied intelligent simulation platform and included three types of tasks: toy sorting and grasping, package handling and weighing, and conveyor belt parts sorting. The average closed-loop evaluation score of 100 independent tests was used as the evaluation result for each type of task.
[0038] Test results show that the basic motion block converter strategy achieved a total score of 235.43 points across the three task categories, while the present invention achieved a total score of 259.63 points, representing an improvement of 24.20 points. Specifically, the score for the package handling and weighing task increased from 67.93 points to 80.03 points, indicating that the present invention improves motion continuity and short-term control performance during multi-stage operations involving dynamic grasping, intermediate placement, secondary grasping, and target area transfer. This result is consistent with the effect of exponentially decaying motion block loss highlighting the training weight of near-future motions. The score for the conveyor belt parts sorting task increased from 83.00 points to 93.20 points, indicating that the present invention improves responsiveness and continuous operation capabilities for continuously moving targets. This result is consistent with the effect of the temporal proximity integration strategy improving the prediction fusion weight of newer motions.
[0039] The above comparison results demonstrate that the method mentioned in this invention can improve the overall closed-loop execution capability of the robotic arm in multi-target recognition, multi-stage contact operation, and dynamic target response tasks.
[0040] To support a broader scope of protection, this invention is not limited to a particular plug-in shape, cabinet type, or robotic arm structure. Without altering the technical concepts of multi-view visual enhancement, near-future motion weighted training, and temporally proximate motion integration coefficients, it can be applied to the operation scenarios of industrial equipment with functional plug-ins of different sizes, drawer-type modules, plug-in control units, or other constraints such as limited space, alignment constraints, and plug-in / plug-out direction constraints.
[0041] The second aspect of the present invention also provides an autonomous motion planning system for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, configured to execute an autonomous motion planning method for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet.
[0042] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, characterized in that, Includes the following steps: S1. Acquire multi-view images of the power distribution cabinet plug-in / plug-out operation environment and the robot's body state. Construct a robot body state vector based on the robot's body state. Acquire expert teaching actions for plug-in insertion and plug-in removal; S2, for multi-view images and robot body state vectors The expert demonstration actions are preprocessed, and then expert action blocks are constructed. ; S3. Based on the preprocessed multi-view images, generate visual features corresponding to the camera viewpoints. Then, input the visual features from multiple camera viewpoints into the action policy network to obtain fused observation features. ; S4, Action Policy Network based on latent variables and by fusing observation features Predict future action blocks based on conditions Then based on future action blocks Calculate the overall training loss Based on overall training loss The action policy network is dynamically adjusted until the training of the action policy network is completed; S5. Use the trained motion policy network to perform autonomous motion planning for the robotic arm, generating multiple candidate actions; perform weighted fusion of all candidate actions to obtain the current action to be executed. ; S6. Change the currently executing action The command is sent to the robotic arm controller and the robotic arm is controlled to execute the command. After execution, the command returns to S5 and enters the next control cycle until the robotic arm's autonomous motion planning task is completed.
2. The method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in claim 1, is characterized in that... S1 specifically includes the following steps: S11. Install several environmental cameras at designated locations on the cabinet to obtain images of the overall layout of the power distribution cabinet and the positional relationship of the robotic arm relative to the cabinet. S12. Install several wrist cameras at the end of the robotic arm to acquire local images of the plug handle, slot edge, gripper and adjacent devices; S13. Synchronously acquire the robot's body state, including the joint position, joint velocity, end effector pose, and gripper state of the robotic arm, and construct the robot's body state vector based on the robot's body state. Then, using the images obtained in S11 and S12, as well as the robot's body state vector... Build t Complete set of observations at time ; ; in, express t Time of the first M Images from a single camera perspective; M This represents the sum of environmental cameras and wrist cameras; S14. Obtain expert teaching actions for plug-in insertion and plug-in removal.
3. The method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in claim 2, is characterized in that... S2 specifically includes the following steps: S21. Organize multi-view images and robot body state vectors according to a unified timestamp. Align with the actions demonstrated by experts; S22. Solve for the aligned robot body state vectors respectively. The mean and standard deviation of the expert-taught actions are used to align the robot's body state vector. The mean and standard deviation of the aligned robot body state vector Standardization is performed; the expert demonstration actions are standardized using the mean and standard deviation; the formula used for standardization is: ; in, Represents the standardized robot body state vector Or the first step in the expert demonstration. One variable; This indicates a small positive number that prevents the denominator from being zero. This represents the mean; Indicates standard deviation; At this point, the multi-view images and robot body state vectors have been completed. And the preprocessing of expert demonstration movements; S23, at the current time t Starting from the standardized expert demonstration movements, a length of [length missing] is extracted. T Future continuous actions constitute expert action blocks .
4. The method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in claim 3, is characterized in that... S3 specifically includes the following steps: S31, the preprocessed first m Images from a camera perspective The image patch feature sequence is resampled to a resolution divisible by the patch size of the visual encoder DINOv2 and input into the visual encoder DINOv2. ; S32, Image patch feature sequence Reconstruct a two-dimensional feature grid based on the original positions of the image patches. ; S33. Using 1×1 convolution to transform the two-dimensional feature grid The feature dimension C is mapped to the hidden layer dimension d of the policy network. , Obtain the visual features of the m-th camera viewpoint ; S34, repeat S31 to S33 until you get M Visual features from a single camera perspective; S35, will M Visual features from each camera viewpoint are concatenated in viewpoint order and then compared with the preprocessed robot body state vector. coding features A common input action policy network is used to form fused observation features. .
5. The autonomous motion planning method for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet according to claim 4, characterized in that, The two-dimensional feature grid The expression is: ; in, Indicates the first m Image block feature sequences output from a camera viewpoint image by the visual encoder DINOv2; Represents a sequence of image patch features The two-dimensional feature mesh is restored according to the spatial arrangement of image patches in the original image; Reshape indicates the restoration of the two-dimensional mesh according to the original spatial order of the image patches. Represents the set of real numbers; B Indicates batch size; C The feature dimension representing the features of each image patch; N Indicates the number of image patches, and ; This represents the height of the two-dimensional feature grid, i.e., the number of rows in the two-dimensional feature grid; This represents the width of the two-dimensional feature grid, i.e., the number of columns in the image patch grid; m Indicates the camera viewpoint number. m=1,2,…,M ; The visual features The expression is: ; in, Represents a 1×1 convolution; The fusion observation features The expression is: ; Among them, Transformer represents the feature fusion module based on the Transformer model in the action policy network, which is used to encode and fuse the stitched multi-camera view visual features and robot body state encoding features. The encoder represents the robot's body state. Indicates the concatenation of feature sequences; Indicates the first M Visual features from a camera's perspective.
6. The method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in claim 5, is characterized in that... S4 specifically includes the following steps: S41, Latent variable encoder based on expert action block and their corresponding fusion observation features Estimate the posterior distribution of latent variables In the prior distribution Under the constraints, using the posterior distribution Obtaining latent variables ; latent variables Combined observation features The common input action block decoder generates future action blocks; S42, Based on the posterior distribution Prior distribution Calculate divergence loss ; S43, Targeting Future Actions For each future action within the time frame, calculate the normalized time distance separately; S44. Calculate the unnormalized exponential decay weight of each future action based on the normalized time distance of each future action, and further obtain the normalized weight. S45. Calculate the exponentially decaying action reconstruction loss based on the normalized weights of each future action. ; S46. Divergence loss Reconstruction loss with exponential decay action Combined, the overall training loss is obtained. Determine the overall training loss If the loss is within the set range, output the trained action policy network and pass it to S5; otherwise, use the overall training loss. Dynamically adjust the action policy network; S47. Return to S3 and continue training the action policy network.
7. The autonomous motion planning method for a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet according to claim 6, characterized in that, The future action block The expression is: ; in, Indicates that the trainable parameters are Action block decoder in action policy network, This represents the trainable parameters of the action block decoder. The action policy network is based on time. t The first observation predicted T -1 future action; The divergence loss The expression is: ; in, Represents the relative entropy divergence operation, used to measure the posterior distribution of latent variables. With prior distribution The differences between them; Represents the trainable parameters of the latent variable encoder; S43 i The expression for the normalized time distance of a future action is: ; in, Indicates the first i Normalized time distance of a future action; S44 i The expression for the unnormalized exponentially decaying weight of each future action is: ; in, Indicates the first i The unnormalized exponentially decaying weight of each future action, Represents the natural exponential function; This refers to the weight decay coefficient during action block training. S44 i The expression for the normalized weights of a future action is: ; in, Indicates the first i Normalized weights for future actions, Indicates the summation index, = 0, 1, 2, ..., T -1; Indicates the first Unnormalized exponentially decaying weights of future actions; normalized weights The sum of the normalized weights of all future actions is 1 under different values of the action block training weight decay coefficient λ, as expressed by: ; Exponential decay action reconstructs loss The expression is: ; in, The action policy network is based on time. t The first observation predicted i A future action; Indicates expert action block The Middle t+i A series of future actions; Represents predicted future actions and expert action blocks. The middle corresponds to the absolute error of future continuous actions in each control dimension.
8. The autonomous motion planning method for a robotic arm in confined space plug-in / plug-out operations in a power distribution cabinet according to claim 7, characterized in that, S5 specifically includes the following steps: S51. Use the trained motion policy network to perform autonomous motion planning for the robotic arm and generate multiple candidate actions. S52. Then, for each candidate action, construct unnormalized temporal proximity weights, where the first... The expression for the unnormalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Unnormalized temporal proximity weights for each candidate action Indicates the time proximity integration coefficient; Indicates the first One candidate action; j =0,1,…, K -1; K Indicates the total number of candidate actions; S53. Based on the unnormalized temporal proximity weights of each candidate action, construct the corresponding normalized temporal proximity weights, where the first... The expression for the normalized time proximity weights of the candidate actions is as follows: ; in, Indicates the first Normalized time proximity weights for each candidate action; n Indicates the summation index, n = 0,1,2,…, K -1; Indicates the first n The unnormalized exponentially decaying weights of each candidate action; the sum of the normalized time proximity weights of all candidate actions is 1, expressed as: ; S54. Perform weighted fusion on all candidate actions to obtain the current action to be executed and sent to the robotic arm controller. The expression is as follows: ; in, Indicates the first One candidate action.
9. The method for autonomous motion planning of a robotic arm for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in claim 8, is characterized in that... S6 specifically includes the following steps: S61. Change the currently executing action The data is sent to the robotic arm controller, which controls the robotic arm to perform corresponding joint movements, end-effector pose adjustments, and gripper control. S62. After execution, proceed to the next control cycle and re-acquire the current complete observation set. Then, it returns to S5 to form a closed-loop motion planning process of continuous perception, action block prediction, candidate action fusion and action execution, until the autonomous motion planning task of the robotic arm is completed.
10. An autonomous motion planning system for a robotic arm performing plug-in / plug-out operations in confined spaces of a power distribution cabinet, characterized in that, The robotic arm is configured to perform an autonomous motion planning method for plug-in / plug-out operations in confined spaces of a power distribution cabinet, as described in any one of claims 1 to 9.
