A cable inspection robot dog path following and obstacle avoidance navigation method and device, a terminal device, and a storage medium

Abstract

This invention discloses a method, device, terminal equipment, and storage medium for path following and obstacle avoidance navigation of a cable inspection robot dog, belonging to the field of path planning technology. The method includes: real-time acquisition of the pose, velocity, trajectory radius of the real robot dog, and the pose of the virtual robot dog, as well as the distance and heading deviation between computers, to construct real robot dog state data; synchronous acquisition of obstacle movement and relative position information, calculation of dangerous areas and comprehensive safe distances, and generation of obstacle state data; and simultaneous deduction of the virtual robot dog's linear velocity and future pose to form virtual robot dog state data. A joint state is obtained by fusing the three types of states, and a velocity command is generated by inputting a deep reinforcement learning network based on a priori long short-term memory network, achieving continuous and stable obstacle avoidance and movement control of the robot dog. This invention solves the problems of lack of continuous motion control and poor movement smoothness in existing technologies.

Description

Technical Field

[0001] This invention relates to the technical field of path planning, and in particular to a method, apparatus, terminal device, and storage medium for a cable inspection robot dog to follow paths and avoid obstacles. Background Technology

[0002] The demand for intelligent operation and maintenance of power infrastructure continues to rise, and the use of robotic dogs to replace manual labor in the inspection of cable tunnels and overhead line corridors has become the mainstream development direction in the industry. During the inspection, the robotic dogs are equipped with precision equipment such as infrared thermal imagers and partial discharge detectors, and carry out autonomous inspection work according to the predetermined cable travel routes.

[0003] Existing technologies mostly employ a hierarchical navigation approach that combines global planning with local tracking. This breaks down the overall inspection path into discrete waypoints, with the robot dog sequentially tracking each waypoint to complete the operation. This method only tracks and controls a single point, failing to provide continuous motion control. Consequently, when the equipment encounters turns or areas with dense waypoints, its movement becomes difficult to transition smoothly, frequently resulting in starts, stops, and sharp turns. This not only slows down the inspection progress but also adversely affects airborne precision instruments. Summary of the Invention

[0004] This invention provides a method, device, terminal equipment, and storage medium for cable inspection robot dog path following and obstacle avoidance navigation. The method can solve the problems of lack of continuous motion control and poor smoothness of movement in the prior art.

[0005] An embodiment of the present invention provides a method for path following and obstacle avoidance navigation for a cable inspection robot dog, comprising: Obtain the first pose, first real-time speed, and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog; Based on the first pose and the second pose, calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog; construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed; Obtain the global position, second real-time velocity, and relative position of the obstacle to the real robot dog; Based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin, calculate the radius of the danger sector of the obstacle; calculate the comprehensive safety distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the danger sector; construct the second state data of the obstacle based on the first pose, global position, radius of the danger sector, relative position with the real robot dog, and comprehensive safety distance. The linear velocity of the virtual dog is calculated based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; the future pose of the virtual dog at the next time step is determined based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; the third state data of the virtual dog is constructed based on the future pose and the first pose. Generate the current joint state data based on the first state data, the second state data, and the third state data; The current joint state data is input into a deep reinforcement learning network based on a priori long short-term memory network, so that the deep reinforcement learning network based on a priori long short-term memory network can generate speed commands for a real robot dog. The movement of the real robot dog is controlled and driven according to the speed commands of the real robot dog.

[0006] Furthermore, based on the linear velocity of the virtual machine dog, the preset global reference path, and the second pose, the future pose of the virtual machine dog at the next time step is determined, including: The distance the virtual machine dog will travel in the future time step is calculated based on the linear velocity of the virtual machine dog. Based on the movement distance and the preset global reference path, determine the future waypoint coordinates of the virtual machine dog in the global reference path at the next time step; The future pose of the virtual machine dog at the next time step is determined based on the coordinates of the future waypoints.

[0007] Next, based on the future pose and the first pose, the third state data of the virtual machine dog is constructed, including: Calculate the second global distance deviation between the real robot dog and the future pose based on the future pose and the first pose. The future waypoint state of the virtual machine dog is constructed based on the future pose and the second global distance deviation, and the third state data is obtained.

[0008] Furthermore, based on the first state data, the second state data, and the third state data, the current joint state data is generated, including: Based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the dangerous sector area, the second state data is sorted to obtain the sorted first state sequence. The third-state data are sorted according to the movement distance from largest to smallest to obtain the sorted second-state sequence; Based on the first state data, the sorted first state sequence, and the sorted second state sequence, the current joint state data is obtained.

[0009] Next, based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the danger sector, the second state data is sorted to obtain the sorted first state sequence, including: Based on the second real-time velocity, calculate the angle of the danger sector of the obstacle; based on the radius and angle of the danger sector, determine the danger sector of the obstacle. Based on the first pose, the first real-time velocity, the global position, and the second real-time velocity, the straight-line distance from the real robot dog to the boundary of the dangerous sector along the relative velocity direction is obtained. Based on the first pose, global position, and dangerous sector area, determine whether the real robot dog is within the dangerous sector area; If the real robot dog is in the dangerous sector area, the escape time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is taken as the first obstacle; If the real robot dog is outside the dangerous sector area, the return time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is used as the second obstacle. For all first obstacles, based on the escape time of the real robot dog, sort the corresponding first obstacles in ascending order to obtain the first sequence; For all second obstacles, based on the return time of the real robot dog, the corresponding second obstacles are sorted in descending order to obtain the second sequence; The first sequence is merged with the second sequence, and the first sequence is placed before the second sequence during the merging process to obtain the sorted first state sequence.

[0010] Furthermore, deep reinforcement learning networks based on prior long short-term memory networks include dynamic obstacle long short-term memory network branches, path tracking long short-term memory network branches, and fully connected neural networks; The current joint state data is input into a deep reinforcement learning network based on a priori long short-term memory (LSTM) to enable the deep reinforcement learning network based on LSTM to generate speed commands for a realistic robot dog, including: Based on the dynamic obstacle long short-term memory network branch, the sequential features of each second state data in the current joint state data are extracted to obtain the first hidden state data of the obstacle; Based on the long short-term memory network branch of the tracking path, the third state data is extracted from the current joint state data to obtain the second hidden state data of the virtual machine dog; Based on a fully connected neural network, high-order joint features are extracted by concatenating the first state data, the first hidden state data, and the second hidden state data in the current joint state to obtain the speed command of the real robot dog.

[0011] Next, the training of the deep reinforcement learning network based on the prior long short-term memory network includes: In each training round, the current joint state data sample is obtained; the current joint state data sample is input into a deep reinforcement learning network based on a prior long short-term memory network to obtain action instructions; state transition samples are generated according to the action instructions; each state transition sample includes: the current joint state data sample, the action instruction sample, the reward sample, and the joint state data sample of the next state; the state transition samples are stored in the experience replay cache pool. Based on the temporal difference error, the sample priority is determined, and the state transition sample is selected by the priority experience replay method. The parameters of the deep reinforcement learning network based on the prior long short-term memory network are iteratively updated. Each time the parameters of the deep reinforcement learning network based on the prior long short-term memory network are updated, a predicted speed command for the real robot dog is generated based on the current joint state data sample and a corresponding reward is obtained. The predicted reward is compared with the expected target reward calculated by temporal difference, and the loss function value is calculated in combination with the temporal difference error. The network parameters are then updated based on the loss function value.

[0012] An embodiment of the present invention also provides a path following and obstacle avoidance navigation device for a cable inspection robot dog, including: a body information acquisition module, a real machine state generation module, an obstacle information acquisition module, an obstacle state generation module, a virtual machine prospective inference module, a joint state fusion module, a control command solving module, and a whole machine execution drive module; The body information acquisition module is used to acquire the first pose, first real-time speed and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog. The real machine state generation module is used to calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog based on the first pose and the second pose; and to construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed. The obstacle information acquisition module is used to obtain the global position, second real-time speed, and relative position of the obstacle with respect to the real robot dog. The obstacle state generation module is used to calculate the radius of the dangerous sector area of ​​the obstacle based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin; calculate the comprehensive safe distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the dangerous sector area; and construct the second state data of the obstacle based on the first pose, global position, radius of the dangerous sector area, relative position with the real robot dog, and comprehensive safe distance. The virtual machine look-ahead simulation module is used to calculate the linear velocity of the virtual dog based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; determine the future pose of the virtual dog at the next time step based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; and construct the third state data of the virtual dog based on the future pose and the first pose. The joint state fusion module is used to generate the current joint state data based on the first state data, the second state data, and the third state data; The control command solving module is used to input the current joint state data into the deep reinforcement learning network based on the prior long short-term memory network, so that the deep reinforcement learning network based on the prior long short-term memory network can generate speed commands for the real robot dog. The whole machine execution drive module is used to control and drive the real robot dog to move according to the speed command of the real robot dog.

[0013] This application also provides a terminal device, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a cable inspection robot dog path following and obstacle avoidance navigation method as described in the above embodiments of the invention.

[0014] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a cable inspection robot dog path following and obstacle avoidance navigation method as described in the above embodiments.

[0015] The following benefits can be obtained by implementing the present invention: This invention first collects the pose, real-time speed, and trajectory radius of a real robot dog and the pose of a virtual robot dog. Based on this, the distance and heading angle deviation between the two are calculated, and relevant parameters are combined to construct the real robot dog's state data. Simultaneously, information such as the position, speed, and relative position of each obstacle is acquired, and a comprehensive safe distance is calculated to construct obstacle state data. Then, based on the deviation and reference speed, the linear velocity of the virtual robot dog is calculated to predict its future pose and form corresponding state data. The three types of state data are integrated to obtain joint state data, which is input into a deep reinforcement learning model that integrates a priori long short-term memory network. The model outputs speed commands, and finally, the motion control of the real robot dog is completed based on the commands. By predicting the running state and global path constraints of the virtual robot dog, and combining environmental obstacle state perception and safe distance judgment, the real robot dog can smoothly complete attitude transitions and obstacle avoidance adjustments in turning, waypoint-dense areas, and complex obstacle scenarios, completely reducing abrupt movement phenomena such as start-stop, sharp turns, and emergency avoidance. It can ensure the continuous progress of inspection operations, improve overall inspection efficiency, effectively avoid the risk of obstacle collisions, and prevent damage to airborne precision instruments caused by violent movements, thus greatly improving the safety, stability and reliability of real robot dog inspection operations. Attached Figure Description

[0016] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a cable inspection robot dog path following and obstacle avoidance navigation method provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the structure of a cable inspection robot dog path following and obstacle avoidance navigation device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0021] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0024] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0025] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0026] See Figure 1To address the problems of lack of continuous motion control and poor smoothness of movement in existing technologies, an embodiment of the present invention provides a path following and obstacle avoidance navigation method for a cable inspection robot dog, comprising: S1. Obtain the first pose, first real-time speed, and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog; Specifically, the system collects multiple motion parameters at the current moment. The first pose refers to the position of the real robot dog in the global coordinate system and its own heading angle. The first real-time velocity is the travel speed of the real robot dog in the global environment. The trajectory radius is the instantaneous radius of curvature corresponding to the current trajectory of the real robot dog, that is, the radius of the arc corresponding to the approximate arc of the real robot dog's travel path. In addition, the second pose of the virtual robot dog, which serves as a path reference, is collected simultaneously. The poses of the real machine and the virtual machine, as well as the motion parameters of the real machine, are summarized to provide basic data for the subsequent construction of the first state data of the real robot dog.

[0027] S2. Based on the first pose and the second pose, calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog; construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed; Specifically, based on the first pose of the real robot dog that has been collected... The second pose of the virtual machine dog Complete the deviation calculation, where the first pose Includes the current global x-coordinate of the real device y-axis With heading angle Second pose This is the current global x-coordinate of the virtual facilitator. y-axis With heading angle According to the distance calculation formula Calculate the distance deviation between the two. That is, the first global distance deviation, and then based on the heading calculation formula. The deviation of the heading angle between the real robot dog and the virtual robot dog can be obtained by solving the problem. After completing the calculation of the two types of deviations, the integration... , Combined with the real device's first real-time speed and trajectory radius Compared with the preset target reference speed The data is combined to form the first state data corresponding to the real robot dog. Among them, the preset target reference speed The preset constant reference cruise speed, which does not change with real-time tracking deviation, serves as the speed reference for the robot dog's trajectory tracking control. When there is no virtual or real position or heading deviation, the system... To ensure standard uniform speed travel, if tracking deviation exists, then... Dynamic velocity correction is performed on the base.

[0028] In this invention, by solving the positional and heading deviations between the real robot dog and the virtual guided robot dog, and integrating the motion parameters of the real machine itself to construct the first state data, the tracking error and motion state characteristics of the real machine relative to the desired path can be accurately quantified. This enables an effective representation of the real machine's real-time following situation, providing accurate state basis for subsequent deep reinforcement learning networks to perceive the tracking state of the aircraft, dynamically plan obstacle avoidance and following strategies, and effectively improving the fit of the robot dog's path tracking and the accuracy of motion control.

[0029] S3. Obtain the global position of the obstacle, the second real-time speed, and the relative position with the real robot dog; Specifically, first, obtain the global position of the obstacle in the global coordinate system. Using the second real-time speed, and relying on the global coordinates of the actual machine, coordinate transformation is performed to obtain the relative position of the obstacle in the actual robot dog's coordinate system. This relative position is the basic parameter for subsequent calculation of safe distance and construction of obstacle status data.

[0030] S4. Calculate the radius of the obstacle's dangerous sector area based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin; calculate the comprehensive safe distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the dangerous sector area; construct the obstacle's second state data based on the first pose, global position, radius of the dangerous sector area, relative position with the real robot dog, and comprehensive safe distance. Specifically, the second real-time speed through obstacles Preset speed influence coefficient With safety margin Through calculation formula The radius of the dangerous sector corresponding to the obstacle is obtained by solving the problem. Among them, the speed influence coefficient A constant proportionality coefficient is preset in the system to quantify the impact of obstacle speed on the expansion of the danger zone. The faster the real-time speed of the obstacle, the greater the impact of the obstacle's movement speed on the expansion of the danger zone. The greater the weighted speed risk increment, the more automatically the warning range can be expanded under high-speed conditions involving dynamic obstacles; safety margin To establish a fixed, preset basic safety margin, representing the inherent basic danger radius of an obstacle in a stationary state, it is used to compensate for safety blind spots caused by detection errors, mechanical vibrations, and control lag, ensuring that low-speed or stationary obstacles still have a reliable safety protection boundary. and The sum of these values ​​gives the overall safe distance between the obstacle and the actual robot dog. Subsequently, multiple parameters were integrated to construct the second state data of the obstacle. The parameters used included the relative position of the obstacle in the real machine coordinate system and the real robot dog, which was calculated from the first pose of the real machine. Global coordinates of obstacles The calculated radius of the danger sector Global distance between obstacles and the real robot dog and comprehensive safety distance Finally, obstacle status data is generated according to the standard format. It fully records information on obstacle locations, danger zones, and safe distances. In this invention, dynamic danger boundaries are quantified by relying on obstacle movement rate correlation parameters, and obstacle state data is constructed by integrating multi-dimensional position parameters. This can accurately characterize the spatial distribution of obstacles and the range of collision risks, making it easier to identify the dangerous working conditions of the real machine and classify obstacle priorities. This provides a quantitative environmental basis for reinforcement learning models to make active obstacle avoidance decisions and improves the safety of the real robot dog's operation.

[0031] S5. Calculate the linear velocity of the virtual dog based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; determine the future pose of the virtual dog at the next time step based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; construct the third state data of the virtual dog based on the future pose and the first pose. Preferably, the future pose of the virtual machine dog at a future time step is determined based on the linear velocity of the virtual machine dog, the preset global reference path, and the second pose, including: The distance the virtual machine dog will travel in the future time step is calculated based on the linear velocity of the virtual machine dog. Based on the movement distance and the preset global reference path, determine the future waypoint coordinates of the virtual machine dog in the global reference path at the next time step; The future pose of the virtual machine dog at the next time step is determined based on the coordinates of the future waypoints.

[0032] Preferably, the third state data of the virtual machine dog is constructed based on the future pose and the first pose, including: Calculate the second global distance deviation between the real robot dog and the future pose based on the future pose and the first pose. The future waypoint state of the virtual machine dog is constructed based on the future pose and the second global distance deviation, and the third state data is obtained.

[0033] Specifically, based on the first global distance deviation The deviation in heading angle between a real robot dog and a virtual robot dog Compared with the preset target reference speed Through formula Solve for the virtual machine dog line speed ; within the formula , is the preset proportional correction coefficient, and sat is the amplitude limiting saturation function used to constrain the correction amplitude of the heading deviation; The `clip` function forces the calculated boot speed to be constrained to the preset lower and upper limits of the virtual machine dongle's running speed. Within the range, to prevent the guide speed from becoming negative or exceeding the speed limit, and to prevent the actual device from deviating too much from the target speed. Automatic speed reduction, based on The kinematic model constrains the laws of motion, among which... Represents the global location of the virtual machine dog. , These are the differentials of position and heading, respectively. It is the global reference path arc length. The unit tangent vector at that point, This is the angular velocity function corresponding to the path curvature, ensuring the virtual robot dog follows the reference path smoothly throughout the entire journey. After obtaining the linear velocity, it is combined with a fixed control period. With forward steps ,pass Calculate the future The arc length of the step , It refers to the distance the virtual machine dog travels in the corresponding future time step.

[0034] Subsequently, based on the cable routing during the inspection, either A* or Dijkstra's algorithm is used to solve for the optimal route from the inspection start point to the end point, generating a global reference path. This path fully preserves the mapping relationship between arc length, planar coordinates, and heading angle. Based on this mapping relationship, the calculated travel distance... Look up the table and interpolate to determine the future number. Step global waypoint position Then, by combining the geometric orientation information inherent in the global reference path, the expected heading angle corresponding to the global waypoint position is extracted. , will be the future The global future pose of the virtual machine dog is determined by integrating the global waypoint position with the corresponding expected heading angle. Then, by converting the first pose and heading angle of the actual robot dog, the pose of the future waypoint in the actual machine coordinate system is calculated. At the same time, through Calculate the global distance between the future waypoint and the actual robot dog. That is, the second global distance deviation, which is ultimately defined according to the future waypoint state of the virtual machine dog. By summarizing the four characteristic parameters, the third state data of the virtual machine dog is completed.

[0035] In this invention, the virtual robot dog's guidance speed is dynamically adjusted through deviation feedback, and its trajectory is constrained by the robot dog's kinematic model. This allows the virtual guide to adaptively match the real robot's following state, consistently maintaining a smooth movement along the globally optimal reference path, effectively avoiding problems such as speed exceeding limits and trajectory deviation. Simultaneously, by forward-looking calculation of future waypoints and pose information and constructing standardized path-looking state data, the future trend and tracking deviation of the global path can be accurately depicted. This provides continuous and accurate forward-looking prior information for subsequent robot dog path-tracking decisions, significantly improving the consistency and trajectory adaptability of the real robot dog's path following.

[0036] S6. Generate the current joint state data based on the first state data, the second state data, and the third state data; Preferably, the current joint state data is generated based on the first state data, the second state data, and the third state data, including: Based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the dangerous sector area, the second state data is sorted to obtain the sorted first state sequence. The third-state data are sorted according to the movement distance from largest to smallest to obtain the sorted second-state sequence; Based on the first state data, the sorted first state sequence, and the sorted second state sequence, the current joint state data is obtained.

[0037] Preferably, based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the danger sector region, the second state data is sorted to obtain a sorted first state sequence, including: Based on the second real-time velocity, calculate the angle of the danger sector of the obstacle; based on the radius and angle of the danger sector, determine the danger sector of the obstacle. Based on the first pose, the first real-time velocity, the global position, and the second real-time velocity, the straight-line distance from the real robot dog to the boundary of the dangerous sector along the relative velocity direction is obtained. Based on the first pose, global position, and dangerous sector area, determine whether the real robot dog is within the dangerous sector area; If the real robot dog is in the dangerous sector area, the escape time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is taken as the first obstacle; If the real robot dog is outside the dangerous sector area, the return time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is used as the second obstacle. For all first obstacles, based on the escape time of the real robot dog, sort the corresponding first obstacles in ascending order to obtain the first sequence; For all second obstacles, based on the return time of the real robot dog, the corresponding second obstacles are sorted in descending order to obtain the second sequence; The first sequence is merged with the second sequence, and the first sequence is placed before the second sequence during the merging process to obtain the sorted first state sequence.

[0038] Specifically, relying on the second real-time velocity of the obstacle Combined with the radius of the already solved dangerous sector region Through the angle calculation formula Obtain the angle of the dangerous sector area The complete danger sector area corresponding to the obstacle is delineated using two parameters: radius and angle. Then, by combining the first pose, first real-time velocity of the real robot dog, the global position of the obstacle, and the second real-time velocity of the obstacle, the relative velocity of the real robot dog with respect to the obstacle is calculated. Then, calculate the straight-line distance from the current position of the actual aircraft to the boundary of the dangerous sector area along the direction of the relative velocity. This distance is the distance within the area. Outside the district Based on the real robot's first pose, the global position of obstacles, and the pre-defined danger sector, the location of the real robot dog is determined: when the real robot falls inside the danger sector, the straight-line distance is recorded as... According to the formula Solve for escape time This obstacle is classified as the first obstacle; when the actual aircraft is outside the danger sector, the calculated distance is recorded as... ,pass Solving for regression time The obstacle was classified as a second obstacle. After classification, priority sorting was performed: all first obstacles were sorted from smallest to largest based on their escape time values, generating a first sequence; all second obstacles were sorted from largest to smallest based on their return time values, generating a second sequence; when merging the two sets of sequences, the first sequence was prioritized and the second sequence was placed last, ultimately forming the first state sequence after obstacle-side sorting. Simultaneously, for the third state data corresponding to the virtual robot dog, referring to the sorting rule of waypoints and distances between the real robot dog and the real robot dog from farthest to closest, the data was sorted according to the corresponding movement distance from largest to smallest, resulting in the sorted second state sequence. Finally, the first state data of the real robot dog, the sorted first state sequence, and the sorted second state sequence were integrated to generate the final current joint state data.

[0039] In this invention, obstacle danger zones are quantified and obstacle danger levels are classified by combining relative motion parameters. Obstacle priority is ranked based on avoidance-related time indicators. Then, the results of the path look-ahead state ranking are fused to generate joint state data. This can distinguish the emergency avoidance weights of different obstacles and the look-ahead priority of path tracking, and achieve orderly organization of environmental hazard information and path prior information. This provides hierarchical input features for the subsequent reinforcement learning network to accurately output control commands, taking into account both the obstacle avoidance safety and path following effect of the robot dog.

[0040] S7. Input the current joint state data into the deep reinforcement learning network based on the prior long short-term memory network so that the deep reinforcement learning network based on the prior long short-term memory network can generate speed commands for the real robot dog. Preferably, the deep reinforcement learning network based on prior long short-term memory network includes a dynamic obstacle long short-term memory network branch, a path tracking long short-term memory network branch, and a fully connected neural network; The current joint state data is input into a deep reinforcement learning network based on a priori long short-term memory (LSTM) to enable the deep reinforcement learning network based on LSTM to generate speed commands for a realistic robot dog, including: Based on the dynamic obstacle long short-term memory network branch, the sequential features of each second state data in the current joint state data are extracted to obtain the first hidden state data of the obstacle; Based on the long short-term memory network branch of the tracking path, the third state data is extracted from the current joint state data to obtain the second hidden state data of the virtual machine dog; Based on a fully connected neural network, high-order joint features are extracted by concatenating the first state data, the first hidden state data, and the second hidden state data in the current joint state to obtain the speed command of the real robot dog.

[0041] Specifically, the deep reinforcement learning network based on prior long short-term memory (LSTM) networks is an optimized and improved LSTM network that integrates task prior control logic. It embeds domain prior constraint information for robot dog trajectory tracking and dynamic obstacle avoidance into the traditional LSTM temporal memory units, specifically enhancing the effective extraction capability of temporal state features and avoiding feature redundancy and temporal forgetting problems in general networks, thus adapting to continuous motion control scenarios for robots. The network's overall architecture is divided into three parts: a dynamic obstacle LSTM network branch, a tracking path LSTM network branch, and a post-connected fully connected neural network. The concatenated current joint state data is completely fed into this deep reinforcement learning network, which then performs layer-by-layer computation to ultimately output the speed control commands for the real robot dog.

[0042] During the step-by-step computation within the network, the dynamic obstacle long short-term memory network branch is run first. This branch receives the first state sequence of obstacles in the joint state after sorting them from high to low danger priority. The LSTM, relying on its own temporal memory structure, reads the state of each obstacle frame by frame along the sorting order of the obstacles, extracts the temporal dependencies of position, speed, and danger level between different obstacles, and obtains the hidden state of the obstacles (i.e., the first hidden state data). The hidden state of the obstacles is a low-dimensional feature vector that summarizes all the temporal environmental features of the obstacles, and is used to characterize the overall dynamic danger distribution law of the obstacles in the environment.

[0043] Then, the Long Short-Term Memory (LSTM) branch for tracking the path is activated. This branch is input with the third state sequence of the virtual machine dog sorted from farthest to nearest. The LSTM reads the state of each future waypoint in the order of distance from the waypoint, and mines the temporal change pattern of the position and heading of the waypoints before and after the path to obtain the hidden state of the virtual machine dog (i.e., the second hidden state data). The hidden state of the virtual machine dog is a feature vector that represents the future trend of the global reference path, and condenses the geometric and positional information of the subsequent path.

[0044] After completing the feature extraction of the two LSTMs, the first state data of the real machine, the hidden state of the obstacle, and the hidden state of the virtual machine dog are concatenated to obtain the joint feature. Then The data is fed into a multi-layer fully connected neural network. Each fully connected layer is followed by a ReLU activation function to complete high-order feature extraction. At the end of the fully connected network, a dual output head is set. The continuous motion head calculates and outputs the linear velocity command and angular velocity command (i.e., velocity command) required by the real robot dog. After physical parameter limiting processing, it can be used as motion execution command at the underlying level of the real machine.

[0045] In this invention, a dual-branch long short-term memory network is used to extract the temporal risk features of obstacles and the prospective change features of the path, and then the multi-dimensional information is fused through a fully connected network to generate control commands. This can fully explore the dynamic laws of environmental obstacles and the future trend of the path, realize the deep feature fusion of environmental information and tracking information, improve the network's perception ability in complex inspection scenarios, and ensure that the output speed commands take into account both the accuracy of the aircraft path following and the reliability of dynamic obstacle avoidance.

[0046] Preferably, in each training round, the current joint state data sample is obtained; the current joint state data sample is input into a deep reinforcement learning network based on a prior long short-term memory network to obtain action instructions; state transition samples are generated according to the action instructions; wherein, each state transition sample includes: the current joint state data sample, the action instruction sample, the reward sample, and the joint state data sample of the next state; the state transition samples are stored in the experience replay cache pool. Based on the temporal difference error, the sample priority is determined, and the state transition sample is selected by the priority experience replay method. The parameters of the deep reinforcement learning network based on the prior long short-term memory network are iteratively updated. Each time the parameters of the deep reinforcement learning network based on the prior long short-term memory network are updated, a predicted speed command for the real robot dog is generated based on the current joint state data sample and a corresponding reward is obtained. The predicted reward is compared with the expected target reward calculated by temporal difference, and the loss function value is calculated in combination with the temporal difference error. The network parameters are then updated based on the loss function value.

[0047] Specifically, the current joint state data sample is collected and fed into a deep reinforcement learning network based on prior long short-term memory. The network then outputs the action command corresponding to the real robot dog. After the real robot dog executes the action command, the environment undergoes a state transition, which is used to construct a state transition sample. A single state transition sample consists of four parts: the current joint state sample, the output action command sample, the real-time reward sample, and the next joint state sample. All state transition samples are stored in the experience pool.

[0048] The real-time reward for each state transition sample is calculated according to the total reward formula. Calculate each item separately. , , The task time penalty is fixed by pre-setting a weight constant. This is used to encourage the robot dog to speed up its work progress; path-following rewards , , The value of this reward is a positive zero tuning constant. The smaller the position and heading deviation between the real machine and the virtual guide, the higher the reward value; collision avoidance reward. Based on the dynamic hazard classification of obstacles, three-segment values ​​are used to dynamically expand the safety radius. + Related to the obstacle's own speed: when (When the real machine is in a high-risk collision zone), a fixed penalty of -0.1 is applied per instance; when (The real machine is in a transitional danger zone), and the reward is calculated based on linear interpolation of the distance; when (With the real machine within a safe range), the collision avoidance reward is set to 0. The segmented collision avoidance reward can guide the model to avoid blindly risky or overly conservative driving strategies, ensuring smooth convergence of training.

[0049] In the iterative training phase, temporal difference error is used as the criterion for prioritizing experience samples. A priority experience replay algorithm is used to select high-value state transition samples from the experience pool for iterative training. In a single training cycle, the network reads the current joint state sample to generate a real-machine prediction speed command. After the real machine runs, the network calculates the immediate reward based on environmental feedback. The difference between the actual predicted reward and the expected target reward calculated by the temporal difference algorithm is calculated, and the loss function is solved using the temporal difference error. The network parameters are then iteratively updated along the backpropagation path based on the value of the loss function. This process is repeated until the loss function converges, ultimately completing the training of the entire deep reinforcement learning network based on prior long short-term memory.

[0050] In this invention, a multi-dimensional segmented reward mechanism is used to precisely constrain the direction of model optimization, guiding the network parameters to continuously update towards the goals of accurate path tracking, efficient task completion, and active obstacle avoidance. Combined with priority experience replay and temporal difference error, the network iterative optimization is completed, which can efficiently screen key training samples, balance path tracking and collision avoidance constraints, avoid extreme training strategies, accelerate model convergence speed, and enable the trained network to adapt to complex inspection environments and stably output control commands that balance following accuracy and operational safety.

[0051] S8. Control and drive the real robot dog to move according to the speed command of the real robot dog.

[0052] Specifically, in practical inspection applications, the system uses a priori long short-term memory deep reinforcement learning network to receive real-time joint state data, consisting of the real machine's status, obstacle status, and virtual guided look-ahead path status. It then outputs speed commands in real-time, composed of linear and angular velocity components. These commands are sent to the underlying motion controller of the real robot dog. The controller adjusts the robot dog's straight-line speed and steering speed based on these values, driving the robot dog to autonomously move and turn along a reference path, completing the on-site inspection operation. After completing a single control cycle Δt, the robot dog's actual global pose changes. The device's positioning and environmental perception modules simultaneously update the real machine's position, surrounding obstacle positions and motion information, and virtual guide point data, reassembling new joint state data and inputting it into the priori long short-term memory deep reinforcement learning network. This continuously generates a new round of control commands in real-time, enabling the entire machine to autonomously follow and avoid obstacles during inspection.

[0053] In this invention, a real-time closed-loop control of the robot dog's inspection process is achieved by relying on a trained reinforcement learning network. By sensing the robot's status, environmental obstacles, and forward path information in real time, precise linear and angular velocity control commands are dynamically output to drive the robot dog to adaptively complete path following and dynamic obstacle avoidance. Simultaneously, by periodically updating environmental and robot status data in real time, optimal control commands are continuously generated iteratively, effectively ensuring the robot dog has continuous, stable, and autonomous operation capabilities in complex inspection scenarios, significantly improving the overall intelligence level, operational safety, and trajectory tracking accuracy of the inspection process.

[0054] See Figure 2 This invention provides a cable inspection robot dog path following and obstacle avoidance navigation device, comprising: a body information acquisition module, a real machine state generation module, an obstacle information acquisition module, an obstacle state generation module, a virtual machine prospective inference module, a joint state fusion module, a control command solving module, and a whole machine execution drive module. The body information acquisition module is used to acquire the first pose, first real-time speed and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog. The real machine state generation module is used to calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog based on the first pose and the second pose; and to construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed. The obstacle information acquisition module is used to obtain the global position, second real-time speed, and relative position of the obstacle with respect to the real robot dog. The obstacle state generation module is used to calculate the radius of the dangerous sector area of ​​the obstacle based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin; calculate the comprehensive safe distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the dangerous sector area; and construct the second state data of the obstacle based on the first pose, global position, radius of the dangerous sector area, relative position with the real robot dog, and comprehensive safe distance. The virtual machine look-ahead simulation module is used to calculate the linear velocity of the virtual dog based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; determine the future pose of the virtual dog at the next time step based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; and construct the third state data of the virtual dog based on the future pose and the first pose. The joint state fusion module is used to generate the current joint state data based on the first state data, the second state data, and the third state data; The control command solving module is used to input the current joint state data into the deep reinforcement learning network based on the prior long short-term memory network, so that the deep reinforcement learning network based on the prior long short-term memory network can generate speed commands for the real robot dog. The whole machine execution drive module is used to control and drive the real robot dog to move according to the speed command of the real robot dog.

[0055] Specifically, in a preferred embodiment of the present invention, the body information acquisition module completes the acquisition of multiple types of motion parameters at the current moment. Its first pose refers to the position of the real robot dog in the global coordinate system and its own heading angle. The first real-time speed is the travel speed of the real robot dog in the global environment. The trajectory radius is the instantaneous radius of curvature corresponding to the current motion trajectory of the real robot dog, that is, the radius of the arc corresponding to the approximate arc of the real robot dog's travel path. In addition, the second pose of the virtual robot dog, which serves as a path reference, is acquired simultaneously. The poses of the real machine and the virtual machine and the motion parameters of the real machine are summarized to reserve basic data for the subsequent construction of the first state data of the real robot dog.

[0056] Specifically, the real machine status generation module uses the first pose of the real robot dog that has been collected. The second pose of the virtual machine dog Complete the deviation calculation, where the first pose Includes the current global x-coordinate of the real device y-axis With heading angle Second pose This is the current global x-coordinate of the virtual facilitator. y-axis With heading angle According to the distance calculation formula Calculate the distance deviation between the two. That is, the first global distance deviation, and then based on the heading calculation formula. The deviation of the heading angle between the real robot dog and the virtual robot dog can be obtained by solving the problem. After completing the calculation of the two types of deviations, the integration... , Combined with the real device's first real-time speed and trajectory radius Compared with the preset target reference speed The data is combined to form the first state data corresponding to the real robot dog. Among them, the preset target reference speed The preset constant reference cruise speed, which does not change with real-time tracking deviation, serves as the speed reference for the robot dog's trajectory tracking control. When there is no virtual or real position or heading deviation, the system... To ensure standard uniform speed travel, if tracking deviation exists, then... Dynamic velocity correction is performed on the base.

[0057] In this invention, by solving the positional and heading deviations between the real robot dog and the virtual guided robot dog, and integrating the motion parameters of the real machine itself to construct the first state data, the tracking error and motion state characteristics of the real machine relative to the desired path can be accurately quantified. This enables an effective representation of the real machine's real-time following situation, providing accurate state basis for subsequent deep reinforcement learning networks to perceive the tracking state of the aircraft, dynamically plan obstacle avoidance and following strategies, and effectively improving the fit of the robot dog's path tracking and the accuracy of motion control.

[0058] Specifically, the obstacle information acquisition module first obtains the global position of the obstacle in the global coordinate system. Using the second real-time speed, and relying on the global coordinates of the actual machine, coordinate transformation is performed to obtain the relative position of the obstacle in the actual robot dog's coordinate system. This relative position is the basic parameter for subsequent calculation of safe distance and construction of obstacle status data.

[0059] Specifically, the obstacle state generation module uses the obstacle's second real-time velocity... Preset speed influence coefficient With safety margin Through calculation formula The radius of the dangerous sector corresponding to the obstacle is obtained by solving the problem. Among them, the speed influence coefficient A constant proportionality coefficient is preset in the system to quantify the impact of obstacle speed on the expansion of the danger zone. The faster the real-time speed of the obstacle, the greater the impact of the obstacle's movement speed on the expansion of the danger zone. The greater the weighted speed risk increment, the more automatically the warning range can be expanded under high-speed conditions involving dynamic obstacles; safety margin To establish a fixed, preset basic safety margin, representing the inherent basic danger radius of an obstacle in a stationary state, it is used to compensate for safety blind spots caused by detection errors, mechanical vibrations, and control lag, ensuring that low-speed or stationary obstacles still have a reliable safety protection boundary. and The sum of these values ​​gives the overall safe distance between the obstacle and the actual robot dog. Subsequently, multiple parameters were integrated to construct the second state data of the obstacle. The parameters used included the relative position of the obstacle in the real machine coordinate system and the real robot dog, which was calculated from the first pose of the real machine. Global coordinates of obstacles The calculated radius of the danger sector Global distance between obstacles and the real robot dog and comprehensive safety distance Finally, obstacle status data is generated according to the standard format. It fully records information on obstacle locations, danger zones, and safe distances. In this invention, dynamic danger boundaries are quantified by relying on obstacle movement rate correlation parameters, and obstacle state data is constructed by integrating multi-dimensional position parameters. This can accurately characterize the spatial distribution of obstacles and the range of collision risks, making it easier to identify the dangerous working conditions of the real machine and classify obstacle priorities. This provides a quantitative environmental basis for reinforcement learning models to make active obstacle avoidance decisions and improves the safety of the real robot dog's operation.

[0060] Preferably, the virtual machine look-ahead simulation module includes: a look-ahead waypoint pose solving unit and a path look-ahead state construction unit; The look-ahead waypoint pose solving unit is used to calculate the movement distance of the virtual machine dog in the future time step based on the linear velocity of the virtual machine dog; determine the future waypoint coordinates of the virtual machine dog in the global reference path in the future time step based on the movement distance and the preset global reference path; and determine the future pose of the virtual machine dog in the future time step based on the future waypoint coordinates.

[0061] The path-lookahead state construction unit is used to calculate the second global distance deviation between the real robot dog and the future pose based on the future pose and the first pose; and to construct the future waypoint state of the virtual robot dog based on the future pose and the second global distance deviation to obtain the third state data.

[0062] Specifically, based on the first global distance deviation The deviation in heading angle between a real robot dog and a virtual robot dog Compared with the preset target reference speed Through formula Solve for the virtual machine dog line speed ; within the formula , is the preset proportional correction coefficient, and sat is the amplitude limiting saturation function used to constrain the correction amplitude of the heading deviation; The `clip` function forces the calculated boot speed to be constrained to the preset lower and upper limits of the virtual machine dongle's running speed. Within the range, to prevent the guide speed from becoming negative or exceeding the speed limit, and to prevent the actual device from deviating too much from the target speed. Automatic speed reduction, based on The kinematic model constrains the laws of motion, among which... Represents the global location of the virtual machine dog. , These are the differentials of position and heading, respectively. It is the global reference path arc length. The unit tangent vector at that point, This is the angular velocity function corresponding to the path curvature, ensuring the virtual robot dog follows the reference path smoothly throughout the entire journey. After obtaining the linear velocity, it is combined with a fixed control period. With forward steps ,pass Calculate the future The arc length of the step , It refers to the distance the virtual machine dog travels in the corresponding future time step.

[0063] Subsequently, based on the cable routing during the inspection, either A* or Dijkstra's algorithm is used to solve for the optimal route from the inspection start point to the end point, generating a global reference path. This path fully preserves the mapping relationship between arc length, planar coordinates, and heading angle. Based on this mapping relationship, the calculated travel distance... Look up the table and interpolate to determine the future number. Step global waypoint position Then, by combining the geometric orientation information inherent in the global reference path, the expected heading angle corresponding to the global waypoint position is extracted. , will be the future The global future pose of the virtual machine dog is determined by integrating the global waypoint position with the corresponding expected heading angle. Then, by converting the first pose and heading angle of the actual robot dog, the pose of the future waypoint in the actual machine coordinate system is calculated. At the same time, through Calculate the global distance between the future waypoint and the actual robot dog. That is, the second global distance deviation, which is ultimately defined according to the future waypoint state of the virtual machine dog. By summarizing the four characteristic parameters, the third state data of the virtual machine dog is completed.

[0064] In this invention, the virtual robot dog's guidance speed is dynamically adjusted through deviation feedback, and its trajectory is constrained by the robot dog's kinematic model. This allows the virtual guide to adaptively match the real robot's following state, consistently maintaining a smooth movement along the globally optimal reference path, effectively avoiding problems such as speed exceeding limits and trajectory deviation. Simultaneously, by forward-looking calculation of future waypoints and pose information and constructing standardized path-looking state data, the future trend and tracking deviation of the global path can be accurately depicted. This provides continuous and accurate forward-looking prior information for subsequent robot dog path-tracking decisions, significantly improving the consistency and trajectory adaptability of the real robot dog's path following.

[0065] Preferably, the joint state fusion module includes: a state sorting and joint construction unit; The state sorting and joint construction unit is used to sort the second state data based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the dangerous sector region to obtain the sorted first state sequence. This includes: calculating the angle of the dangerous sector region of the obstacle based on the second real-time velocity; determining the dangerous sector region of the obstacle based on the radius and angle of the dangerous sector region; obtaining the straight-line distance from the real robot dog to the boundary of the dangerous sector region along the relative velocity direction based on the first pose, first real-time velocity, global position, and second real-time velocity; determining whether the real robot dog is within the dangerous sector region based on the first pose, global position, and dangerous sector region; if the real robot dog is within the dangerous sector region, determining the escape time based on the straight-line distance and relative velocity, and designating the corresponding obstacle as the first obstacle; if... When the real robot dog is outside the danger sector, its return time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is designated as the second obstacle. For all first obstacles, they are sorted in ascending order based on the real robot dog's escape time, resulting in a first sequence. For all second obstacles, they are sorted in descending order based on the real robot dog's return time, resulting in a second sequence. The first and second sequences are merged, with the first sequence placed before the second sequence, resulting in a sorted first state sequence. The third state data is then sorted in descending order of movement distance, resulting in a sorted second state sequence. Based on the first state data, the sorted first state sequence, and the sorted second state sequence, the current joint state data is obtained.

[0066] Specifically, relying on the second real-time velocity of the obstacle Combined with the radius of the already solved dangerous sector region Through the angle calculation formula Obtain the angle of the dangerous sector area The complete danger sector area corresponding to the obstacle is delineated using two parameters: radius and angle. Then, by combining the first pose, first real-time velocity of the real robot dog, the global position of the obstacle, and the second real-time velocity of the obstacle, the relative velocity of the real robot dog with respect to the obstacle is calculated. Then, calculate the straight-line distance from the current position of the actual aircraft to the boundary of the dangerous sector area along the direction of the relative velocity. This distance is the distance within the area. Outside the district Based on the real robot's first pose, the global position of obstacles, and the pre-defined danger sector, the location of the real robot dog is determined: when the real robot falls inside the danger sector, the straight-line distance is recorded as... According to the formula Solve for escape time This obstacle is classified as the first obstacle; when the actual aircraft is outside the danger sector, the calculated distance is recorded as... ,pass Solving for regression time The obstacle was classified as a second obstacle. After classification, priority sorting was performed: all first obstacles were sorted from smallest to largest based on their escape time values, generating a first sequence; all second obstacles were sorted from largest to smallest based on their return time values, generating a second sequence; when merging the two sets of sequences, the first sequence was prioritized and the second sequence was placed last, ultimately forming the first state sequence after obstacle-side sorting. Simultaneously, for the third state data corresponding to the virtual robot dog, referring to the sorting rule of waypoints and distances between the real robot dog and the real robot dog from farthest to closest, the data was sorted according to the corresponding movement distance from largest to smallest, resulting in the sorted second state sequence. Finally, the first state data of the real robot dog, the sorted first state sequence, and the sorted second state sequence were integrated to generate the final current joint state data.

[0067] In this invention, obstacle danger zones are quantified and obstacle danger levels are classified by combining relative motion parameters. Obstacle priority is ranked based on avoidance-related time indicators. Then, the results of the path look-ahead state ranking are fused to generate joint state data. This can distinguish the emergency avoidance weights of different obstacles and the look-ahead priority of path tracking, and achieve orderly organization of environmental hazard information and path prior information. This provides hierarchical input features for the subsequent reinforcement learning network to accurately output control commands, taking into account both the obstacle avoidance safety and path following effect of the robot dog.

[0068] It is understood that the above-described device embodiments correspond to the method embodiments of the present invention, and can realize the cable inspection robot dog path following and obstacle avoidance navigation method provided by any of the above-described method embodiments of the present invention.

[0069] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0070] Accordingly, one embodiment of this application also provides a terminal device, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a cable inspection robot dog path following and obstacle avoidance navigation method as described above.

[0071] The processor is used to control the overall operation of the terminal device to complete all or part of the steps of the above-described cable inspection robot dog path following and obstacle avoidance navigation method.

[0072] Memory is used to store various types of data to support the operation of the terminal device. This data may include, for example, instructions for any application or method operating on the terminal device, as well as application-related data. The memory can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0073] In an exemplary embodiment, the terminal device may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform a cable inspection robot dog path following and obstacle avoidance navigation method as described in any of the above embodiments, and achieve the same technical effect as the above method.

[0074] Accordingly, the present invention provides a computer-readable storage medium including a computer program, which, when executed by a processor, implements the steps of a cable inspection robot dog path following and obstacle avoidance navigation method as described in any of the above embodiments. For example, the computer-readable storage medium may be the aforementioned memory including the computer program, which may be executed by a processor of a terminal device to complete the cable inspection robot dog path following and obstacle avoidance navigation method as described in any of the above embodiments, and achieve the same technical effects as the aforementioned method.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for path following and obstacle avoidance navigation in a cable inspection robot dog, characterized in that, include: Obtain the first pose, first real-time speed, and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog; Based on the first pose and the second pose, calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog; construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed; Obtain the global position, second real-time velocity, and relative position of the obstacle to the real robot dog; Based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin, calculate the radius of the danger sector of the obstacle; calculate the comprehensive safety distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the danger sector; construct the second state data of the obstacle based on the first pose, global position, radius of the danger sector, relative position with the real robot dog, and comprehensive safety distance. The linear velocity of the virtual dog is calculated based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; the future pose of the virtual dog at the next time step is determined based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; the third state data of the virtual dog is constructed based on the future pose and the first pose. Generate the current joint state data based on the first state data, the second state data, and the third state data; The current joint state data is input into a deep reinforcement learning network based on a priori long short-term memory network, so that the deep reinforcement learning network based on a priori long short-term memory network can generate speed commands for a real robot dog. The movement of the real robot dog is controlled and driven according to the speed commands of the real robot dog.

2. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 1, characterized in that, The future pose of the virtual machine dog at the next time step is determined based on the linear velocity of the virtual machine dog, the preset global reference path, and the second pose, including: The distance the virtual machine dog will travel in the future time step is calculated based on the linear velocity of the virtual machine dog. Based on the movement distance and the preset global reference path, determine the future waypoint coordinates of the virtual machine dog in the global reference path at the next time step; The future pose of the virtual machine dog at the next time step is determined based on the coordinates of the future waypoints.

3. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 2, characterized in that, The third state data of the virtual machine dog is constructed based on the future pose and the first pose, including: Calculate the second global distance deviation between the real robot dog and the future pose based on the future pose and the first pose. The future waypoint state of the virtual machine dog is constructed based on the future pose and the second global distance deviation, and the third state data is obtained.

4. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 3, characterized in that, The current joint state data is generated based on the first state data, the second state data, and the third state data, including: Based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the dangerous sector area, the second state data is sorted to obtain the sorted first state sequence. The third-state data are sorted according to the movement distance from largest to smallest to obtain the sorted second-state sequence; Based on the first state data, the sorted first state sequence, and the sorted second state sequence, the current joint state data is obtained.

5. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 4, characterized in that, Based on the first pose, first real-time velocity, global position, second real-time velocity, and radius of the danger sector, the second state data is sorted to obtain the sorted first state sequence, including: Based on the second real-time velocity, calculate the angle of the danger sector of the obstacle; based on the radius and angle of the danger sector, determine the danger sector of the obstacle. Based on the first pose, the first real-time velocity, the global position, and the second real-time velocity, the straight-line distance from the real robot dog to the boundary of the dangerous sector along the relative velocity direction is obtained. Based on the first pose, global position, and dangerous sector area, determine whether the real robot dog is within the dangerous sector area; If the real robot dog is in the dangerous sector area, the escape time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is taken as the first obstacle; If the real robot dog is outside the dangerous sector area, the return time is determined based on the straight-line distance and relative speed, and the corresponding obstacle is used as the second obstacle. For all first obstacles, based on the escape time of the real robot dog, sort the corresponding first obstacles in ascending order to obtain the first sequence; For all second obstacles, based on the return time of the real robot dog, the corresponding second obstacles are sorted in descending order to obtain the second sequence; The first sequence is merged with the second sequence, and the first sequence is placed before the second sequence during the merging process to obtain the sorted first state sequence.

6. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 5, characterized in that, Deep reinforcement learning networks based on prior long short-term memory networks include dynamic obstacle long short-term memory network branches, path tracking long short-term memory network branches, and fully connected neural networks; The current joint state data is input into a deep reinforcement learning network based on a priori long short-term memory (LSTM) to enable the deep reinforcement learning network based on LSTM to generate speed commands for a realistic robot dog, including: Based on the dynamic obstacle long short-term memory network branch, the sequential features of each second state data in the current joint state data are extracted to obtain the first hidden state data of the obstacle; Based on the long short-term memory network branch of the tracking path, the third state data is extracted from the current joint state data to obtain the second hidden state data of the virtual machine dog; Based on a fully connected neural network, high-order joint features are extracted by concatenating the first state data, the first hidden state data, and the second hidden state data in the current joint state to obtain the speed command of the real robot dog.

7. The cable inspection robot dog path following and obstacle avoidance navigation method as described in claim 6, characterized in that, Training of deep reinforcement learning networks based on prior long short-term memory networks includes: In each training round, the current joint state data sample is obtained; the current joint state data sample is input into a deep reinforcement learning network based on a prior long short-term memory network to obtain action instructions; state transition samples are generated according to the action instructions; each state transition sample includes: the current joint state data sample, the action instruction sample, the reward sample, and the joint state data sample of the next state; the state transition samples are stored in the experience replay cache pool. Based on the temporal difference error, the sample priority is determined, and the state transition sample is selected by the priority experience replay method. The parameters of the deep reinforcement learning network based on the prior long short-term memory network are iteratively updated. Each time the parameters of the deep reinforcement learning network based on the prior long short-term memory network are updated, a predicted speed command for the real robot dog is generated based on the current joint state data sample and a corresponding reward is obtained. The predicted reward is compared with the expected target reward calculated by temporal difference, and the loss function value is calculated in combination with the temporal difference error. The network parameters are then updated based on the loss function value.

8. A cable inspection robot dog path following and obstacle avoidance navigation device, characterized in that, include: The system includes a main body information acquisition module, a real machine state generation module, an obstacle information acquisition module, an obstacle state generation module, a virtual machine prospective simulation module, a joint state fusion module, a control command solving module, and a whole machine execution drive module. The body information acquisition module is used to acquire the first pose, first real-time speed and trajectory radius of the real robot dog at the current moment, as well as the second pose of the virtual robot dog. The real machine state generation module is used to calculate the first global distance deviation and heading angle deviation between the real robot dog and the virtual robot dog based on the first pose and the second pose; and to construct the first state data of the real robot dog based on the first global distance deviation, heading angle deviation, first real-time speed, trajectory radius and preset target reference speed. The obstacle information acquisition module is used to obtain the global position, second real-time speed, and relative position of the obstacle with respect to the real robot dog. The obstacle state generation module is used to calculate the radius of the dangerous sector area of ​​the obstacle based on the second real-time speed, the preset speed influence coefficient, and the preset safety margin; calculate the comprehensive safe distance between the obstacle and the real robot dog based on the trajectory radius and the radius of the dangerous sector area; and construct the second state data of the obstacle based on the first pose, global position, radius of the dangerous sector area, relative position with the real robot dog, and comprehensive safe distance. The virtual machine look-ahead simulation module is used to calculate the linear velocity of the virtual dog based on the first global distance deviation, the preset target reference velocity, and the heading angle deviation; determine the future pose of the virtual dog at the next time step based on the linear velocity of the virtual dog, the preset global reference path, and the second pose; and construct the third state data of the virtual dog based on the future pose and the first pose. The joint state fusion module is used to generate the current joint state data based on the first state data, the second state data, and the third state data; The control command solving module is used to input the current joint state data into the deep reinforcement learning network based on the prior long short-term memory network, so that the deep reinforcement learning network based on the prior long short-term memory network can generate speed commands for the real robot dog. The whole machine execution drive module is used to control and drive the real robot dog to move according to the speed command of the real robot dog.

9. A terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement a cable inspection robot dog path following and obstacle avoidance navigation method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a cable inspection robot dog path following and obstacle avoidance navigation method as described in any one of claims 1-7.

Images