Generative AI-based expressway body intelligent inspection tour controller construction method

Abstract

The invention discloses a method for constructing an intelligent highway inspection tour controller based on a generative AI (Artificial Intelligence). Basic gait learning is realized through motion redirection and simulation reinforcement learning by utilizing motion track data of an open source robot, a basic level control network based on VQ-VAE is constructed, an expressway environment is disassembled into typical unit scenes such as rugged ground, obstacles, holes and hurdling, and various inspection skills such as crawling, jumping, climbing and obstacle avoiding are trained in each scene. A multi-expert knowledge distillation method is adopted, a plurality of skill-level controllers are fused to form a unified inspection-level controller, and the robot is driven to execute a full-coverage inspection task in a complex road environment in combination with a coverage path generation algorithm. The system has higher maneuverability, adaptability and stability, can realize high-coverage-rate and high-efficiency autonomous inspection in complex scenes such as tunnels, bridges and side slopes, and is suitable for safety monitoring and maintenance of large-scale intelligent traffic infrastructures.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation and robotics, specifically to a method for constructing a highway embodied intelligent inspection controller based on generative AI. Background Technology

[0002] Highway inspection is a crucial link in ensuring the safe operation of transportation infrastructure. Due to the complexity of the environment, inspection tasks not only involve the carriageway and shoulders, but also often include narrow areas such as drainage ditches, rugged terrain such as slopes, and temporary obstacles such as falling rocks, posing significant challenges to stability and full coverage. Traditional manual inspection not only suffers from high subjectivity, limited coverage area, and low efficiency, but also faces significant safety risks during execution.

[0003] Currently, autonomous inspection technologies, including ground vehicles and drones, are gradually being applied to highway inspections. While these technologies have improved inspection performance and reduced labor costs to some extent, their application scenarios remain quite limited. Inspection vehicles rely on ground transportation, and their movement is hindered when facing complex terrain such as unpaved roads, limiting their inspection range. Although drone inspections offer greater mobility, they are easily affected by factors such as endurance and weather conditions like strong winds and rain, making it difficult to guarantee stability during long-duration, long-distance inspection missions. Therefore, existing autonomous inspection technologies alone cannot meet the comprehensive inspection needs of diverse and complex highway environments. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing highway inspection methods, such as insufficient adaptability, poor stability, and limited coverage, and to provide a method for constructing an embodied intelligent inspection controller for highways based on generative AI, so as to achieve efficient inspection of all elements and the entire environment of highways.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] Step 1: Build a basic motion controller:

[0007] The inspection robot is a quadruped robot; the motion trajectories of several self-moving open-source robots are redirected onto the shape of the inspection robot;

[0008] A frozen basic motion controller consisting of a discrete latent space and a decoder is constructed using VQ-VAE, and the gait learning of the inspection robot is completed based on the imitation reinforcement learning method.

[0009] Step 2: Construct a skill-level motion controller:

[0010] The actual highway environment is broken down into several typical inspection unit scenarios, and parameterized representations are constructed based on feature parameters to serve as a learning environment for specific skills.

[0011] Construct several skill-level motion controllers, each of which includes a general environment embedding network and a corresponding specific skill sub-network;

[0012] In each typical unit scenario, a specific skill sub-network is trained based on reinforcement learning methods, and the frozen basic motion controller is reused to complete the learning of specific skills of the inspection robot.

[0013] Step 3: Construct an inspection-level motion controller:

[0014] Construct an inspection path planning generator that does not rely on prior obstacle information, generate an inspection target point sequence based on a coverage planning algorithm, and plan the inspection path.

[0015] A multi-expert distillation method is used to integrate several skill-level motion controllers into a unified inspection-level controller, which works in conjunction with a frozen basic-level motion controller network to achieve full-area inspection of highways.

[0016] The present invention has the following beneficial effects:

[0017] (1) A quadruped robot inspection method adapted to complex highway environments is proposed to replace the traditional ground inspection vehicle and UAV highway inspection methods, which have stronger mobility, stability and efficiency; (2) A multi-level and phased controller training method is constructed to realize the discrete embedding expression and effective reuse of motion knowledge; (3) A method for modular decomposition of highways and parameterized unit scene construction is proposed to realize adaptive training for different environmental elements, which significantly improves the inspection coverage and execution efficiency; (4) In the simulation environment, the robot can achieve 100% coverage and key point hit rate, and the average inspection speed exceeds 2m / s, which is significantly better than the existing manual inspection or autonomous inspection methods, verifying the practicality and superiority of the method. Attached Figure Description

[0018] Figure 1 This is a diagram of the motion learning framework for the three-level highway inspection system of the present invention.

[0019] Figure 2 This is a diagram illustrating a typical unit scenario and its corresponding specific skills.

[0020] Figure 3 This is a schematic diagram of the control strategy structure of a basic motion controller.

[0021] Figure 4 This is a schematic diagram of the control strategy structure for a skill-level motion controller.

[0022] Figure 5 This is a schematic diagram of the distillation training structure of the inspection-level motion controller. Detailed Implementation

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and specific preferred embodiments.

[0024] In the description of this invention, it should be understood that the terms "left side," "right side," "upper part," "lower part," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. "First," "second," etc., do not indicate the importance of the components, and therefore should not be construed as a limitation of this invention. The specific dimensions used in this embodiment are only for illustrating the technical solution and do not limit the scope of protection of this invention.

[0025] like Figures 1-5 As shown, a method for constructing a highway embodied intelligent inspection controller based on generative AI includes the following steps:

[0026] Step 1: Build a basic motion controller:

[0027] The inspection robot is a quadruped robot; the motion trajectories of several self-moving open-source robots are redirected onto the shape of the inspection robot; a frozen basic motion controller consisting of a discrete latent space and a decoder is constructed using VQ-VAE, and the gait learning of the inspection robot is completed based on the imitation reinforcement learning method.

[0028] In this embodiment, multiple open-source robots that have already achieved self-motion are used as reference motion trajectories, which are then mapped onto the inspection robot's form through motion redirection. Each inspection robot is assigned to perform five gaits in the simulation environment: walking, jogging, backward movement, rotation, and jumping, each repeated 4-6 times. The root pose (position and orientation) and the positions (angles) of the 12 joints are captured at 120fps during the motion. Inverse kinematics is used to redirect the expert keypoint sequence to the joint space of the target inspection robot, obtaining a pose sequence for the inspection robot and providing a reference trajectory for subsequent imitation learning. The specific solution formula is as follows.

[0029]

[0030] Specifically, This represents the 3D position of keypoint i on the open-source robot at each time step t. The source keypoints include the positions of the feet and hips, totaling eight, and are paired with the corresponding target keypoints on the inspection robot. The target keypoints of the inspection robot are... The posture of the inspection robot The root orientation and joint positions are determined. Then, inverse kinematics will be applied to obtain the pose sequence of the inspection robot. .

[0031] The basic motion controller is implemented based on vector quantization variational autoencoder (VQ-VAE), which encodes the perceived state information into a discrete latent embedding space, compresses motion data into reusable prior knowledge, and reconstructs the output control signal through a decoder to ensure that the inspection robot can flexibly reproduce the motion behavior of the reference robot.

[0032] A basic motion controller based on the VQ-VAE architecture is constructed to output control actions based on the input state of the inspection robot. The encoder is built based on MLP and accepts... As input, and output latent variables Define a learnable latent embedding space. This is called a codebook, consisting of K D-dimensional vectors. Composition. To achieve vector quantization, i.e., to obtain the discretized latent embedding of high-dimensional continuous features, the continuous latent vector output by the encoder is... With all vectors in the codebook Calculate the distance and select the closest embedding vector. As a representation, that is .

[0033]

[0034] Subsequently, the decoder receives the status. and latent embedding vector As input, output action .

[0035] The VQ-VAE controller enables the discrete representation and reuse of gait knowledge. The encoder receives state input and outputs continuous latent variables; it calculates the distance with K D-dimensional vectors in the codebook and selects the nearest embedded vector as the discrete representation; the decoder generates actions with "state + discrete embedding" as input, realizing the reconstruction of expert actions and the stable reproduction of basic gait.

[0036] The inspection robot utilizes a reinforcement learning approach, learning basic gait through proprioception to achieve stable motion control. The training of the basic motion controller is based on imitative learning of reference motion, using a planar simulation environment as the training scenario. The various elements of reinforcement learning are designed as follows:

[0037] (1) Environment. The simulation environment and the open-source robot data acquisition process are consistent, both being planar.

[0038] (2) State. Primitive level state. It consists of two parts: the physical state of the inspection robot. and the reference target pose within the next k time steps Composition. Among them... It includes continuous pose, velocity, and historical motion over three time steps from t-2 to t. The pose of a single time step includes 12 joint positions and root orientation and root position. Includes reference joint position, reference root position, and orientation for the next 0.03s, 0.06s, 0.3s, and 1s.

[0039] (3) Actions. Actions Defined as the target position residual for each joint. This involves adding the current motion to the current joint position to form the target position, which is then sent to the underlying PD controller, ultimately outputting the specific motor torque command. The control strategy network queries for new motions at 100 Hz, while the PD controller's control frequency is 500 Hz.

[0040] (4) Reward. The reward function is used to track the reference trajectory time by time. Training is terminated when attitude instability occurs, the reference trajectory ends, or the deviation from the reference is too large.

[0041] The rewards in the initial stage will encourage the control strategy to track and imitate the reference motion trajectory at each time step, as defined below.

[0042]

[0043]

[0044] Root position reward Defined as

[0045]

[0046] and These represent the target root position and the inspection robot root position, respectively. This represents the rotational angle error between the robot's current root orientation and the root orientation of the reference trajectory, expressed through the target root orientation. And inspection robot root orientation relative rotation The axis-angle representation is obtained.

[0047] Root speed reward Defined as

[0048]

[0049] and Let represent the root linear velocity of the target and the root linear velocity of the inspection robot, respectively. and These represent the root angular velocity of the target and the root angular velocity of the inspection robot, respectively.

[0050] End effector tracking reward Defined as

[0051]

[0052] and These represent the target position and actual position of the k-th toe relative to the root, respectively.

[0053] Joint position reward Defined as

[0054]

[0055] and These represent the target angle and the actual angle of the j-th joint, respectively.

[0056] Joint speed bonus Defined as

[0057]

[0058] and These represent the target angular velocity and the actual angular velocity of the j-th joint, respectively.

[0059] (5) Termination conditions. The current episode is considered to have ended if the inspection robot meets one of the following conditions.

[0060] ① Posture instability: If the robot's roll angle or pitch angle exceeds the threshold (corresponding to roll > ±45° and pitch > 60° respectively), it is considered to have fallen and training is stopped;

[0061] ② End of reference trajectory: When the imitation trajectory is completed (e.g., when the motion generator-based motion playback ends), the episode will automatically end;

[0062] ③ Excessive deviation between reference trajectory and motion trajectory: When the current reference robot form deviates significantly from the trained inspection robot in posture or position (angle error > 1 rad, position error > 1 m²), the process is forcibly terminated.

[0063] (6) Optimization Algorithm. This study uses the PPO (Proximal Policy Optimization) algorithm to optimize the policy. PPO is an improved version of the policy gradient method, which can ensure the stability of policy updates and improve the convergence speed by introducing mechanisms such as probability ratio clipping.

[0064] Step 2: Construct a skill-level motion controller:

[0065] The actual highway environment is broken down into several typical inspection unit scenarios, and a parameterized representation is constructed based on feature parameters to serve as a learning environment for specific skills. Several skill-level motion controllers are constructed, each of which includes a general environment embedding network and a corresponding specific skill sub-network. In each typical unit scenario, the corresponding specific skill sub-network is trained based on reinforcement learning methods, and the frozen basic motion controller is reused to complete the learning of specific skills for the inspection robot.

[0066] In this embodiment, as Figure 2 As shown, the complex problem environment of highways, such as damaged guardrails, gaps, cracks, rugged ground, and trees and boulders on the road, is decomposed and constructed into typical inspection unit scenarios: holes, hurdles, rugged ground, and obstacles. The terrain undulations, channel scale, obstacle size and distribution, etc. are represented by parameter vectors, which serve as a learning environment for specific skills and realize the abstract expression of environmental features.

[0067] Among them, damaged guardrails are constructed as holes; gaps and cracks are constructed as hurdles; rugged ground is constructed as rugged ground; and trees and boulders on the road are constructed as obstacles.

[0068] Specific skills include crawling, jumping, climbing, and obstacle avoidance. Cave scenarios are used to train crawling skills, hurdle scenarios to train jumping skills, rugged terrain scenarios to train climbing skills, and obstacle scenarios to train obstacle avoidance skills.

[0069] During training, different instances are randomly sampled from the parameter space, introducing environment randomization to improve the generalization and robustness of the policy. Specifically, each scene has a parameter vector containing multiple key feature parameters. During each round of training, the system randomly samples parameters from the parameter space C. This allows for the dynamic generation of scene instances with different combinations of features. This randomization strategy effectively increases the diversity of the training environment, enabling the robot to fully cope with various constantly changing scenarios and master more adaptive and robust inspection skills.

[0070] A skill-level controller is constructed. In addition to the robot's own state, the network also accepts the environmental state and the target point command as inputs. It outputs a latent embedding vector and drives the frozen base-level controller to generate the final motion action.

[0071] The skill-level controller consists of a general environment embedding network and individual skill subnetworks. Environment-aware state information. First, the input is embedded into a network. This module encodes high-dimensional, heterogeneous environmental perception information into a low-dimensional, compact representation. This enhances the ability of subsequent sub-skill modules to analyze environmental features, while preventing the inspection robot from overly relying on environmental information for motion control, ensuring that it can still perform complete and stable inspections based on its own perception even when external perception is interfered with. Subsequently, With the current task instructions and proprioceptive information The inputs are fed into each skill-specific subnetwork, generating a candidate latent embedding distribution p, which represents the probability of each vector in the codebook being selected. Subsequently, specific embedding vectors are obtained through sampling. The data is then fed into the frozen primitive level controller to generate the final joint motion. During the training phase, a random sampling strategy based on probability distribution is typically used to encourage policy diversity and exploration; during the testing or real-world deployment phase, the embedding vector with the highest probability is selected to ensure the stability and safety of the motion.

[0072] The skill controller is trained using reinforcement learning. In each scene unit, a skill-level motion controller is trained independently using reinforcement learning to enable the robot to master specific inspection skills. For example, four scene units will serve as simulation environments to train a separate skill-level inspection motion controller to master specific inspection skills. The training process is also implemented using reinforcement learning, with the specific configuration as follows.

[0073] (1) State. Primitive level state It consists of three parts: the physical state of the inspection robot. This remains consistent with the original stage. Command status. This includes the expected target point's x and y coordinates and the expected velocity. The target point's height is not considered to avoid inducing unnecessary jumps in the robot's control strategy, ensuring that the behavior focuses on ground inspection in the road environment rather than aerial alignment. Environmental State It simulates the perception information of LiDAR and depth cameras. It includes a 25 × 13 map to perceive terrain height within a 2.4 m × 1.2 m square area around the robot; the 25 × 13 map also perceives the depth in front of an area with a width × height × depth of 0.5 m × 0.6 m × 3 m; and 128 uniformly split 360-degree light rays are emitted horizontally from the robot's center to perceive the distance between the robot and its surroundings.

[0074] (2) Action. Consistent with the original level, it represents the target position residuals for each joint.

[0075] (3) Rewards. The rewards for skill level stages will drive the inspection robot to follow commands and reach the target, as defined below.

[0076]

[0077]

[0078] Direction follows reward Defined as

[0079]

[0080] in This is the robot's current root yaw angle. The angle corresponding to the target direction.

[0081] Distance Advancement Rewards Defined as

[0082]

[0083] in The distance between the robot and the target point at the initial moment. and These are the residual distances from the previous step and the current moment, respectively.

[0084] Speed ​​follows reward The robot will only reach the target point if the distance is less than the threshold. Calculated in time. Defined as...

[0085]

[0086] in The average velocity along the target direction within the episode. The desired target speed.

[0087] (4) Termination conditions. The current episode is terminated when one of the following conditions is met.

[0088] ① Attitude instability. Consistent with the assessment in the initial stage.

[0089] ② Step limit. The current step count has reached the preset maximum step count.

[0090] ③ If the two-dimensional Euclidean distance between the robot's position and the target position is less than 0.5 meters, it is considered to have reached the target point, and the task is successfully completed.

[0091] The four unit scenarios described above will serve as simulation environments to train a skill-level inspection motion controller, enabling it to master specific inspection skills. In the "hole" scenario, the robot crawls to navigate through narrow spaces; in the "obstacle" scenario, the robot uses a detour strategy to avoid obstacles; and in the "hurdle" scenario, the robot learns to jump across drainage ditches. After training, the robot possesses multiple typical skills and can maintain stability and adaptability in different scenarios. The state consists of three parts: the robot's own state, the command state, and the environmental state. Actions are still considered as joint target position residuals; rewards emphasize "direction adherence, distance advancement, and speed adherence," and termination conditions such as posture instability, step limit, and reaching a threshold are set.

[0092] Step 3: Construct an inspection-level motion controller:

[0093] An inspection path planning generator that does not rely on prior obstacle information is constructed. Based on the coverage planning algorithm, an inspection target point sequence is generated to plan the inspection path. A multi-expert distillation method is used to integrate skill-level motion controllers into a unified inspection-level controller, which works in conjunction with the frozen basic-level motion controller network to achieve full-area inspection of highways.

[0094] In this embodiment, for a complete and complex highway inspection scenario, an inspection coverage path generator is introduced. Based on a coverage planning algorithm, a sequence of inspection target points is generated, which serves as the input controller for the inspection path instructions. The path planning process employs a "plowing by oxen"-like coverage strategy. This method represents the area to be inspected using a grid map and ensures that the inspection path traverses the entire target area without blind spots by moving back and forth row by row. In each traversal, the robot moves in a fixed direction in the current row. When it encounters an obstacle or boundary at the end of the row, it switches direction and moves to the next row, thus efficiently completing the coverage of the entire area. Unlike traditional coverage path planning, this study does not assume known obstacle information during the path generation stage. This is because in actual inspection tasks, obstacles are often unpredictable and random. Incorporating obstacle information into the planning in advance would cause the model to rely on unrealistic prior conditions, weakening its adaptability in real-world environments. The path generator only provides an idealized sequence of cover target points. During actual robot execution, if an obstacle is encountered, its action strategy will be generated in real-time by the skill network sensing the environmental state, thus maintaining the flexibility and robustness of the task.

[0095] The inspection path generator employs a grid-map-based reciprocating coverage strategy. Specifically, the highway scene to be inspected is modeled as a gridded 2D map M, and a starting point is set as the initial position of the inspection robot. The system defines a direction variable to represent the robot's horizontal movement (e.g., right or left). During execution, the inspection robot moves grid by grid along its current direction, adding its current position to the path queue after each move. When encountering map boundaries or obstacles, the system instructs the robot to move to the next row and automatically reverses its lateral movement direction, thus forming a reciprocating coverage pattern. If the inspection robot continues to move to the next row and exceeds the map boundary, the entire area is considered traversed. Finally, the path generator outputs a list containing a sequence of target points. These points are used sequentially as input instructions for the inspection task, ensuring complete coverage of the area to be inspected.

[0096] By employing a multi-expert distillation method, skill controllers such as crawling, jumping, climbing, and obstacle avoidance are integrated into a unified inspection-level controller. This controller drives the basic-level controller to perform autonomous inspection tasks and detect typical risks in the highway environment, such as slope cracking, boulder accumulation, infrastructure damage, and pavement distress.

[0097] To ensure the effectiveness of the distillation process, the inspection-level controller completely reuses the input format and output interface of the skill-level network in its structural design. Its inputs still include: robot body state, environmental perception state, and task command state. The probability distribution of the embedding vector selection for the output dimension K is also considered. .

[0098] A statically supervised multi-expert distillation process is employed to collect supervised data from episodes successfully completed by experts of each skill level under typical conditions, constructing an expert dataset. For each skill-level expert network... In its corresponding characteristic environment The process executes multiple complete and successful episodes, generating a series of trajectories. Record the input state at each time step t. and the corresponding discrete latent embedding probability distribution All data generated by skill experts are uniformly constructed into an expert dataset.

[0099]

[0100] The training objective is to enable the inspection-level network Given the same input state, the embedding distribution of the output. To approximate the supervised distribution provided by the expert as closely as possible, a unified inspection-level controller is trained using KL divergence as the loss function, with the objective of minimizing KL divergence. This controller outputs an embedding distribution consistent with that of the expert under the same conditions.

[0101]

[0102] The trained inspection-level controller will work in conjunction with the frozen base-level motion controller to continuously generate smooth and stable joint control commands, completing comprehensive inspections and detections of targets such as slope cracking, boulder accumulation, facility damage, and road surface defects. This learning framework possesses good scalability and versatility. If new inspection scenario units are encountered in subsequent applications, only additional skill-level controllers need to be trained and compressed into the unified network through the above process, without requiring reconstruction of the overall control architecture or large-scale retraining.

[0103] This invention utilizes a hierarchical, multi-skill integrated motion learning framework—from basic to skill-inspection levels—to enable quadruped robots to achieve efficient, stable, and intelligent autonomous inspection in complex highway environments. It is applicable to various scenarios such as bridges, tunnels, and slopes, and has promising engineering application prospects and promotional value.

[0104] This invention relates to a method for constructing an embodied intelligent inspection controller based on a quadruped robot design. This method enables the inspection robot to have stronger terrain adaptability, allowing it to flexibly traverse various complex and diverse highway scenarios and achieve the maximum inspection coverage area. It also possesses high stability and accuracy, enabling it to reconstruct environmental information through its own perception even when external sensor information is interfered with, reducing the possibility of self-damage and inspection interruption. Furthermore, it maximizes inspection safety by effectively avoiding personal injury risks and reducing interference with road traffic by flexibly adjusting the inspection path, thus avoiding secondary risks and accidents.

[0105] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details of the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A method for constructing a highway embodied intelligent inspection controller based on generative AI, characterized in that, The method comprises the following steps: Step 1, constructing a basic level motion controller: The inspection robot is a quadruped robot; the motion trajectories of several open-source robots realizing self-motion are redirected to the form of the inspection robot; A frozen basic level motion controller is constructed by using VQ-VAE to construct a discrete latent space and a decoder, and the gait learning of the inspection robot is completed based on an imitation reinforcement learning method; Step 2, constructing a skill level motion controller: The actual environment of the expressway is disassembled into several typical unit scenes for inspection, and a parameterized representation is constructed based on feature parameters as a learning environment of a specific skill; A plurality of skill level motion controllers are constructed, each of which comprises a general environment embedding network and a corresponding specific skill sub-network; In each typical unit scene, the corresponding specific skill sub-network is trained based on a reinforcement learning method, and the frozen basic level motion controller is reused to complete the specific skill learning of the inspection robot; Step 3, constructing an inspection level motion controller: An inspection path planning generator independent of prior obstacle information is constructed, an inspection target point sequence is generated based on a coverage planning algorithm, and the inspection path is planned; A plurality of skill level motion controllers are fused into a unified inspection level controller by using a multi-expert distillation method, and the frozen basic level motion controller network is cooperated to realize the full-domain inspection of the expressway.

2. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 1, the construction method of the basic level motion controller is as follows: The encoder is constructed based on an MLP, accepts as input, and outputs latent variables ; Defining a learnable latent embedding space For the codebook, by K D-dimensional vectors comprise; successive latent vectors output by the encoder all vectors in the codebook compute distances, pick closest embedding vector as a representation, i.e. latent embedding vector . Decoder accepts state And potential embedding vector As input, output action .

3. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 1, the specific configuration of the imitation reinforcement learning method is as follows: A planar simulation environment is used as a training scene; State Body state of the inspection robot and reference target poses for the next k time steps ; wherein contains consecutive poses, velocities and historical actions for the past three time steps t -2 to t; a single time step pose contains 12 joint positions and root orientation, root position; contains reference joint positions, reference root position and orientation for the future 0.03s, 0.06s, 0.3s and 1s; Actions The control policy network defines the target position residual for each joint, queries a new action at 100 Hz, and the underlying PD controller converts the target position to motor torque at 500 Hz. The reward is defined as: where the root location reward is defined as and target root position and inspection robot root position, respectively; represents the rotational angle error between the current root orientation of the robot and the root orientation of the reference trajectory, obtained by the relative rotation of the target root orientation and the inspection robot root orientation in axis-angle representation; Root speed reward defined as and and respectively represent the root linear velocity of the target and the root linear velocity of the inspection robot, and and respectively represent the root angular velocity of the target and the root angular velocity of the inspection robot; End effector tracking reward defined as and tkand tkdenote the target and actual positions of the kth toe with respect to the root, respectively; Joint position reward defined as and θj and θj represent the target angle and the actual angle of the jth joint, respectively; Joint velocity reward defined as and ωj,ωj denote the target and actual angular velocities of the jth joint, respectively; The termination condition is that the posture of the inspection robot is unstable, or the reference trajectory is ended, or the reference trajectory-motion trajectory deviation is significant.

4. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 2, the typical unit scenes include a hole scene, a hurdle scene, a rugged ground scene and an obstacle scene; the specific skills include a crawling skill, a jumping skill, a climbing skill and an obstacle avoidance skill; The hole scene is used to train the crawling skill, and the hurdle scene is used to train the jumping skill; The rugged ground scene is used to train the climbing skill, and the obstacle scene is used to train the obstacle avoidance skill.

5. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 2, the environment parameter randomization technology is introduced in the construction of each typical unit scene, specifically as follows: Each typical unit scenario has a parameter vector containing a number of key characteristic parameters In each round of training, the system randomly samples from the parameter space C Thus dynamically generating typical unit scenario instances with different characteristic combinations.

6. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 2, the construction method of the skill level motion controller is as follows: sensing state information input general-purpose environment embedding network; Encoding high-dimensional, heterogeneous environment perception information into a low-dimensional compact representation , With the current task instructions and the ontology perception information are input into each specific skill sub-network together, and generate candidate potential embedding distribution p, p is specifically represented as the probability of each vector being selected in the codebook; By sampling to get specific embedding vectors and inputting them into the frozen base-level motion controller to finally generate the action of a specific skill.

7. The method of claim 6, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In the training phase, a random sampling strategy based on probability distribution selects embedding vectors ; in the testing or real deployment phase, the embedding vector with the highest probability is selected .

8. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 2, the specific configuration of the reinforcement learning method is as follows: State including the body state of the inspection robot , the command state and the environment state ; Ontological state Consistent with base level motion controller; command state Contains expected target point xy coordinates and expected velocity; environment state Simulated perception information from LIDAR and depth camera Actions For each joint, the target position error is consistent with the base level motion controller; reward r t The computational formula is defined as follows: The direction follows the reward is defined as: wherein is the current root yaw angle of the robot, is the corresponding angle for the target direction; Distance advancement reward is defined as: wherein is the distance of the robot and the target point at the initial time instant, and is the residual distance at the previous time instant and the current time instant, respectively. Speed follows reward is defined as: wherein is the average speed along the target direction within the episode, is the desired target speed; is computed only if the robot successfully reached the target point, i.e. the distance is less than a threshold . The termination condition is that the posture of the inspection robot is unstable, or the number of steps reaches the upper limit, or the target position is reached.

9. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 3, the coverage planning algorithm is as follows: A reciprocating coverage strategy based on a grid map is used to generate a target point sequence: the inspection area is discretized into a two-dimensional grid, and the grid is pushed forward along the current direction from the starting point; when a boundary or an obstacle is encountered, the line is changed and the direction is reversed, realizing reciprocating coverage line by line without dead angle coverage; The path generation stage does not assume prior obstacle information, and dynamic obstacle avoidance is realized by the skill network based on environment perception during the execution period; the generated target point sequence and speed command are input into the controller in sequence to ensure the area coverage rate.

10. The method of claim 1, wherein the method is based on a generative AI-based highway embodied intelligent inspection controller construction method. In step 3, the construction method of the inspection level controller is as follows: The input form and output interface of the skill level motion controller network are reused. The input includes: robot body state, environment perception state, task command state; the output is an embedding vector selection probability distribution of dimension K ; Adopting static supervised multi-expert distillation procedure; for each skill level expert network , perform multiple full and successful episodes in its corresponding characteristic environment , generating a series of trajectories ; record the input state and the corresponding discrete latent embedding probability distribution at each time step t; all skill expert generated data are uniformly constructed into an expert dataset: Enable inspection-level network Under the same input state, the embedding distribution of the output The supervised distribution given by the expert is approximated; KL divergence is used as the loss function, and the goal is to minimize KL divergence to train a unified inspection-level controller so that it outputs an embedding distribution consistent with that of the expert under the same conditions; 。

Images