A robot positioning and navigation control method and system

Abstract

This invention provides a robot positioning, navigation, and control method and system. The method involves: S1 acquiring the robot's perception data and full-body state vector, extracting a minimum interaction feature set and target device identification information from the perception data; S2 constructing a task primitive library, comparing the target device identification information with the task primitive library, and activating task primitives matching the target device identification information; S3 constructing a dynamic environment model based on a reduction strategy, inputting the minimum interaction feature set, full-body state vector, and task primitives into the dynamic environment model for short-time-domain forward simulation to obtain multiple sets of control command sequences; S4 filtering these multiple sets of control command sequences to obtain the optimal control command sequence; and S5 driving the robot to complete the current control cycle, entering the next control cycle, and repeating S1-S5. This robot positioning, navigation, and control method achieves precise and efficient navigation and control in inspection scenarios.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and more specifically, to a robot positioning, navigation, and control method and system. Background Technology

[0002] Power inspection robots often need to operate in complex outdoor terrain. Their core task is to accurately approach and inspect power equipment such as insulators and towers. Existing robot navigation and control methods mostly adopt the classic map-planning-tracking paradigm, which has poor real-time navigation performance, slow response to dynamic environments, and difficulty in balancing the requirements of navigation accuracy and inspection quality.

[0003] Therefore, there is an urgent need for a robot localization and navigation control method that optimizes the entire process of perception, modeling, and control, so as to achieve accurate and efficient navigation control in inspection scenarios. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a robot positioning and navigation control method, the method comprising:

[0005] S1: Acquire the sensory data and whole-body state vector of the inspection robot, and extract the minimum interaction feature set and target device identification information from the sensory data;

[0006] S2: Construct a task primitive library, compare the target device identification information with the task primitive library, and activate the task primitive that matches the target device identification information;

[0007] S3: Based on the reduction strategy, a dynamic environment model is constructed. The minimum interaction feature set, the whole body state vector and the task primitive are input into the dynamic environment model to perform short-time domain forward simulation and obtain multiple sets of control command sequences.

[0008] S4: Based on task primitives, filter multiple sets of control instruction sequences to obtain the optimal control instruction sequence;

[0009] S5: Drive the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle and repeat S1-S5.

[0010] As a further aspect of the present invention, the sensory data and whole-body state vector of the inspection robot are acquired, and the minimum interaction feature set and target device identification information are extracted from the sensory data, including:

[0011] The inspection robot collects perception data from its perception front end, which includes a laser sensor and a multispectral vision sensor. The perception data includes laser echo intensity data and multispectral visual texture data.

[0012] The whole-body state vector of the inspection robot is collected synchronously, including the center of mass pose, joint angles, and actuator torques.

[0013] Feature extraction is performed on the perceived data to obtain the minimum interaction feature set and target device identification information. The minimum interaction feature set includes local terrain height, the first derivative of local terrain height, the second derivative of local terrain height, and the ground attribute estimation vector of the robot contact point set.

[0014] The first and second derivatives of the local terrain height include the slope and curvature of the local terrain, the robot contact point set represents the robot's expected landing point, and the ground data estimation vector includes the subsidence coefficient, shear modulus, and internal friction angle.

[0015] As a further aspect of the present invention, a task primitive library is constructed, and the task primitive library is compared with the target device identification information to activate the task primitive that matches the target device identification information, including:

[0016] Construct a task primitive library, which includes proximity detection primitives and terrain traversal primitives;

[0017] All task primitives in the task primitive library are parameterized policy vectors that include the target device, the desired observation pose, and the stable time window.

[0018] The target device identification information is matched with the task primitive library and the corresponding task primitive is activated.

[0019] As a further aspect of the present invention, a dynamic environment model is constructed based on a reduction strategy. The minimum interaction feature set, the whole-body state vector, and the task primitives are input into the dynamic environment model for short-time forward simulation to obtain multiple sets of control command sequences, including:

[0020] The dynamic environment model includes a data processing layer, a core simulation layer, and an output adaptation layer.

[0021] The data processing layer standardizes the minimum interaction feature set, whole-body state vector, and task primitives of the input dynamic environment model.

[0022] The core simulation layer uses the minimum interaction feature set and the whole-body state vector as initial state parameters, the task primitives as simulation boundaries, and combines dynamic functions to perform simulation and obtain simulation results.

[0023] Specifically, the Ross function is reduced based on navigation control accuracy to obtain the Ross reduced model, and the Ross reduced model is locally linearized to obtain the dynamic function;

[0024] The output adaptation layer receives and parses the simulation results, generating multiple sets of control command sequences.

[0025] As a further aspect of the present invention, multiple sets of control instruction sequences are filtered based on task primitives to obtain the optimal control instruction sequence, including:

[0026] Using the parameterized policy vector in the task primitive as the evaluation standard, multiple sets of control instruction sequences are evaluated and scored, and the control instruction sequence with the highest evaluation score is selected as the optimal control instruction sequence.

[0027] When the task primitive is a proximity detection primitive, the evaluation items include the alignment error between the camera line of sight and the target, the time to reach the target pose, and the overall energy consumption; when the task primitive is a terrain crossing primitive, the evaluation items include the center of mass stability, joint torque margin, and slippage risk prediction.

[0028] As a further aspect of the present invention, the robot is driven to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle, and cycle through S1-S5, including:

[0029] The first control instruction of the optimal control instruction sequence is sent to the robot actuator, which then drives the robot to perform the corresponding action.

[0030] Collect the robot's actual motion state data and feed the actual motion state data back to the dynamic environment model to update the dynamic environment model;

[0031] After completing this control cycle, the next control cycle begins, and S1 to S5 are executed repeatedly until the power inspection task is completed.

[0032] As a further aspect of the present invention, the method further includes:

[0033] If the task objective does not change during the loop, there is no need to activate the task primitive; the task primitive from the previous control cycle will be used instead.

[0034] If the task objective changes, the task primitive corresponding to the changed task objective will be activated.

[0035] Furthermore, embodiments of the present invention also provide a robot positioning and navigation control system, comprising:

[0036] The acquisition module is used to acquire the sensory data and whole-body state vector of the inspection robot;

[0037] The extraction module extracts a minimum set of interactive features and target device identification information from the perceived data;

[0038] The construction module is used to build a task primitive library and construct a dynamic environment model based on a reduction strategy.

[0039] An activation module, which compares the target device identification information with the task primitive library and activates the task primitive that matches the target device identification information;

[0040] The simulation module performs short-time forward simulation of the minimum interaction feature set, the whole-body state vector, and the task primitive input dynamic environment model to obtain multiple sets of control command sequences.

[0041] A filtering module filters multiple sets of control instruction sequences based on task primitives to obtain the optimal control instruction sequence;

[0042] The loop module drives the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle, and repeat S1-S5. Attached Figure Description

[0043] Figure 1 This is a flowchart of the steps of a robot positioning, navigation and control method according to the present invention;

[0044] Figure 2 This is a schematic diagram of a robot positioning and navigation control system according to the present invention. Detailed Implementation

[0045] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0046] like Figure 1 As shown, a robot positioning, navigation, and control method includes the following steps:

[0047] Step S1: Obtain the sensory data and whole-body state vector of the inspection robot, and extract the minimum interaction feature set and target device identification information from the sensory data.

[0048] Step S1 includes:

[0049] Step S1-1: Collect perception data based on the perception front end of the inspection robot. The perception front end includes a laser sensor and a multispectral vision sensor. The perception data includes laser echo intensity data and multispectral visual texture data.

[0050] Specifically, the sensing front end is a component for the inspection robot to acquire information about the external environment. It is equipped with a laser sensor and a multispectral vision sensor. The laser sensor is mainly used to collect the spatial distribution and physical reflection characteristics of the local terrain in the scene and output laser echo intensity data. This data can reflect the density of the ground material, such as the difference in echo intensity between hard soil and soft mud. The multispectral vision sensor is used to collect multi-band visual information of the target area and output multispectral visual texture data. This data can accurately distinguish between power equipment and the background environment, providing texture feature support for target equipment identification. At the same time, the two sensors start and collect data synchronously during the acquisition process to ensure data consistency and adapt to subsequent synchronous processing requirements.

[0051] In some possible embodiments, taking the power inspection robot performing the inspection task of insulator No. 12 as an example, when the robot moves to the vicinity of the area where insulator No. 12 is located, the laser sensor at the sensing front end continuously emits laser signals to the ground and the surrounding area of ​​the insulator, receives the reflected laser signals and generates laser echo intensity data, wherein the echo intensity of the insulator surface is significantly higher than that of the soft mud area, and the echo intensity of the hard soil area is higher than that of the soft mud area; at the same time, the multispectral vision sensor is aimed at the direction of the insulator and collects multispectral visual texture data including the insulator, the ground and the surrounding weeds, wherein the texture of the insulator is smooth and milky white, the texture of the weeds is rough and green, and the texture of the ground is uneven and grayish brown, and the texture of the insulator is significantly different from the texture of the weeds and the texture of the ground.

[0052] Step S1-2: Synchronously collect the whole-body state vector of the inspection robot. The whole-body state vector includes the center of mass pose, joint angles, and actuator torques.

[0053] Specifically, the whole-body state vector is a set of parameters reflecting the current motion state and posture of the inspection robot. The acquisition process is synchronized with the acquisition of perception data to ensure that the two are matched in time. Among them, the center of mass pose is acquired by the robot's onboard IMU sensor, including the three-dimensional spatial coordinates and posture angles of the robot's center of mass, which is used to determine the robot's current balance state. The joint angles are acquired by the joint encoder, mainly targeting the robot's quadruped joints, gimbal joints and other moving joints, reflecting the real-time rotation angle of each joint, which is used to determine the joint motion state. The actuator torque is acquired by the torque sensor, mainly targeting the actuators that drive the track movement and gimbal rotation, reflecting the real-time driving force of the actuator, which is used to determine the robot's current power output state.

[0054] In some possible embodiments, continuing the inspection task of insulator No. 12, as the robot approaches insulator No. 12, it simultaneously collects the whole-body state vector, including: the IMU sensor collects the three-dimensional coordinates of the robot's center of mass in real time as (X=100m, Y=50m, Z=0.8m), with a pitch angle of 3°, a roll angle of 1°, and a yaw angle of 180°, indicating that the robot is currently in a slightly forward-leaning state and the overall posture is stable; the joint encoder collects the rotation angle of the leg joint as 5° and the rotation angle of the gimbal joint as 10°, indicating that the legs are in a slow forward movement state and the gimbal is slowly adjusting the angle to align with the insulator; the torque sensor collects the real-time torque of the track actuator as 20 N·m and the real-time torque of the gimbal actuator as 5 N·m, indicating that the actuator power output is stable and can support the robot's continuous and stable movement and gimbal posture adjustment.

[0055] Steps S1-3 involve extracting features from the perceived data to obtain a minimum interaction feature set and target device identification information. The minimum interaction feature set includes local terrain height, the first derivative of local terrain height, the second derivative of local terrain height, and the ground attribute estimation vector of the robot contact point set.

[0056] The first and second derivatives of the local terrain height include the slope and curvature of the local terrain, the robot contact point set represents the robot's expected landing point, and the ground data estimation vector includes the subsidence coefficient, shear modulus, and internal friction angle.

[0057] Specifically, feature extraction is performed on the collected laser echo intensity data and multispectral visual texture data to select key features directly related to robot navigation and environmental interaction, i.e., the minimum interaction feature set. Simultaneously, specific information about the target power equipment, i.e., target equipment identification information, is identified. Understandably, local terrain height, the first derivative of local terrain height, and the second derivative of local terrain height are extracted from the laser echo intensity data. The local terrain height reflects the terrain undulations along the robot's current path; the first derivative reflects the terrain's inclination, primarily in terms of slope; and the second derivative reflects the terrain's curvature, primarily in terms of curvature. The robot contact point set represents the expected landing point of the robot's legs. Its ground attribute estimation vector is obtained by regressing from the laser echo intensity data and multispectral visual texture data using a lightweight neural network. The subsidence coefficient reflects the degree of ground subsidence after pressure, the shear modulus reflects the ground's shear resistance, and the internal friction angle reflects the frictional characteristics between ground particles. Target equipment identification information is extracted from the multispectral visual texture data. Through algorithms such as texture matching and shape recognition, key information such as the type and number of the target equipment is identified.

[0058] Understandably, using the Z-axis coordinate of the quadruped robot's current center of mass in the world coordinate system as a reference, a local terrain coordinate system is constructed. The X-axis represents the robot's direction of travel, the Y-axis represents the horizontal direction, and the Z-axis represents the vertical direction. The laser echo data collected by the laser sensor within the travel path range is pre-defined as a rectangular area 0.5m-5m in front of the robot and 0.3m to the left and right. Through coordinate transformation and interpolation, this area is mapped to a local terrain height matrix. ,in , These represent the number of sampling points in the local region along the X and Y axes, respectively. Each element in the matrix... Represents coordinates in a local topographic coordinate system The actual terrain height at a given location; specifically, the local terrain height is quantified by a height matrix in the local coordinate system, directly reflecting the terrain undulations at each point on the current robot's path.

[0059] Furthermore, the first derivative of local terrain height is divided into the derivative along the X-axis. and the derivative along the Y-axis All were calculated using the first-order finite difference method, and the specific calculation formula is as follows: , ,in , The sampling interval in the X and Y axes is preset to 0.05m in this embodiment, and the derivative unit is radians. Slope is essentially the rate of change of terrain height with horizontal distance, i.e., the steepness of the terrain inclination. The first derivative of local terrain height is the rate of change of height with the horizontal coordinate. The first derivative in the X-axis direction corresponds to the terrain slope in the robot's travel direction, and the first derivative in the Y-axis direction corresponds to the terrain slope in the robot's lateral direction. The combined derivative of the two... This is the actual slope value of the terrain at that point. The larger the absolute value of the derivative, the steeper the terrain slope. The quadruped robot needs to adjust the angle of its leg joints and the force of landing to maintain balance.

[0060] Furthermore, similarly, the second-order difference method can be used to obtain the second derivative of local terrain height. Curvature reflects the degree of curvature of the terrain outline, such as gentle convexity, concavity, and right-angle bends, while the second derivative is the rate of change of the height change rate. When the second derivative is positive, it indicates that the rate of change of terrain height is increasing, corresponding to a convex curve, such as the top of a small mound; when the second derivative is negative, it indicates that the rate of change of terrain height is decreasing, corresponding to a concave curve, such as the bottom of a small pit; when the second derivative approaches 0, it indicates that the rate of change of terrain height remains stable, corresponding to a gentle slope.

[0061] Furthermore, the ground attribute estimation vector of the robot's contact point set, i.e. the expected landing points of the four legs of the quadruped robot, combined with the robot's gait planning preset, is usually obtained by regressing the 4-6 feasible points closest to the current centroid in the local terrain through a lightweight neural network, such as using a CNN-LSTM lightweight architecture. The input is the fusion feature of laser echo intensity data and multispectral visual texture data, and the output is a 3D vector. Among them, the subsidence coefficient reflects the degree of subsidence of the ground after pressure. The larger the coefficient, the softer the ground, and the more obvious the subsidence when the quadruped robot lands. The shear modulus reflects the shear resistance of the ground. The larger the modulus, the stronger the ground's resistance to deformation, and the greater the leg torque that can be supported. The internal friction angle reflects the friction characteristics between ground particles. The larger the angle, the greater the ground friction, and the lower the risk of the quadruped robot's legs slipping. The target equipment identification information is extracted from the multispectral visual texture data. Through texture matching, shape recognition and other algorithms, the type of target equipment, such as suspension insulators, pin insulators, towers, number and other key information, is accurately identified.

[0062] In some possible embodiments, the inspection task of insulator No. 12 continues, and feature extraction is performed on the collected perception data. This includes extracting the local terrain height of the robot's current path from the laser echo intensity data. It is found that there is a slight protrusion 5m ahead, with a height difference of 0.3m. The first derivative slope of the local terrain height is 5°, and the second derivative curvature is 0.02° / m, indicating that the terrain of the protrusion is gentle and will not affect the robot's passage. The expected landing points of the robot's legs are determined to be 2m, 3m, 4m, and 5m ahead. The ground attribute estimation vectors of these four contact points are obtained by lightweight neural network regression. The settlement coefficient is 0.1, the shear modulus is 1000Pa, and the internal friction angle is 30°, indicating that the ground in this area is hard soil with strong bearing capacity and is not prone to settlement and slippage. The target equipment identification information is extracted from the multispectral visual texture data. Through texture matching and shape recognition, the target equipment is clearly identified as insulator No. 12, which is a suspension insulator.

[0063] Step S2: Construct a task primitive library, compare the target device identification information with the task primitive library, and activate the task primitive that matches the target device identification information.

[0064] Step S2 includes:

[0065] Step S2-1: Construct a task primitive library, which includes proximity detection primitives and terrain traversal primitives.

[0066] In this context, all task primitives in the task primitive library are parameterized policy vectors that include the target device, the desired observation pose, and the stable time window.

[0067] Specifically, based on typical power line inspection scenarios, two types of core task primitives were selected: proximity detection primitives and terrain traversal primitives, covering the two major requirements of traversing complex terrain and proximity detection during the inspection process. All task primitives are constructed in the form of parameterized policy vectors. The parameters of each task primitive include three parts: target device, expected observation pose, and stable time window. The target device is used to clarify the specific power equipment targeted by the task, the expected observation pose is used to clarify the optimal position and posture of the robot when completing the task, and the stable time window is used to clarify the time that the robot needs to maintain stability after reaching the expected observation pose.

[0068] It should be noted that in power line inspection scenarios, target equipment types are diverse and terrain varies greatly. Furthermore, the movement posture and leg landing planning of the quadruped robot need to be precisely matched with the real-time terrain and target location. If a fixed instruction sequence is used, it cannot adapt to the detection requirements of different equipment and the access requirements of different terrains. It is necessary to design task primitives separately for each scenario, resulting in redundancy of the task primitive library, high maintenance costs, and an inability to quickly respond to scenario changes. Therefore, this embodiment adopts a parameterized strategy vector. By simply adjusting the core parameters, the same type of task primitive can be adapted to different scenarios without redesigning the task logic. This reduces the construction and maintenance costs of the task primitive library and improves the adaptability and response speed of the task primitives.

[0069] In some possible embodiments, continuing the inspection task of insulator No. 12, the constructed task primitive library includes a proximity detection primitive designed for devices requiring close-range inspection, such as insulators and towers. Its parameterized strategy vector is proximity detection, such as (target device: insulator No. 12, desired observation pose: 5m distance from insulator, azimuth angle 30°, pitch angle 10°, stabilization time window: 2s). This clarifies that the robot needs to approach insulator No. 12 to a distance of 5m, adjust its posture to align the camera with the insulator, and maintain stability for 2s to complete the inspection. The terrain crossing primitive is designed for complex terrains such as slopes and mud in the inspection path. Its parameterized strategy vector is terrain crossing, such as (target device: no specific device, desired observation pose: keep the center of mass horizontal, the heading angle unchanged, stabilization time window: 1s). This clarifies that the robot needs to maintain a stable posture when crossing the terrain to ensure smooth passage and provide path support for approaching insulator No. 12.

[0070] Step S2-2: Match the target device identification information with the task primitive library and activate the corresponding task primitive.

[0071] Specifically, the matching of target device identification information with the task primitive library is based on the type and number of the target device being inspected. The most suitable task primitive is selected from the task primitive library and activated to ensure that the robot's subsequent actions can conform to the current task objective. The matching process adopts the principle of precise target device matching, extracting key parameters such as device type and number from the target device identification information, and then comparing them with the target device parameters of each task primitive in the task primitive library to select task primitives whose parameters are completely matched or highly adapted. After activation, the parameterized strategy vector of the task primitive will serve as the core basis for subsequent dynamic environment model simulation and control command selection, guiding the robot to complete the corresponding task.

[0072] In some possible embodiments, the inspection task of insulator No. 12 continues, and the target equipment identification information extracted in steps S1-3 is matched with the task primitive library. By comparing the task primitives in the task primitive library, it is found that the target equipment parameter of a certain proximity detection primitive is insulator No. 12, which is a perfect match with the current target equipment identification information; the target equipment parameter of the terrain crossing primitive is no specific equipment, which does not match the current target equipment identification information. Therefore, it is determined to activate the proximity detection primitive. After activation, the parameterized strategy vector of the proximity detection primitive (target equipment: insulator No. 12, expected observation pose: distance from insulator 5m, azimuth angle 30°, pitch angle 10°, stabilization time window: 2s) will be used as the guiding parameter for subsequent steps.

[0073] Step S3: Construct a dynamic environment model based on the reduction strategy. Input the minimum interaction feature set, the whole body state vector and the task primitive into the dynamic environment model to perform short-time domain forward simulation and obtain multiple sets of control command sequences.

[0074] Step S3 includes:

[0075] Step S3-1: The dynamic environment model includes a data processing layer, a core simulation layer, and an output adaptation layer.

[0076] It should be noted that the dynamic environment model is a module for the robot to predict the interaction results between itself and the environment and generate control command sequences. It adopts a three-layer hierarchical framework design including a data processing layer, a core simulation layer, and an output adaptation layer. The data processing layer is mainly responsible for the preprocessing of input data to ensure that the input data is standardized and consistent in timing, providing input data for the core simulation layer. The core simulation layer is responsible for the dynamic coupling simulation of the robot and the environment, predicting the robot's motion state under different control commands. The output adaptation layer is responsible for parsing the simulation results and mapping the motion state obtained from the simulation into a feasible control command sequence, providing a basis for subsequent control command selection.

[0077] Furthermore, the dynamic environment model separates data preprocessing, core simulation, and instruction mapping into independent layers to avoid mutual interference between the various stages and reduce the overall debugging difficulty of the model. If a certain stage needs to be optimized later, such as improving the accuracy of data standardization or optimizing the simulation algorithm, the corresponding layer can be adjusted separately without modifying the entire model, thus improving the model's scalability. At the same time, it adapts to the dynamic scene's adaptive capability. The three-layer collaborative working mode can quickly respond to changes in input data, such as sudden changes in local terrain, fine adjustments to the target device's position, and fluctuations in the quadruped robot's own posture. The data processing layer quickly completes the standardization of new data, the core simulation layer updates the simulation boundary and initial parameters in real time, and the output adaptation layer generates new control instruction sequences in a timely manner, ensuring that the model can adjust its output according to the dynamic changes in the inspection scene, thereby improving the adaptability of the quadruped robot's navigation control.

[0078] In step S3-2, the data processing layer standardizes the minimum interaction feature set, whole-body state vector, and task primitives of the input dynamic environment model.

[0079] Specifically, the data processing layer standardizes the minimum interaction feature set, whole-body state vector, and task primitives of the input dynamic environment model, thereby solving problems such as inconsistent dimensions, temporal deviations, and chaotic parameter ranges in the input data, ensuring that the data can adapt to the simulation requirements of the core simulation layer. The standardization process includes data temporal alignment, parameter range standardization, and task primitive decomposition. Specifically, data temporal alignment means synchronizing the acquisition time of the three types of data to ensure that they all correspond to the initial state of the same control cycle. Parameter range standardization means standardizing the terrain parameters of the minimum interaction feature set, the attitude parameters of the whole-body state vector, and the observation pose parameters of the task primitives to a preset range, such as 0-1, to eliminate dimensional differences. Task primitive decomposition means decomposing the parameterized strategy vector of the task primitives into quantifiable simulation constraints, such as decomposing the desired observation pose into three-dimensional coordinate constraints and attitude angle constraints, providing a clear simulation boundary for the core simulation layer.

[0080] In some possible embodiments, the inspection task of insulator No. 12 continues, and the data processing layer performs standardization processing on the input data, including synchronously calibrating the acquisition time of the minimum interaction feature set (ground subsidence coefficient 0.1, slope 5°), the whole body state vector (pitch angle 3°, leg torque 20 N·m), and the proximity detection primitive (distance 5m, azimuth angle 30°) to ensure that they all correspond to the initial state of the current control cycle (2ms); the subsidence coefficient 0.1 is standardized to 0 by a normalization method such as min-max. 1. Parameters such as slope of 5° is standardized to 0.25, pitch angle of 3° is standardized to 0.15, leg torque of 20 N·m is standardized to 0.8, and distance of 5m is standardized to 0.5, and are uniformly standardized to the preset range of 0-1; the proximity detection primitive is decomposed into quantitative constraints, including distance from the insulator of 5m±0.1m, azimuth angle of 30°±1°, pitch angle of 10°±1°, and settling time ≥2s, to provide clear simulation boundaries for the core simulation layer and ensure that the simulation process revolves around the target of approaching insulator No. 12.

[0081] In step S3-3, the core simulation layer uses the minimum interaction feature set and the whole-body state vector as initial state parameters, the task primitive as the simulation boundary, and combines the dynamic function to perform simulation and obtain simulation results.

[0082] Specifically, the Ross function is reduced based on navigation control accuracy to obtain the Ross-reduced model, and the Ross-reduced model is locally linearized to obtain the dynamic function.

[0083] Specifically, the simulation process of the core simulation layer uses the minimum interaction feature set and the whole-body state vector as initial state parameters to clarify the robot's current self-state and environmental state; it uses the quantized constraints after task primitive decomposition as simulation boundaries to ensure that the simulation results fit the current task objectives; and it uses the dynamic function constructed based on the Ross function as the simulation core to achieve short-time domain high-speed simulation of robot-environment dynamic coupling. The construction of the dynamic function includes two steps: First, based on navigation control accuracy requirements, the classic Ross function is reduced by removing redundant degrees of freedom and higher-order coupling terms, retaining the core parameters related to robot navigation and environmental interaction, thus constructing a reduced Ross model and reducing computational complexity; second, the reduced Ross model is locally linearized, approximating the nonlinear coupling terms as linear relationships near the current working point, further reducing computational complexity, and finally obtaining the dynamic function. The simulation results mainly include the predicted short-time domain motion states of the robot under different control commands, including changes in the center of mass pose, joint angles, and actuator torque.

[0084] Understandably, the reduction of the classical Ross function includes clarifying the reduction premise and core objective, decomposing and filtering the Ross function degrees of freedom, identifying and eliminating higher-order coupling terms, and parameter calibration and model verification. The reduction premise, in clarifying the reduction premise and core objective, is to meet the navigation control accuracy requirements of the quadrupedal bionic inspection robot, i.e., navigation positioning error less than or equal to 0.01m and attitude control error less than or equal to 0.1°. The core objective is to eliminate redundant degrees of freedom and higher-order coupling terms unrelated to navigation control, reduce the computational complexity of the Ross function, and ensure that the reduced model accurately reflects the coupling relationship between the robot's leg movements, center of mass balance, and the environment. The decomposition and filtering of the Ross function degrees of freedom... The selection process involves decomposing the robot's degrees of freedom (DOFs) contained in the classic Ross function. For example, a quadruped robot has 12 joint DFs, 3 translational DFs, and 3 rotational DFs, totaling 18 DFs. DFs directly related to navigation and control are selected, redundant DFs are eliminated, and 4 core leg joint DFs, 3 translational DFs, and 3 rotational DFs are retained, totaling 10 core DFs. Eight redundant DFs are eliminated, such as the two auxiliary joint DFs per leg, which are only responsible for minor posture adjustments and have minimal impact on overall navigation and control, and can be indirectly fitted through the core DF motion. The identification and reduction of higher-order coupling terms specifically refers to the identification and reduction of the core DFs retained after selection. The Ross function terms were analyzed to identify and remove higher-order coupling terms, i.e., coupling terms with an order greater than or equal to 3. Only first- and second-order core coupling terms were retained. This is mainly because higher-order coupling terms primarily include third-order coupling terms between different leg joint degrees of freedom and third-order coupling terms between the center of mass attitude degree of freedom and the leg joint degrees of freedom. These coupling terms have high computational complexity and contribute little to improving navigation control accuracy, thus requiring removal. The retained first- and second-order coupling terms mainly include first-order coupling terms between the leg joint degrees of freedom and the center of mass translational degree of freedom, and second-order coupling terms between the center of mass attitude degree of freedom and the ground reaction force. These coupling terms directly affect the robot's walking stability and navigation accuracy. The specific table for core parameter calibration and model verification is as follows: The Ross function, after removing redundant degrees of freedom and higher-order coupling terms, is used as the initial reduced model. Combined with the quadruped robot's hardware parameters, including the leg actuator's limit torque, joint rotation range, and center of mass coordinate range, the core parameters in the model are calibrated, such as inertial parameters and damping coefficients, to ensure that the model parameters are consistent with the robot's actual hardware characteristics. Then, preliminary simulation verification is performed, comparing the simulation results of the reduced model with those of the classic Ross function, including changes in center of mass pose and leg joint angles. If the error is less than or equal to ±0.005m and ±0.05°, the construction of the reduced Ross model is complete. If the error exceeds the threshold, the process returns to fine-tuning the degree of freedom selection range or the proportion of higher-order coupling term removal until the verification criteria are met.

[0085] Furthermore, the Ross reduction model undergoes local linearization, approximating the nonlinear coupling terms as linear near the current working point, further reducing computational complexity and ensuring high-speed simulation while maintaining simulation accuracy. The specific linearization steps include determining the local linearization working point. Using the typical working state of the quadrupedal bionic inspection robot as a benchmark, the working point is determined. The selection of the working point must closely match the actual power inspection scenario, ensuring coverage of the robot's main motion postures. Specifically, the working point is selected as the center of mass posture angle, the rotation angle of the core leg joints, and the actuator torque. This working point range covers the robot's motion states when approaching equipment and traversing flat terrain, and linearization can adapt to most inspection scenarios. Simultaneously, the robot's actual state in the current control cycle is used as an auxiliary working point, combined with the whole-body state vector, to ensure the linearization result closely matches the real-time motion state. The process involves identifying and decomposing nonlinear terms. Specifically, this means decomposing the Ross reduction model and identifying all nonlinear coupling terms, primarily the retained second-order coupling terms, such as the second-order coupling terms between the center of mass attitude degree of freedom and the ground reaction force, and the second-order coupling terms between the leg joint degrees of freedom and the center of mass translational degree of freedom. The Ross reduction model is then decomposed into linear and nonlinear terms. The linear terms are directly retained as the basic framework of the locally linearized model; the nonlinear terms are extracted separately and subsequently linearized for approximation. Then, a Taylor expansion and approximation of the nonlinear terms are performed. Specifically, a first-order Taylor expansion method is used to expand the extracted nonlinear coupling terms near a preset locally linearized working point, ignoring second-order and higher-order infinitesimal terms in the Taylor expansion, retaining only the first-order terms to achieve linearization approximation of the nonlinear terms. The pitch angle θ of the center of mass attitude degree of freedom and the ground reaction force are used as the basis for the linearization approximation. Second-order coupling term For example, where k represents the proportionality coefficient at the operating point Expanding nearby, such as The Taylor expansion is After ignoring higher-order terms, the nonlinear term is approximated as a linear function of θ, thus reducing computational complexity. All nonlinear coupling terms are linearized and approximated using the above method. Finally, the linearized model is integrated and verified, specifically by integrating the retained linear terms with all linearized approximated nonlinear terms to construct a complete locally linearized dynamic model. Subsequently, linearization verification is performed, comparing the simulation results of the linearized model with the Ross reduction model, mainly including changes in the center of mass pose, leg joint angle, and actuator torque. The verification error must meet the navigation control accuracy requirements. If the error exceeds the threshold, the process returns to the step of adjusting the operating point range or the step of optimizing the approximate accuracy of the Taylor expansion until the verification criteria are met. If the error is within the allowable range, the local linearization process is complete.

[0086] It should be noted that the scaling factor k is an inherent parameter of the quadrupedal bionic inspection robot, determined jointly by the robot's hardware characteristics and the calibration parameters of the Ross reduction model. The robot's hardware characteristics include the stiffness of the leg actuators, joint damping coefficients, and robot center-of-mass inertial parameters, such as mass and moment of inertia. When constructing the Ross reduction model, the scaling factor is initially calibrated by fitting the simulation data of the classical Ross function and the reduced model, such as the center-of-mass pose, leg joint angles, and measured and predicted values ​​of ground reaction forces, in conjunction with the robot's hardware characteristics, to ensure that the scaling factor matches the robot's actual motion characteristics.

[0087] In some possible embodiments, the inspection task of insulator No. 12 continues. The standardized minimum interaction feature set, such as hard ground terrain, slope of 5°, and whole-body state vector, such as centroid coordinates and pitch angle of 3°, are used as initial state parameters to clarify that the robot is currently in hard ground terrain, slightly tilted forward, and in a stable walking state. The constraint conditions after the proximity detection primitive is decomposed are used as simulation boundaries, such as distance of 5m and azimuth angle of 30°, to ensure that the robot's movement is always towards insulator No. 12 during the simulation. The dynamic function adopts a function constructed based on Ross function reduction and local linearization to quickly simulate the robot's motion state under different leg torque and gimbal rotation angle commands. Multiple sets of motion states are predicted in the short time domain. For example, when the leg torque is 22 N·m and the gimbal rotation angle is 12°, the robot's centroid will move 0.1m towards the insulator, the pitch angle will remain at 3°, and the posture will be stable. When the leg torque is 25 N·m and the gimbal rotation angle is 15°, the robot's movement speed increases, but the posture fluctuation increases, the time to approach the insulator is shortened, but the vibration intensifies. Finally, the motion state prediction results corresponding to multiple sets of different control commands are obtained.

[0088] Step S3-4: The output adaptation layer receives and parses the simulation results to generate multiple sets of control command sequences.

[0089] Specifically, the output adaptation layer receives multiple sets of motion state prediction results from the core simulation layer, and parses and maps the simulation results, converting the motion state prediction values ​​into a sequence of control commands that the robot actuator can directly execute. The parsing process mainly includes: screening the simulation results to remove motion state prediction results that do not meet the task primitive constraints, such as not facing the target device, or do not meet the hardware safety constraints, such as the actuator torque exceeding the limit; then, command mapping is performed, and the filtered valid motion state prediction results are back-mapped into the corresponding control command sequences. Each set of control command sequences corresponds to continuous control commands in the short time domain, such as track torque and joint rotation angle commands for each control cycle, and finally, multiple sets of feasible control command sequences are generated.

[0090] In some possible embodiments, the inspection task of insulator No. 12 continues, and the simulation results of the adaptation layer screening are output. Predictions with excessive robot posture fluctuations or actuator torque exceeding limits are discarded, while valid results that meet the proximity detection primitive constraints and have stable posture are retained. Then, the valid simulation results are back-mapped into a sequence of control commands, generating a total of 50 feasible sequences. Each sequence corresponds to 250 control commands within 0.5 seconds, and the control period is set to 2ms. Therefore, there are 250 control commands within 0.5 seconds. For example, one of them... The control command sequence is as follows: control cycles 1-50 (0-0.1 seconds): leg torque 22 N·m, gimbal rotation angle 11°; control cycles 51-150 (0.1-0.3 seconds): leg torque 23 N·m, gimbal rotation angle 12°; control cycles 151-250 (0.3-0.5 seconds): leg torque 22 N·m, gimbal rotation angle 13°. The simulation result corresponding to this sequence is that the robot smoothly approaches the No. 12 insulator, with stable posture, and closely approaches the target of the detection task.

[0091] Step S4: Based on the task primitives, filter multiple sets of control instruction sequences to obtain the optimal control instruction sequence.

[0092] Step S4 includes:

[0093] Using the parameterized policy vector in the task primitive as the evaluation standard, multiple sets of control instruction sequences are evaluated and scored, and the control instruction sequence with the highest evaluation score is selected as the optimal control instruction sequence.

[0094] It should be noted that the evaluation process adopts a quantitative scoring and comprehensive ranking method. Based on the type of activated task primitive, the corresponding evaluation items and the weight of each evaluation item are determined. Each evaluation item of each control command sequence is quantitatively scored with a maximum score of 100 points. The comprehensive score is calculated based on the weights. All control command sequences are sorted in descending order of comprehensive scores, and the control command sequence with the highest comprehensive score is selected as the optimal control command sequence. If multiple sequences have the same comprehensive score, the sequence with lower energy consumption in the proximity detection primitive or lower risk of slippage in the terrain crossing primitive is selected first to ensure the practicality and reliability of the optimal sequence.

[0095] Step S4-1: When the task primitive is the proximity detection primitive, the evaluation items include the alignment error between the camera line of sight and the target, the time to reach the target pose, and the overall energy consumption.

[0096] Specifically, when the activated task primitive is the proximity detection primitive, the core of the selection and evaluation of the control command sequence is to align with the proximity detection task objective, that is, to ensure that the robot can approach the target device quickly, smoothly, and accurately, while reducing energy consumption and providing good posture support for subsequent detection work. Therefore, three evaluation items are set, each with clear quantitative standards. The weights are set according to task priority, with alignment error having the highest weight to ensure detection accuracy; arrival time and energy consumption have the next highest weights to balance efficiency and energy consumption. Among them, the alignment error between the camera line of sight and the target evaluates the deviation between the camera line of sight and the center of the target device after the robot reaches the desired observation pose. The smaller the deviation, the higher the detection accuracy. The time to reach the target pose evaluates the time it takes for the robot to reach the desired observation pose from the current state. The shorter the time, the higher the inspection efficiency. The overall energy consumption evaluates the total energy consumption of the robot during the execution of the control command sequence. The lower the energy consumption, the stronger the robot's endurance.

[0097] In some possible embodiments, the inspection task of insulator No. 12 continues. Currently, the proximity detection element is activated, and 50 sets of control command sequences are evaluated. Quantitative standards are set as follows: alignment error less than or equal to 0.1 mm has a weight of 0.5, arrival time less than or equal to 10 s has a weight of 0.3, and overall energy consumption less than or equal to 200 J has a weight of 0.2. For example, the 10th control command sequence has an alignment error of 0.08 mm, an arrival time of 8 s, and an overall energy consumption of 180 J, with a comprehensive score of 0.92; the 25th control command sequence has an alignment error of 0.15 mm, an arrival time of 7 s, and an overall energy consumption of 170 J, with a comprehensive score of 0.75; and the 38th control command sequence has an alignment error of 0.09 mm, an arrival time of 11 s, and an overall energy consumption of 160 J, with a comprehensive score of 0.78. Through quantitative evaluation, candidate sequences with higher comprehensive scores are initially screened.

[0098] Step S4-2: When the task primitive is the terrain crossing primitive, the evaluation items include center of mass stability, joint torque margin, and slippage risk prediction.

[0099] Specifically, when the activated task primitive is the terrain traversal primitive, the selection and evaluation of the control command sequence is aligned with the terrain traversal task objective, namely, ensuring that the robot can smoothly and safely traverse complex terrain, avoiding risks such as rollover and slippage, while protecting the actuators and joints and extending the equipment's lifespan. Therefore, three evaluation items are set, each with clear quantitative standards, and their weights are set according to safety priorities. Among them, center of mass stability and slippage risk have the highest weights to ensure safety; joint torque margin has the second highest weight to protect the equipment. Center of mass stability evaluates the fluctuation range of the robot's center of mass posture during the execution of the sequence; the smaller the fluctuation, the better the stability. Joint torque margin evaluates the difference between the actuator torque and the joint's limit torque; the larger the difference, the higher the safety redundancy of the joint and actuator. Slippage risk prediction evaluates the probability of robot track slippage based on ground attribute estimation vectors and simulation results; the lower the probability, the higher the safety.

[0100] In some possible embodiments, continuing the inspection task of insulator No. 12, suppose that the robot encounters a slope with an 8° gradient while approaching insulator No. 12, activates the terrain traversal primitive, and evaluates the newly generated 50 sets of control command sequences. Suppose that the weights for the centroid pitch angle fluctuation of less than 2° are set to 0.4, the weights for the joint torque margin of greater than or equal to 5 N·m are set to 0.2, and the weights for the slippage risk of less than or equal to 5% are set to 0.4. For example, the evaluation result for the 8th set of control command sequences is a centroid pitch angle fluctuation of 1. The first set of control command sequences had the following evaluation results: 5° pitch angle fluctuation, 6 N·m joint torque margin, 3% slippage risk, and an overall score of 0.91. The second set of control command sequences had the following evaluation results: 2.5° pitch angle fluctuation, 7 N·m joint torque margin, 2% slippage risk, and an overall score of 0.76. The third set of control command sequences had the following evaluation results: 1.8° pitch angle fluctuation, 4 N·m joint torque margin, 4% slippage risk, and an overall score of 0.79. The sequences with higher overall scores were selected to ensure that the robot could smoothly cross the slope and continue to approach the No. 12 insulator.

[0101] Step S5: Drive the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle, and repeat S1-S5.

[0102] Step S5 includes:

[0103] Step S5-1: Send the first control instruction of the optimal control instruction sequence to the robot actuator, and the robot actuator drives the robot to perform the corresponding action.

[0104] Specifically, only the first control instruction of the optimal control instruction sequence is sent to the actuator, rather than the complete sequence, to ensure that the robot can cope with real-time changes in the dynamic environment. The optimal control instruction sequence corresponds to continuous control instructions in a short time domain. The first control instruction corresponds to the action instruction of the current control cycle. After being sent to the robot's core actuators such as the track actuator and the gimbal actuator, the actuator outputs the corresponding power and action according to the instruction, driving the robot to complete the movement of the current control cycle, such as the track moving forward and the gimbal adjusting the angle, to ensure that the robot's current action conforms to the optimal prediction, while providing actual motion data for the state update of the next control cycle.

[0105] In some possible embodiments, continuing the inspection task of insulator No. 12, the current optimal control command sequence is the 10th sequence, which contains 250 control commands, corresponding to 0.5 seconds. Only the first control command, a leg torque of 22 N·m and a gimbal rotation angle of 11°, is sent to the robot actuator. After receiving the command, the leg actuator outputs a torque of 22 N·m to drive the leg to move forward slowly. After receiving the command, the gimbal actuator adjusts the rotation angle to 11° so that the camera is further aligned with insulator No. 12, completing the action of the current 2ms control cycle.

[0106] Step S5-2: Collect the robot's actual motion state data and feed the actual motion state data back to the dynamic environment model to update the dynamic environment model.

[0107] Specifically, the robot collects actual motion state data after executing the first control command through IMU sensors, joint encoders, torque sensors, etc. The parameters of the actual motion state data are consistent with those of the whole body state vector, including the center of mass pose, joint angles, and actuator torques. The actual motion state data is compared with the motion state data predicted by the core simulation layer, and the deviation values ​​are calculated, such as center of mass pose deviation and joint angle deviation. The parameters of the dynamic environment model are updated online based on the deviation values, including the ground attribute estimation vector and the linearization coefficient of the dynamic function.

[0108] In some possible embodiments, after the robot executes the first control command, it collects actual motion state data, including the IMU sensor collecting the actual distance the center of mass moves (0.0075m, predicted value 0.008m), the pitch angle (actual value 2.9°, predicted value 3°), the joint encoder collecting the actual gimbal rotation angle (10.9°, predicted value 11°), and the torque sensor collecting the actual leg torque (21.8 N·m, predicted value 22 N·m). The actual data is compared with the predicted data, and the deviation values ​​are calculated to be within the allowable range. Based on the deviation value, the dynamic environment model updates the ground attribute estimation vector online to a settlement coefficient of 0.098, where the original settlement coefficient was 0.1. The linearization coefficient of the dynamic function is fine-tuned so that subsequent simulations can more accurately predict the motion state of the robot as it approaches the No. 12 insulator.

[0109] Step S5-3: After completing this control cycle, proceed to the next control cycle and repeat steps S1 to S5 until the power inspection task is completed.

[0110] During the loop, if the task objective does not change, there is no need to activate the task primitive, and the task primitive of the previous control cycle is used; if the task objective changes, the task primitive corresponding to the new task objective is activated.

[0111] Specifically, during the loop, if the task target does not change, such as always approaching and detecting insulator No. 12, then there is no need to repeat the task element matching and activation in step S2. The task element activated in the previous control cycle can be directly used to improve loop efficiency. If the task target changes, such as switching to detecting insulator No. 13 after completing the detection of insulator No. 12, then step S2 is executed again. The corresponding task element is activated through the new target device identification information to ensure that the loop logic fits the new task target. The loop continues until the robot completes all the preset power inspection tasks and the loop terminates.

[0112] In some possible embodiments, the inspection task of insulator No. 12 continues. After completing the current 2ms control cycle, the next control cycle begins, and the S1-S5 process is repeated. In step S1, the sensing data and the whole-body state vector are re-acquired, and the extracted target device identification information is still insulator No. 12, that is, the task target has not changed. In step S2, there is no need to re-match and activate the task primitive, and the proximity detection primitive activated in the previous cycle is directly used. In steps S3-S4, based on the updated dynamic environment model, the optimal control command sequence is regenerated and filtered. In step S5, the first control command of the new optimal sequence is executed, the actual state is fed back and the model is updated, and this cycle is repeated. This cycle continues until the robot reaches the expected observation pose of insulator No. 12, maintains stability for 2s to complete the detection, and then the task target is switched to detect insulator No. 13. At this time, the proximity detection primitive is reactivated, and the target device is insulator No. 13. The cycle continues until all insulator inspection tasks are completed.

[0113] Figure 2 The diagram illustrates a robot positioning and navigation control system that can implement the ideas of this application, according to some embodiments of this application.

[0114] Specifically, a robot positioning and navigation control system includes:

[0115] The acquisition module is used to acquire the sensory data and whole-body state vector of the inspection robot;

[0116] The extraction module extracts a minimum set of interactive features and target device identification information from the perceived data;

[0117] The construction module is used to build a task primitive library and construct a dynamic environment model based on a reduction strategy.

[0118] An activation module, which compares the target device identification information with the task primitive library and activates the task primitive that matches the target device identification information;

[0119] The simulation module performs short-time forward simulation of the minimum interaction feature set, the whole-body state vector, and the task primitive input dynamic environment model to obtain multiple sets of control command sequences.

[0120] A filtering module filters multiple sets of control instruction sequences based on task primitives to obtain the optimal control instruction sequence;

[0121] The loop module drives the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle, and repeat S1-S5.

[0122] The specific usage and function of this embodiment are explained below:

[0123] First, the robot's perception data and full-body state vector are acquired via S1. The minimum interaction feature set and target device identification information are extracted from the perception data. Next, a task primitive library is constructed based on S2. The target device identification information is compared with the task primitive library, and task primitives matching the target device identification information are activated. Then, a dynamic environment model based on a reduction strategy is constructed via S3. The minimum interaction feature set, full-body state vector, and task primitives are input into the dynamic environment model for short-time-domain forward simulation to obtain multiple sets of control command sequences. Finally, S4 filters these multiple sets of control command sequences to obtain the optimal control sequence. The system generates a sequence of commands, and finally drives the robot to move through S5 to complete the current control cycle, entering the next control cycle and repeating S1-S5. This scheme uses a task primitive library of parameterized strategy vectors, which can adjust the core parameters to adapt to different target devices and complex terrains, thereby improving the scenario adaptability of navigation control. High-speed simulation is achieved by combining a dynamic environment model with a Ross reduction model and local linearization processing to meet the robot's real-time requirements. At the same time, hierarchical division of labor and parameter calibration ensure simulation accuracy. Through this robot positioning and navigation control method, accurate and efficient navigation control can be achieved in inspection scenarios.

[0124] This embodiment provides an electronic device, which may include: at least one processor, at least one network interface, a user interface, a memory, and at least one communication bus.

[0125] The following is a detailed introduction to the various components of the electronic device:

[0126] The communication bus can be used to enable communication between the various components mentioned above.

[0127] The user interface may include buttons, and optional user interfaces may also include standard wired interfaces and wireless interfaces.

[0128] The network interface may include, but is not limited to, Bluetooth modules, NFC modules, Wi-Fi modules, etc.

[0129] The processor may include one or more processing cores. It connects various parts of the electronic device via various interfaces and lines, executing instructions, programs, code sets, or instruction sets stored in memory, and accessing data stored in memory to perform various functions and process data. Optionally, the processor can be implemented using at least one hardware form of DSP, FPGA, or PLA. The processor can integrate one or more of the following: CPU, GPU, and modem, for example, one or more digital signal processors (DSPs) or one or more field-programmable gate arrays (FPGAs). The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor.

[0130] The memory may include RAM or ROM. Optionally, the memory may include a non-transitory computer-readable medium. The memory may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (e.g., touch function, sound playback function, image playback function, etc.), instructions for implementing the various method embodiments described above, etc.; the data storage area may store data involved in the various method embodiments described above, etc. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor. The memory, as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an evaluation application. The processor may be used to call the evaluation application stored in the memory and execute the method steps mentioned in the foregoing embodiments.

[0131] It should be noted that the above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0132] The above embodiments can be implemented, in whole or in part, through software, hardware (such as circuits), firmware, or any other combination thereof.

[0133] When implemented using software, the above embodiments can be implemented in whole or in part as a computer program product, which includes one or more computer instructions or computer programs; when the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present invention are generated in whole or in part.

[0134] It is understood that the computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device; the computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via transmission methods such as infrared, wireless, or microwave; the computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0135] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0136] It should be understood that, in the embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0137] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A robot positioning and navigation control method, characterized by, It includes the following steps: S1: Acquire the perception data and whole-body state vector of the inspection robot, and extract the minimum interaction feature set and target device identification information from the perception data; the minimum interaction feature set includes local terrain height, the first derivative of local terrain height, the second derivative of local terrain height, and the ground attribute estimation vector of the robot contact point set; Wherein, the first derivative and second derivative of the local terrain height include the slope and curvature of the local terrain, the robot contact point set represents the robot's expected landing point, and the ground attribute estimation vector includes the subsidence coefficient, shear modulus, and internal friction angle. S2: Construct a task primitive library, compare the target device identification information with the task primitive library, and activate the task primitive that matches the target device identification information; S3: Based on the reduction strategy, a dynamic environment model is constructed. The minimum interaction feature set, the whole body state vector and the task primitive are input into the dynamic environment model for short-time domain forward simulation to obtain multiple sets of control command sequences. The dynamic environment model includes a data processing layer, a core simulation layer and an output adaptation layer. The data processing layer standardizes the minimum interaction feature set, whole-body state vector, and task primitives of the input dynamic environment model. The core simulation layer uses the minimum interaction feature set and the whole-body state vector as initial state parameters, the task primitives as simulation boundaries, and combines dynamic functions to perform simulation and obtain simulation results. Specifically, the Ross function is reduced based on navigation control accuracy to obtain the Ross reduced model, and the Ross reduced model is locally linearized to obtain the dynamic function; The output adaptation layer receives and parses the simulation results, generating multiple sets of control command sequences. S4: Based on task primitives, filter multiple sets of control instruction sequences to obtain the optimal control instruction sequence; S5: Drive the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle and repeat S1-S5.

2. The method of claim 1, wherein, Acquire the inspection robot's perception data and full-body state vector, and extract the minimum interaction feature set and target device identification information from the perception data, including: The inspection robot collects perception data from its perception front end, which includes a laser sensor and a multispectral vision sensor. The perception data includes laser echo intensity data and multispectral visual texture data. The whole-body state vector of the inspection robot is collected synchronously, including the center of mass pose, joint angles, and actuator torques. Feature extraction is performed on the perceived data to obtain the minimum set of interactive features and target device identification information.

3. The method of claim 1, wherein, Construct a task primitive library, compare the target device identification information with the task primitive library, and activate the task primitives that match the target device identification information, including: Construct a task primitive library, which includes proximity detection primitives and terrain traversal primitives; All task primitives in the task primitive library are parameterized policy vectors that include the target device, the desired observation pose, and the stable time window. The target device identification information is matched with the task primitive library and the corresponding task primitive is activated.

4. The method of claim 1, wherein, Based on task primitives, multiple sets of control instruction sequences are filtered to obtain the optimal control instruction sequence, including: Using the parameterized policy vector in the task primitive as the evaluation standard, multiple sets of control instruction sequences are evaluated and scored, and the control instruction sequence with the highest evaluation score is selected as the optimal control instruction sequence. When the task primitive is a proximity detection primitive, the evaluation items include the alignment error between the camera line of sight and the target, the time to reach the target pose, and the overall energy consumption; when the task primitive is a terrain crossing primitive, the evaluation items include the center of mass stability, joint torque margin, and slippage risk prediction.

5. The method of claim 1, wherein, The robot moves based on the optimal control command sequence to complete the current control cycle, then enters the next control cycle and repeats S1-S5, including: The first control instruction of the optimal control instruction sequence is sent to the robot actuator, which then drives the robot to perform the corresponding action. Collect the robot's actual motion state data and feed the actual motion state data back to the dynamic environment model to update the dynamic environment model; After completing this control cycle, the next control cycle begins, and S1 to S5 are executed repeatedly until the power inspection task is completed.

6. The method of claim 5, wherein, The method further includes: If the task objective does not change during the loop, there is no need to activate the task primitive; the task primitive from the previous control cycle will be used instead. If the task objective changes, the task primitive corresponding to the changed task objective will be activated.

7. A robot positioning and navigation control system for implementing the method of any one of claims 1-6, characterized by include: The acquisition module is used to acquire the sensory data and whole-body state vector of the inspection robot; The extraction module extracts a minimum set of interactive features and target device identification information from the perceived data; The construction module is used to build a task primitive library and construct a dynamic environment model based on a reduction strategy. An activation module, which compares the target device identification information with the task primitive library and activates the task primitive that matches the target device identification information; The simulation module performs short-time forward simulation of the minimum interaction feature set, the whole-body state vector, and the task primitive input dynamic environment model to obtain multiple sets of control command sequences. A filtering module filters multiple sets of control instruction sequences based on task primitives to obtain the optimal control instruction sequence; The loop module drives the robot to move based on the optimal control command sequence to complete the current control cycle, enter the next control cycle, and repeat S1-S5.

Images