Quadruped robot obstacle avoidance control method and system based on path planning

Abstract

The invention discloses a quadruped robot obstacle avoidance control method and system based on path planning, and relates to the technical field of quadruped robot obstacle avoidance control, and the method comprises the steps: obtaining environment point cloud data, image data and quadruped robot motion speed data, and carrying out the terrain semantic segmentation and semantic pixel back projection of the image data, thereby obtaining a quadruped robot obstacle avoidance result; generating semantic point cloud data, combining the semantic point cloud data with the environment point cloud data subjected to motion compensation, and constructing a semantic annotation grid map; performing global path search on the semantic annotation grid map to generate a global smooth path; when the quadruped robot advances along the global smooth path, path nodes on the global smooth path are extracted at a fixed step pitch, the terrain category and the grid occupation state of each path node are judged according to the semantic annotation grid map, and a local reference path is generated; and executing a multi-stage obstacle avoidance strategy on the local reference path, and generating a foot end track sequence.

Description

Technical Field

[0001] The present invention relates to the technical field of quadruped robot obstacle avoidance control, and in particular to a quadruped robot obstacle avoidance control method and system based on path planning. Background Art

[0002] In the field of autonomous obstacle avoidance on mobile platforms, quadruped robots generally use lidar or RGB-D sensors to obtain environmental point clouds or depth maps, construct two-dimensional or three-dimensional occupancy grid maps, and generate smooth paths through global path search algorithms. Subsequently, most solutions use local obstacle avoidance algorithms such as dynamic window methods or time-domain boundaries to avoid static and dynamic obstacles around the path in real time. During execution, the robot's motion state is estimated through sensors such as joint encoders or IMUs, achieving foot-end trajectory tracking and adjusting the travel speed based on the preset speed and obstacle detection results. This type of technology has been widely used in complex indoor and outdoor environments, providing quadruped robots with basic path planning and obstacle avoidance capabilities.

[0003] However, conventional approaches lack dynamic switching of gait strategies based on terrain type, typically using fixed gaits that struggle to balance energy efficiency and stability. Furthermore, monitoring of foot contact status and obstacle avoidance trigger frequency is relatively coarse, making it impossible to proactively replan global or local paths or regenerate foot trajectories based on error or slip recovery. These shortcomings limit the stability and adaptability of quadruped robots during continuous locomotion in complex terrain and changing environments. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a quadruped robot obstacle avoidance control method and system based on path planning to solve the problems of lack of dynamic gait switching in complex terrain and changing environments and lack of adaptive adjustment of path and trajectory based on sole contact status and obstacle avoidance triggering conditions.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, the present invention provides a quadruped robot obstacle avoidance control method based on path planning, which comprises: Obtain environmental point cloud data, image data, and quadruped robot motion speed data, generate semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merge it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map; Perform global path search on semantically annotated raster maps to generate global smooth paths; When the quadruped robot moves along a global smooth path, it extracts path nodes on the global smooth path with a fixed step distance. It then determines the terrain type and grid occupancy status at each path node based on a semantically annotated grid map to generate a local reference path. Execute multi-level obstacle avoidance strategies on the local reference path and generate a sequence of foot-end trajectories; Foot trajectory tracking is performed through joint drive, and the foot trajectory tracking error and grid occupancy status are monitored in real time. At the same time, the triggering frequency of the multi-level obstacle avoidance strategy is monitored, and global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration are performed according to the triggering frequency.

[0007] As a preferred solution of the path planning-based obstacle avoidance control method for a quadruped robot described in the present invention, the generation of semantic point cloud data includes performing terrain semantic segmentation on image data, classifying color image pixels into four semantic labels: ground, low obstacles, high obstacles, and traversable obstacles, and performing semantic pixel back-projection on the semantic labels and depth images to generate semantic point cloud data.

[0008] As a preferred solution of the path planning-based obstacle avoidance control method of the quadruped robot described in the present invention, the construction of a semantically annotated grid map includes performing motion compensation on the environmental point cloud data based on the motion speed data of the quadruped robot, and merging the semantic point cloud data with the motion-compensated environmental point cloud data to construct a semantically annotated grid map.

[0009] As a preferred solution of the quadruped robot obstacle avoidance control method based on path planning described in the present invention, the steps of generating a global smooth path are as follows: The semantically labeled grid map is filtered according to the foot-end height of the quadruped robot, and the filtered voxels are projected onto the horizontal plane to construct a two-dimensional grid. Perform pass determination on the 2D grid based on semantic labels and voxel occupancy probabilities; Based on the traffic judgment results, a path search is performed on the two-dimensional grid, and redundant inflection points are eliminated through local sight distance for smoothing to generate a global smooth path.

[0010] As a preferred solution of the quadruped robot obstacle avoidance control method based on path planning of the present invention, the steps of generating a local reference path are as follows: On the global smooth path, the straight-line distances between adjacent path nodes on the horizontal plane are accumulated in sequence to generate a cumulative arc length sequence; Perform segmentation on the cumulative arc length sequence to generate equidistant sampling arc lengths; For the sampling arc length, find the first path segment that is greater than or equal to the sampling arc length in the cumulative arc length sequence, and interpolate the starting point and end point of the path segment along a straight line according to the ratio of the sampling arc length to the starting distance of the path segment to obtain the sampling coordinates; Map the sampling coordinates to the grid of the semantically annotated raster map, and perform passability judgment on the path nodes based on the semantic category and voxel occupancy probability of the grid; All traversable nodes are connected in ascending order of sampled arc lengths to generate a local reference path.

[0011] As a preferred embodiment of the path planning-based obstacle avoidance control method for a quadruped robot according to the present invention, the multi-level obstacle avoidance strategy includes executing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation for all nodes in the local reference path; In the execution process of the multi-level obstacle avoidance strategy, gait switching is completed before static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation; After gait switching is completed, the triggering conditions of static obstacle avoidance, dynamic obstacle avoidance, slip recovery and sensor degradation are monitored in parallel within the same obstacle avoidance execution cycle; When static obstacle avoidance is triggered, the local reference path is corrected for obstacle avoidance; When dynamic obstacle avoidance is triggered, adjust the quadruped robot's travel speed; When slip recovery is triggered, roll back the target landing point and insert the recovery node; When sensor degradation is triggered, the posture propulsion mode of the quadruped robot is switched.

[0012] As a preferred solution of the path planning-based obstacle avoidance control method for a quadruped robot described in the present invention, the generation of the foot-end trajectory sequence includes calculating the foot-end trajectory coordinates for the current traversable node and the next traversable node on the local reference path of the quadruped robot, combined with the swing height and swing period of the current gait switching, and generating the foot-end trajectory sequence according to the trajectory update period.

[0013] As an optimal solution of the path planning-based obstacle avoidance control method of a quadruped robot described in the present invention, the real-time monitoring of the foot-end trajectory tracking error and the grid occupancy status, and the simultaneous monitoring of the triggering frequency of the multi-level obstacle avoidance strategy include real-time monitoring of the foot-end trajectory tracking error and the grid occupancy status, and the simultaneous monitoring of the triggering frequency of the multi-level obstacle avoidance strategy, and counting the triggering frequency, which is recorded as the foot-end trajectory tracking error count, the grid occupancy event count, the static obstacle avoidance trigger count, the dynamic obstacle avoidance trigger count, the slip recovery trigger count, and the sensor degradation trigger count.

[0014] As a preferred solution of the path planning-based obstacle avoidance control method for a quadruped robot according to the present invention, the execution of global smooth path replanning, local reference path replanning, and foot end trajectory sequence regeneration according to the trigger frequency includes: When the grid occupancy event trigger count accounts for the highest proportion of all trigger events, global smooth path replanning is performed and all trigger counts are cleared; When the static obstacle avoidance trigger count accounts for the highest proportion of all trigger events, global smooth path replanning is performed and all trigger counts are cleared; When the dynamic obstacle avoidance trigger count accounts for the highest proportion of all trigger events, local reference path replanning is performed and all trigger counts are cleared; When any one of the foot-end trajectory tracking error count, the slip recovery trigger count, and the sensor degradation trigger count accounts for the highest proportion in all trigger events, the foot-end trajectory sequence is regenerated and all trigger counts are cleared.

[0015] In a second aspect, the present invention provides a quadruped robot obstacle avoidance control system based on path planning, comprising: The map construction module is used to obtain environmental point cloud data, image data, and quadruped robot motion speed data. It generates semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merges it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map. The smooth path generation module is used to perform global path search on the semantically annotated raster map and generate a global smooth path; The reference path generation module is used to extract path nodes on the global smooth path with a fixed step distance when the quadruped robot moves along the global smooth path. It also determines the terrain type and grid occupancy status at each path node based on the semantically annotated grid map to generate a local reference path. The multi-level obstacle avoidance execution module is used to execute the multi-level obstacle avoidance strategy on the local reference path and generate the foot end trajectory sequence; The path planning and adjustment module is used to perform foot-end trajectory tracking through joint drive and monitor the foot-end trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the triggering frequency of the multi-level obstacle avoidance strategy and performs global smooth path replanning, local reference path replanning and foot-end trajectory sequence regeneration according to the triggering frequency.

[0016] The present invention achieves the following beneficial effects: by performing terrain semantic segmentation and semantic pixel backprojection on image data, and fusing the motion-compensated environmental point cloud to construct a semantically annotated grid map, the unified expression of geometric and semantic information is achieved, improving the accuracy of path generation. By implementing real-time obstacle monitoring, speed adjustment, and terrain-based gait switching on a local reference path, along with a multi-level obstacle avoidance strategy, the quadruped robot achieves adaptive motion in complex terrain, enhancing its continuity and safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Flowchart of the obstacle avoidance control method for a quadruped robot based on path planning.

[0019] Figure 2 Schematic diagram of the quadruped robot obstacle avoidance control system based on path planning.

[0020] Figure 3 Flowchart generated for a global smoothing path.

[0021] Figure 4 Flowchart generated for a local reference path. DETAILED DESCRIPTION

[0022] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0023] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0025] Reference Figures 1 to 4, is an embodiment of the present invention, which provides a quadruped robot obstacle avoidance control method based on path planning, comprising the following steps: S1. Obtain environmental point cloud data, image data, and quadruped robot motion speed data, generate semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merge it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map.

[0026] Furthermore, a lidar, an RGB-D sensor (a composite camera device capable of simultaneously acquiring color and depth images), and a velocity sensor were installed on the top and front of the quadruped robot body and on the main drive shaft to acquire environmental point cloud data, image data, and quadruped robot motion velocity data. The specific steps are as follows: The laser radar is installed on the rigid platform on the top of the quadruped robot body to achieve 360-degree unobstructed environmental point cloud data collection. At the same time, it is connected to the trigger pulse signal output by the external clock to obtain drift-free environmental point cloud data. The RGB-D sensor is mounted at the center of the front of the fuselage to obtain a wide forward visual field and synchronously capture color images and corresponding depth images using the same external clock. A velocity sensor is installed on the robot's main drive shaft. Leg joint encoders are used to collect real-time velocity data of the quadruped robot, including the angular velocity of each leg joint, namely the hip roll axis angular velocity, hip pitch axis angular velocity, and knee pitch axis angular velocity read by the leg joint encoders, for a total of twelve joint angular velocities. The sampling time is calibrated using the trigger pulse signal output by an external clock to obtain velocity sensor data with a timestamp. Among them, the external clock refers to the unified trigger pulse signal output by a dedicated clock distributor (such as a GPS synchronized clock), which is connected to the acquisition interface of each sensor through a cable to ensure time synchronization and unified timestamp of multi-sensor data acquisition.

[0027] Furthermore, the image data is subjected to terrain semantic segmentation using a semantic segmentation network model (such as the DeepLabv3 model). Each pixel in the color image is classified into four semantic labels: ground, low obstacle, high obstacle, and traversable obstacle. The semantic labels are then back-projected onto the depth image to generate semantic point cloud data. At the same time, the environmental point cloud data is motion-compensated based on the quadruped robot's motion speed data. The semantic point cloud data is then merged with the motion-compensated environmental point cloud data and input into a 3D octree grid to construct a semantically labeled grid map. The specific steps are as follows: Using multiple sets of checkerboard images and Zhang Zhengyou's calibration method, checkerboard images of known size are captured. The captured checkerboard images should be acquired from different angles and positions to ensure that the checkerboard corner information in the checkerboard images is rich and clearly visible. The checkerboard corners in each image are extracted through image processing algorithms to obtain the pixel coordinates of the corners in the image. An optimization algorithm for camera calibration, such as the least squares method, is used to calculate the known physical size of the checkerboard and the pixel coordinates of the corners in the image to obtain the intrinsic parameters of the RGB-D sensor, including the horizontal focal length, vertical focal length, and principal point coordinates. At the same time, the distortion parameters of the image are calculated. , including radial distortion and tangential distortion of the camera lens. The distortion parameters will cause the pixel coordinates of the image to shift, affecting the geometric shape of the image. Based on the internal parameters and distortion parameters of the RGB-D sensor, the remapping function is used to map the pixel coordinates of each original image to the corrected coordinates, and a pixel correction mapping table is generated. The color image and depth image collected by the RGB-D sensor are mapped pixel by pixel through the pixel correction mapping table, and the original coordinates of each pixel are converted to the corrected coordinates to eliminate the lens distortion and achieve pixel-level alignment of the color image and the depth image, thereby obtaining the corrected color image and the corrected depth image. The corrected color image is input into the pre-trained DeepLabv3 model. The DeepLabv3 model is based on the ResNet-101 backbone network and the dilated spatial pyramid pooling decoder. It combines the local features of the image with the global context information through deep learning methods for pixel-level classification. Specifically, the DeepLabv3 model uses a convolutional neural network to extract features from each pixel and calculate the category probability of each pixel. During the pixel-level classification process, the DeepLabv3 model expands the receptive field through dilated convolution operations, which can simultaneously capture local details and the global context information of the image, thereby improving the classification ability of complex scenes. Finally, the DeepLabv3 model assigns four types of semantic labels to each pixel of the color image, including ground, low obstacles, high obstacles, and crossable obstacles, and generates a semantic label map with the same resolution as the corrected color image. The ground refers to the area of ​​the obstacle or terrain surface where the local mean height difference relative to the current support plane of the quadruped robot does not exceed one twentieth of the quadruped robot's leg length. It is generally flat or slightly uneven terrain and is suitable for direct walking without avoiding or adjusting gait. Low obstacles refer to obstacles or terrain surfaces with a local mean height difference relative to the quadruped robot's current support plane that is greater than one twentieth and no more than one fifth of the quadruped robot's leg length, such as small rocks and low steps. High obstacles refer to obstacles or terrain surfaces whose local mean height difference relative to the quadruped robot's current support plane is greater than one-fifth and no more than one-third of the quadruped robot's leg length, such as large rocks and high steps. A traversable obstacle is an obstacle or terrain surface whose local mean height difference relative to the quadruped robot's current support plane is greater than one-third and no more than three-quarters of the quadruped robot's leg length, and whose width does not exceed the quadruped robot's maximum horizontal distance in a jumping motion. Examples include low fences and the edge of a wide platform. The local mean of the current support plane refers to the support plane formed by the landing point of the quadruped robot's current standing support leg as a reference benchmark. The average height value is obtained by performing local smooth fitting calculations on the support plane. The average height value is used as a reference benchmark for relative height to evaluate the height difference between the target position and the reference plane.

[0028] Traverse the semantic label map pixel by pixel. When the semantic label of the color image pixel in the semantic label map belongs to any of the categories of ground, low obstacle, high obstacle and crossable obstacle, read the depth value of the corresponding color image pixel in the rectified depth image, mark the color image pixel coordinates and depth value as semantic pixels, and accumulate them to generate a semantic pixel set. In the semantic pixel set, the horizontal and vertical coordinates of each semantic pixel and the corresponding depth value are first read. Then, combined with the horizontal focal length, vertical focal length and principal point coordinates obtained by RGB-D sensor calibration, according to the view back-projection principle, the horizontal component of the 3D point is generated by combining the pixel horizontal coordinate, principal point horizontal coordinate, depth value and horizontal focal length. The vertical component of the 3D point is generated by combining the pixel vertical coordinate, principal point vertical coordinate, depth value and vertical focal length. The depth value itself is the vertical component of the point. In this way, the 3D point of each semantic pixel in the RGB-D sensor coordinate system is obtained. These 3D points form the semantic point cloud data in the RGB-D sensor coordinate system. Taking the semantic point cloud data in the RGB-D sensor coordinate system as input, the extrinsic parameters of the RGB-D sensor and quadruped robot body coordinate systems, namely the rotation matrix and translation vector that describe the RGB-D sensor coordinate system relative to the quadruped robot body coordinate system, are used to transform each 3D point by first performing a rotation matrix transformation and then superimposing the translation vector. This method maps the semantic point cloud data from the RGB-D sensor coordinate system to the quadruped robot body coordinate system, obtaining the semantic point cloud data in the quadruped robot body coordinate system. At the same time, according to the twelve-way joint angular velocity collected by the velocity sensor and combined with the geometric parameters of the quadruped robot's legs, each joint angular velocity is converted into the corresponding foot-end linear velocity, then the foot end in the supporting foot is identified, and its foot-end linear velocity is extracted on the supporting foot respectively. The arithmetic average of the foot-end linear velocity in the forward direction of the fuselage is obtained to obtain the forward linear velocity, and the arithmetic average of the lateral linear velocity in the lateral direction of the fuselage is obtained to obtain the lateral linear velocity. Then, the forward linear velocity difference of the supporting feet on both sides is combined with the fuselage width to obtain the yaw angular velocity. In order to eliminate point cloud distortion, motion compensation is performed on each frame of environmental point cloud data based on uniform motion, and the maximum forward velocity and maximum yaw angular velocity of the quadruped robot on an unobstructed flat ground are used as the upper limit of motion compensation. Specifically, the maximum forward velocity and the maximum yaw angular velocity are assumed to remain constant within the laser frame sampling period, and the posture increment is calculated by numerical integration, and the posture increment is applied to the coordinates of each point in the original environmental point cloud data to eliminate the distortion caused by the movement of the quadruped robot, thereby obtaining the environmental point cloud data after motion compensation. The maximum safe forward speed refers to the highest linear speed at which a quadruped robot can maintain its center of mass within the support polygon and maintain a stable foot contact force distribution when traveling on flat ground. When the forward speed exceeds the maximum safe forward speed, the quadruped robot may experience excessive center of mass deviation, slippage of the supporting legs, or joint overload during the switching process between the three-legged and four-legged support phases, resulting in unstable movement. The upper limit of the maximum safe forward speed is determined by factors such as the quadruped robot's own center of gravity height, leg length, and leg driving ability, as well as the friction coefficient between the sole of the foot and the ground and the dynamic load distribution during the switching phase. It is usually determined by comparing the stability margin and foot force distribution at different speeds through dynamic simulation and ground experiments. The maximum safe yaw rate is the highest angular velocity at which the foot contact force does not exceed the friction limit and the center of mass projection does not deviate from the support polygon when the quadruped robot is turning laterally or rotating in place. When the yaw rate exceeds the maximum safe yaw rate, the quadruped robot's legs may slip during turning due to insufficient friction, or the support leg combination may become unbalanced due to steering inertia, causing posture instability. The upper limit of the maximum safe yaw rate depends on parameters such as the quadruped robot's body width, leg extension angle, foot friction performance, and joint drive response speed, and must be determined through steering dynamic testing and friction limit testing. The laser frame sampling period is determined based on the rated scanning frequency of the lidar and the terminal point cloud processing delay test, ensuring that at least one to five point clouds are collected within each laser frame sampling period, while leaving a calculation buffer of no less than 10 milliseconds to achieve a balance between real-time performance and processing load.

[0029] The motion-compensated environmental point cloud data and the semantic point cloud data in the quadruped robot body coordinate system are inserted into the three-dimensional octree grid. For each voxel, its logarithmic probability is updated by the observed hit logarithmic probability and the observed miss logarithmic probability, and the voxel occupancy probability is calculated in combination with the prior logarithmic probability. Then, the occupancy threshold is adaptively calculated through a calibration experiment. Specifically, the calibration experiment first records the true state of each voxel in a scene containing a rigid plane or column (real occupied area) and an unobstructed area (real idle area). Then, multiple frames of environmental point cloud data are continuously collected, and the ratio of the number of times each real occupied voxel is judged to be occupied to the total number of times it is detected is counted, which is defined as the hit probability. The ratio of the number of times each real idle voxel is incorrectly judged to be idle to the total number of times it is detected is counted, which is defined as the miss probability. The occupancy threshold is calculated by combining the complementary terms of the hit probability and the miss probability and normalizing the combined result, which is expressed as: ; in, represents the occupation threshold, represents the probability of hitting, represents the probability of loss; The idle threshold is obtained by the complementary calculation of the occupied threshold; When the voxel occupancy probability is higher than the occupancy threshold, it is marked as occupied; when it is lower than the idle threshold, it is marked as idle; the rest remain unknown; after the voxel occupancy status is marked, the stored semantic labels are statistically counted inside each voxel, and the one with the highest count is used as the semantic label of the voxel, finally constructing a semantically labeled raster map containing occupancy status and semantic labels of ground, low obstacle, high obstacle and traversable obstacle.

[0030] It should be noted that the DeepLabv3 model adopts the following training process in this method: First, ResNet-101 (a convolutional neural network structure composed of 101 layers of deep residual blocks) is selected as the backbone network to extract deep and shallow image features, and the void space pyramid pooling module and upsampling decoder are connected on the top of ResNet-101 to take into account both the global receptive field and the restoration of boundary details; the DeepLabv3 model has four input channels, corresponding to the three-channel color image output by the RGB-D sensor and the one-channel linearly normalized depth image (for example, the original depth value is linearly mapped from 0-5 meters to the range of 0-1); the training data comes from the self-built semantic segmentation dataset, which contains 10,000 frames of color and depth images with a resolution of 640×480, and is accompanied by four types of semantic labels: pixel-level ground, low obstacles, high obstacles, and traversable obstacles. The amount of training data ensures both scene diversity and training time. Training efficiency; To accelerate the convergence of the DeepLabv3 model and improve scene generalization ability, 100 rounds of pre-training were first performed on the Cityscapes public dataset. During the pre-training, the DeepLabv3 model parameters were optimized using the stochastic gradient descent method and all training samples were traversed in each round. Then, the self-built semantic segmentation dataset was fine-tuned for 150 rounds using the stochastic gradient descent optimizer with a learning rate of 0.001, a momentum of 0.9, and a weight decay of 0.0001. The fine-tuning batch size was set to 4 to adapt to the mainstream GPU memory; the image enhancement strategy used during training included random horizontal flipping of the input image (flip probability 50%), random scaling (for example, enlarging or reducing to 0.5 to 2 times the original size), and random cropping to a fixed size (for example, 512×512 pixels) to cover different shooting distances and perspective changes; the training goal was to minimize the pixel-level cross entropy loss and ultimately achieve an average intersection-over-union ratio of 85% on the validation set; Among them, the learning rate, the example value range is 0.0001 to 0.01, which is determined based on the trade-off between convergence speed and gradient stability. Specifically, a grid search is performed with learning rates of 0.0001, 0.001, 0.005 and 0.01, and the decrease rate of pixel-level cross entropy loss and the change of gradient norm on the validation set are monitored. The results show that when the learning rate is between 0.0001 and 0.01, the loss curve does not show obvious oscillation and can converge to a stable level within a limited number of training rounds. When it is lower than 0.0001, the convergence is too slow, and when it is higher than 0.01, the training is prone to divergence; Momentum, for example, ranges from 0.8 to 0.95. This is determined based on the need for oscillation suppression and convergence acceleration. Specifically, by comparing four candidate values ​​of 0.8, 0.9, 0.95, and 0.99, and observing the oscillation amplitude and iterative convergence speed during training, the optimal range of 0.8 to 0.95 is determined. A value below 0.8 decreases the convergence speed, while a value above 0.95 exacerbates gradient oscillation. Weight decay, the example value range is 0.00001 to 0.001, which is determined based on preventing overfitting and maintaining the capacity balance of the DeepLabv3 model. Specifically, weight decay is verified in 0.00001, 0.0001, 0.001 and 0.01. By comparing the accuracy of the validation set and the stability of the loss curve, 0.00001 to 0.001 is selected as the appropriate range, which can effectively suppress overfitting without weakening the feature expression ability of the DeepLabv3 model due to excessive regularization.

[0031] S2. Perform a global path search on the semantically annotated raster map to generate a global smooth path.

[0032] Furthermore, the Theta* algorithm is used on the semantically annotated raster map to search for paths at any angle, and redundant vertex points are eliminated by local sight distance to achieve fast smoothing and generate a global smooth path. The specific steps are as follows: To limit the reach of the quadruped's foot, we first reserve voxels within a vertical range of –0.05 to 0.15 meters from the semantically annotated grid map. The –0.05-meter range is based on slight ground depressions, and the 0.15-meter range is based on the maximum obstacle height required for the quadruped to cross. The reserved voxels are then projected onto a horizontal plane, and a two-dimensional grid is constructed with a grid spacing of 0.05 meters. This grid spacing is based on the matching requirement between the RGB-D sensor point cloud resolution and the foot's positioning accuracy. A traffic check is performed on each 2D grid. Specifically, the semantic label and voxel occupancy probability of each grid are simultaneously determined. When the semantic label of the voxel corresponding to the grid is ground or crossable obstacle and the voxel occupancy probability is lower than the occupancy threshold, the grid is marked as free. When the semantic label is low obstacle or high obstacle or the voxel occupancy probability is higher than or equal to the occupancy threshold, the grid is marked as obstructed. In other cases, the grid is marked as unknown. After marking the two-dimensional grids, the Theta* algorithm is applied to all grids marked as free or unknown to perform arbitrary-angle path search. First, the starting grid is placed in the open list and its cumulative cost is set to zero. The following steps are repeated until the target grid is visited or the open list is empty. The grid with the smallest sum of the cumulative cost and the heuristic cost is selected from the open list as the current grid. The reachability of the current grid's eight adjacent grids is checked one by one. First, a check is performed to determine whether the straight line connection between the current grid and the adjacent grid passes through any grid marked as an obstruction. If the straight line connection does not pass through an obstruction grid, the sum of the current grid's cumulative cost and the straight-line distance between the two grid centers on the horizontal plane is used as the new migration cost for the adjacent grid. Otherwise, the sum of the current grid's cumulative cost and the standard distance for eight-connectivity (i.e., vertical, horizontal, and diagonal adjacency) is used as the migration cost. The horizontal straight-line distance from the adjacent grid to the target grid is then used as the heuristic cost. If the adjacent grid is not already in the open list or the new migration cost is less than the adjacent grid's original cumulative cost, the adjacent grid's cumulative cost and parent grid pointer are updated, and the adjacent grid is added to the open list. After the target grid is accessed, the initial path is formed by tracing back from the target grid along the parent grid pointer to the starting grid. To reduce the number of path vertices, the initial path is thinned according to the visibility rule of three adjacent grids. Specifically, for any three consecutive grids in the initial path, if the grids passing along the straight line between the first and last grids are not marked as obstacles, the middle grid is deleted. After traversing the entire path, a global smooth path with the fewest vertices and the smoothest curve is obtained.

[0033] It should be noted that the Theta* algorithm is an arbitrary-angle path search algorithm based on the A* algorithm. Its main difference from the A* algorithm lies in the line-of-sight judgment of path connectivity when expanding nodes, thereby generating paths with fewer inflections and closer to a straight line. Specifically, the Theta* algorithm maintains the same open and closed list structures as the A* algorithm, but when expanding each adjacent grid of the current grid, it first checks the straight-line connectivity between the parent grid and the adjacent grid. If all intermediate grids between the parent grid and the adjacent grid are marked as free or unknown, it directly attempts to relax the adjacent grid through the parent grid, calculates the true Euclidean distance from the parent grid to the adjacent grid, and adds the accumulated cost of the parent grid as the new migration cost of the adjacent grid. Otherwise, it falls back to the traditional A* method, using the adjacent distance from the current grid to the adjacent grid plus the accumulated cost of the current grid as the migration cost. After calculating the migration cost of each adjacent mesh, the Euclidean distance from the adjacent mesh to the target mesh is used as a heuristic estimate. The sum of the two determines the mesh's priority. If an adjacent mesh already exists in the open list and the new migration cost is smaller, the cumulative cost and parent pointer are updated; otherwise, the adjacent mesh is inserted into the open list. By dynamically switching between the "straight-line relaxation" and "eight-connected relaxation" rules, the Theta* algorithm significantly reduces path breakpoints while ensuring search integrity. Ultimately, the initial path formed by backtracking along the parent pointer after the target is visited is a breakpoint-optimized path that connects at any angle.

[0034] S3. When the quadruped robot moves along the global smooth path, it extracts path nodes on the global smooth path with a fixed step distance, and judges the terrain category and grid occupancy status at each path node based on the semantically annotated grid map to generate a local reference path.

[0035] Furthermore, the straight-line distances between adjacent path points on the horizontal plane are calculated sequentially from the starting point of the global smooth path, and all straight-line distances are sequentially added to form a cumulative arc length sequence. The cumulative arc length sequence represents the travel distance of any position on the global smooth path relative to the starting point. The cumulative arc length sequence is segmented using an example step length of 0.2 meters as the node sampling interval. The first segmentation at the starting point of the global smooth path corresponds to a distance of 0.2 meters, and so on to generate a set of equidistant sampling arc lengths. The example step length of 0.2 meters is determined based on the field test results of the optimal gait stability and travel efficiency of the quadruped robot. For each sampling arc length, the cumulative position where the sampling arc length is first reached or exceeded is searched from the cumulative arc length sequence, and the corresponding path segment is located. The path segment consists of two adjacent path nodes, the start point and the end point. Within the path segment, linear interpolation is performed between the start point and the end point of the path segment according to the distance ratio between the sampling arc length and the cumulative arc length at the start of the path segment. The sampling coordinates are evenly calculated between the start point and the end point according to the distance ratio between the sampling arc length and the cumulative arc length at the start of the path segment, thereby accurately reflecting the equidistant nodes on the global smooth path.

[0036] Furthermore, each sampling coordinate is projected onto the corresponding grid of the semantically annotated grid map, the semantic label and voxel occupancy probability of the grid are read, and the path node passability is determined according to the following rules: When the voxel occupancy probability is lower than the idle threshold and the semantic label is ground or crossable obstacle, it is determined to be a passable node; when the voxel occupancy probability is not lower than the occupancy threshold or the semantic label is low obstacle or high obstacle, it is determined to be an obstruction node; in other cases, it is determined to be an uncertain node; The sampling coordinates of all nodes judged to be traversable are connected in ascending order of sampling index to generate a local reference path. The sampling coordinates of nodes judged to be uncertain are retained for subsequent processing by the multi-level obstacle avoidance strategy. The obtained local reference path is then used for dynamic obstacle avoidance and foot trajectory generation.

[0037] S4. Execute a multi-level obstacle avoidance strategy on the local reference path and generate a foot trajectory sequence.

[0038] Furthermore, the multi-level obstacle avoidance strategy includes performing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation for all nodes in the local reference path. The specific steps are as follows: Gait switching is always at the top of the multi-level obstacle avoidance strategy. Before judging static obstacle avoidance, dynamic obstacle avoidance, slip recovery and sensor degradation, gait switching is performed on all sampling nodes on the local reference path. Specifically, two sample landing points are taken on the local reference path, and the corresponding voxel semantic label is read from the semantically labeled grid map for each landing point. When the semantic label is ground, a fast gait is selected and the example swing height is set to one twentieth of the leg length of the quadruped robot; when the semantic label is low obstacle, a slow gait is selected and the example swing height is set to one fifth of the leg length of the quadruped robot; when the semantic label is high obstacle, a crawling gait is selected and the example swing height is set to one third of the leg length of the quadruped robot; when the semantic label is surmountable obstacle, a jumping gait is selected and the example swing height is set to three quarters of the leg length of the quadruped robot; The swing height of the example fast gait is taken as one twentieth of the quadruped robot's leg length, the slow gait is taken as one fifth of the quadruped robot's leg length, the crawling gait is taken as one third of the quadruped robot's leg length, and the jumping gait is taken as three quarters of the quadruped robot's leg length. These proportions are determined based on the kinematic reach and static stability margin analysis. Assuming the quadruped robot's leg length is 250 mm, to ensure that the foot can cross a maximum of about 40 mm ground undulation without rubbing, the swing height needs to be set to one twentieth of the quadruped robot's leg length. For low obstacles with a height not exceeding one-fifth of the quadruped robot's leg length, a swing height of one-fifth of the quadruped robot's leg length can be used to smoothly cross them without significantly shrinking the three-legged support polygon. For high obstacles with a height of about one-third of the quadruped robot's leg length, a swing height of one-third of the quadruped robot's leg length can ensure that the center of mass projection still falls within the support area during the single-leg support phase. When the semantic label indicates that the obstacle can be crossed and the obstacle height does not exceed three-quarters of the quadruped robot's leg length, a jumping gait is adopted and the swing height is set to three-quarters of the quadruped robot's leg length to ensure that sufficient safety clearance is maintained between the foot and the ground when switching between swinging and landing. The above height ratios are obtained through kinematic derivation and stability margin analysis and do not require additional field testing and verification. After completing the gait switching of all landing points, the triggering conditions of static obstacle avoidance, dynamic obstacle avoidance, slip recovery and sensor degradation are monitored in parallel within the same obstacle avoidance execution cycle; when static obstacle avoidance is triggered, the local reference path is corrected for obstacle avoidance; when dynamic obstacle avoidance is triggered, the moving speed of the quadruped robot is adjusted; when slip recovery is triggered, the target landing point is retracted and a recovery node is inserted; when sensor degradation is triggered, the posture propulsion mode of the quadruped robot is switched; all corrections are completed based on the gait switching results, and a foot-end trajectory sequence matching the current gait and swing height is uniformly generated before the end of the current obstacle avoidance execution cycle.

[0039] When the quadruped robot moves along the local reference path, the example obstacle avoidance execution cycle is 10 milliseconds as the time interval, and the voxel occupancy probability of all grids within the example 1 meter range in front of the local reference path, including the grids where the nodes marked as obstacles and uncertain are located, is read in sequence. When the occupancy probability of any grid voxel is greater than or equal to the occupancy threshold, static obstacle avoidance is triggered; the D*Lite algorithm is run in the local grid area with a radius of 1 meter and a trigger grid as the starting point and the trigger grid as the center, to recalculate the avoidance corridor; the D*Lite algorithm recalculates the minimum path cost only for the grid area affected by the new obstacle by maintaining a priority queue and incremental cost differential propagation, and dynamically updates the cost value of the adjacent grid in combination with the heuristic distance until a shortest avoidance path that bypasses the obstacle and connects to the boundary of the area is found in the local area; finally, the corresponding path segment in the local reference path is replaced with the shortest avoidance path to ensure that the new static obstacle is avoided; The obstacle avoidance execution cycle is 10 milliseconds, which is determined based on the external clock trigger frequency of the speed sensor. This duration ensures that the voxel occupancy probability acquisition and obstacle avoidance decision-making can be completed within one cycle to meet real-time response. The 1-meter range in this example is based on a quadruped robot with a step length of approximately 0.2 meters. This is the minimum reserved distance required to complete obstacle detection and avoidance within four steps, providing sufficient space for path reconnection.

[0040] During the same obstacle avoidance execution cycle, dynamic obstacle avoidance first reads the real-time position and speed of each dynamic obstacle within a 1-meter range in front of the local reference path. It then calculates the safe obstacle avoidance distance based on the speed sensor perception and decision delay time, plus the obstacle avoidance safety margin, expressed as: ; in, Indicates the safe obstacle avoidance distance. represents the speed of dynamic obstacles, It represents the delay time between speed sensor perception and decision making. The specific time is determined by testing the delay from sensor data acquisition to processing and combining it with the information transmission delay in the decision-making process. represents the obstacle avoidance safety margin; When the horizontal distance between the current position of the quadruped robot and the position of the dynamic obstacle is less than the safe obstacle avoidance distance, a correction vector in the direction of the quadruped robot away from the obstacle is superimposed on the preset reference speed. That is, the direction of the correction vector is opposite to the direction of the obstacle line. The newly generated obstacle avoidance speed vector is obtained, which is expressed as: ; in, represents the newly generated obstacle avoidance velocity vector, Indicates the horizontal distance between the robot and the obstacle, Indicates the current position coordinates of the dynamic obstacle, Indicates the current position coordinates of the quadruped robot, The preset reference speed is the target linear speed of the quadruped robot traveling along a path on a flat, obstacle-free surface. For example, 0.3 m / s is used, which is determined based on the path tracking stability test. The newly generated obstacle avoidance velocity vector is applied to the quadruped robot's travel speed, thereby achieving real-time avoidance of dynamic obstacles. When static avoidance and dynamic avoidance are triggered simultaneously, the local reference path is updated first and then the newly generated obstacle avoidance velocity vector is used to control the speed.

[0041] During the same obstacle avoidance execution cycle, the plantar normal force of each leg of the quadruped robot is monitored by the plantar force sensor of each leg. When the plantar normal force of a leg maintains the typical level of the support phase and lasts long enough to confirm that the leg has completed the support phase switch, the leg is determined to be in the support state. If any of the conditions are met at the same time in the support state, the plantar normal force shows a significant downward trend or the horizontal speed of the foot suddenly increases, then the slip recovery is triggered. After the slip recovery is triggered, the target landing point of the slipping leg is immediately retracted to the nearest passable node in the local reference path, and two equidistant recovery nodes are inserted after the node according to the example step distance of 0.2 meters to maintain stable movement. Among them, the horizontal velocity of the foot end is calculated based on the horizontal displacement difference of the foot end in the coordinate system of the quadruped robot body and the sampling period.

[0042] During the same obstacle avoidance execution cycle, the RGB-D sensor output frame rate is continuously monitored. When the RGB-D sensor continuously drops frames and does not receive new color or depth images, sensor degradation is triggered. After sensor degradation is triggered, the robot estimates the body velocity and propulsion based solely on the leg odometry and laser data. Based on the twelve joint angular velocities collected by the velocity sensor, combined with the quadruped robot's leg geometric parameters and the current gait cycle, the joint angular velocity of each supporting leg during the stance phase is converted into a horizontal displacement increment at the foot end. The displacement increments of all supporting legs are then arithmetic averaged on the horizontal plane to obtain the estimated body linear velocity. The position increment is calculated as the product of the laser frame sampling period and the estimated body linear velocity, and the quadruped robot's posture is advanced along the local reference path. When the RGB-D sensor frame rate is restored, sensor degradation is exited, and the original multi-level obstacle avoidance strategy and foot end trajectory generation process are restored.

[0043] Furthermore, for each swing of the quadruped robot, a cubic polynomial is used to generate a smooth foot trajectory based on the current traversable node and the next traversable node on the local reference path, combined with the swing height and example swing period corresponding to the current gait, and the normalized time ratio is defined, which is expressed as: ; ; in, represents the normalized time ratio, Indicates the swing period, with a range of 0.4 to 0.6 seconds, determined based on gait cycle and dynamic stability tests. Indicates that the foot is at the moment The three-dimensional trajectory coordinates of Indicates the coordinates of the current landing point of the swing foot, Indicates the coordinates of the landing point of the swinging foot, represents the swing height under the corresponding gait, represents the vertical upward unit vector; During the trajectory update cycle (Example value range 0.01 to 0.05 seconds, determined based on the balance between travel control frequency and calculation load), the time Starting from scratch As the step size increases, d(t) is calculated and collected in sequence to form a foot end trajectory sequence, which is used for the lower-level driver to track and execute in time sequence.

[0044] S5. Execute foot-end trajectory tracking through joint drive, and monitor the foot-end trajectory tracking error and grid occupancy status in real time. At the same time, monitor the triggering frequency of the multi-level obstacle avoidance strategy, and perform global smooth path replanning, local reference path replanning, and foot-end trajectory sequence regeneration based on the triggering frequency.

[0045] Furthermore, within the same obstacle avoidance execution cycle, the foot trajectory sequence is converted into corresponding joint angle commands through inverse kinematics and sent to the joint drivers in sequence; at the same time, the actual foot position of each swinging foot is calculated by inverse kinematics using the real-time read joint angles and the quadruped robot leg geometric parameters (such as thigh, calf and knee lengths). Specifically, inverse kinematics learns to accumulate the positions of these joints in the quadruped robot body coordinate system layer by layer based on the known joint angles by calculating the transformation matrix of the relative position of each joint, and finally obtains the actual foot position of each swinging foot; the actual foot position of each swinging foot is combined with the expected foot position (the expected foot position is the foot position at time The three-dimensional trajectory coordinates of the foot are calculated and the foot end trajectory tracking error is expressed as: ; in, represents the foot end trajectory tracking error, Indicates the actual foot end position of each swinging foot; When the foot-end trajectory tracking error is greater than the error threshold, the foot-end trajectory tracking error count is accumulated once. If the foot-end trajectory tracking error is less than or equal to the error threshold, the count is not counted. The error threshold, for example, ranges from 0.01 to 0.05 meters, with a lower limit of 0.01 meters. This is based on the typical foot-end positioning accuracy of approximately ±0.005 meters and the minimum increment of the control command of approximately 0.003 to 0.007 meters. To avoid unnecessary error counts caused by measurement noise and slight command jitter, the lower limit of the error threshold is set to twice the noise level, thereby effectively filtering out these interference factors and ensuring that the error count is only triggered when the actual deviation of the foot-end trajectory is large. The upper limit is used to ensure that the foot-end deviation does not exceed 1 / 4 of the step length, which can fully respond to trajectory deviations without frequently triggering replanning due to slight jitter.

[0046] At the same time, within the same obstacle avoidance execution cycle, the voxel occupancy probability of the grid where the next foot landing point is located on the local reference path is read. When the voxel occupancy probability is greater than or equal to the occupancy threshold, the grid occupancy event trigger count is accumulated once.

[0047] Furthermore, the number of events triggered by the multi-level obstacle avoidance strategy is monitored in parallel, including static obstacle avoidance trigger count, dynamic obstacle avoidance trigger count, slip recovery trigger count, and sensor degradation trigger count. The above counts are accumulated when each trigger is triggered in the corresponding process; After the same obstacle avoidance execution cycle ends, the proportion of each type of trigger count in all trigger events is evaluated in turn to determine whether to update the path or trajectory. All trigger counts are reset to zero after any replanning or regeneration is completed. The specific rules are as follows: If the grid occupancy event trigger count accounts for the highest proportion of all trigger events, then global smooth path replanning is performed and all trigger counts are cleared; Otherwise, if the static obstacle avoidance trigger count accounts for the highest proportion of all trigger events, global smooth path replanning is performed and all trigger counts are cleared; Otherwise, if the dynamic obstacle avoidance trigger count accounts for the highest proportion of all trigger events, local reference path replanning is performed and all trigger counts are cleared; Otherwise, if any of the foot-end trajectory tracking error count, slip recovery trigger count, or sensor degradation trigger count accounts for the highest proportion of all trigger events, the foot-end trajectory sequence is regenerated and all trigger counts are cleared.

[0048] After any replanning or regeneration is completed, all trigger counts are reset to zero, and the original content is replaced by the newly generated global smooth path, local reference path, or foot trajectory sequence. The obstacle avoidance execution cycle enters the next execution cycle and continues to perform obstacle judgment and foot trajectory tracking based on the updated path or trajectory.

[0049] This embodiment also provides a quadruped robot obstacle avoidance control system based on path planning, comprising: The map construction module is used to obtain environmental point cloud data, image data, and quadruped robot motion speed data. It generates semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merges it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map. The smooth path generation module is used to perform global path search on the semantically annotated raster map and generate a global smooth path; The reference path generation module is used to extract path nodes on the global smooth path with a fixed step distance when the quadruped robot moves along the global smooth path. It also determines the terrain type and grid occupancy status at each path node based on the semantically annotated grid map to generate a local reference path. The multi-level obstacle avoidance execution module is used to execute the multi-level obstacle avoidance strategy on the local reference path and generate the foot end trajectory sequence; The path planning and adjustment module is used to perform foot-end trajectory tracking through joint drive and monitor the foot-end trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the triggering frequency of the multi-level obstacle avoidance strategy and performs global smooth path replanning, local reference path replanning and foot-end trajectory sequence regeneration according to the triggering frequency.

[0050] This embodiment also provides a computer device, which is suitable for the case of a quadruped robot obstacle avoidance control method based on path planning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the quadruped robot obstacle avoidance control method based on path planning proposed in the above embodiment.

[0051] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0052] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the path planning-based obstacle avoidance control method for a quadruped robot as proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0053] In summary, the present invention achieves unified representation of geometric and semantic information and improves the accuracy of path generation by performing terrain semantic segmentation and semantic pixel backprojection on image data and fusing it with a motion-compensated environmental point cloud to construct a semantically annotated grid map. Furthermore, by implementing real-time obstacle monitoring, speed adjustment, and terrain-based gait switching on a local reference path, along with a multi-level obstacle avoidance strategy, the robot achieves adaptive motion in complex terrain, enhancing both continuity and safety.

[0054] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A quadruped robot obstacle avoidance control method based on path planning, characterized by: include, Obtain environmental point cloud data, image data, and quadruped robot motion speed data, generate semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merge it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map; Perform global path search on semantically annotated raster maps to generate global smooth paths; When the quadruped robot moves along a global smooth path, it extracts path nodes on the global smooth path with a fixed step distance. It then determines the terrain type and grid occupancy status at each path node based on a semantically annotated grid map to generate a local reference path. Execute multi-level obstacle avoidance strategies on the local reference path and generate a sequence of foot-end trajectories; Foot trajectory tracking is performed through joint drive, and the foot trajectory tracking error and grid occupancy status are monitored in real time. At the same time, the triggering frequency of the multi-level obstacle avoidance strategy is monitored, and global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration are performed according to the triggering frequency.

2. The quadruped robot obstacle avoidance control method based on path planning according to claim 1, characterized in that: The generation of semantic point cloud data includes performing terrain semantic segmentation on image data, classifying color image pixels into four semantic labels: ground, low obstacles, high obstacles, and crossable obstacles, and performing semantic pixel back-projection on the semantic labels and depth images to generate semantic point cloud data.

3. The quadruped robot obstacle avoidance control method based on path planning according to claim 2, characterized in that: The construction of the semantically annotated grid map includes performing motion compensation on the environmental point cloud data based on the motion speed data of the quadruped robot, and merging the semantic point cloud data with the motion-compensated environmental point cloud data to construct the semantically annotated grid map.

4. The quadruped robot obstacle avoidance control method based on path planning according to claim 3, characterized in that: The steps for generating a global smooth path are as follows: The semantically labeled grid map is filtered according to the foot-end height of the quadruped robot, and the filtered voxels are projected onto the horizontal plane to construct a two-dimensional grid. Perform pass determination on the 2D grid based on semantic labels and voxel occupancy probabilities; Based on the traffic judgment results, a path search is performed on the two-dimensional grid, and redundant inflection points are eliminated through local sight distance for smoothing to generate a global smooth path.

5. The quadruped robot obstacle avoidance control method based on path planning according to claim 4, characterized in that: The steps of generating a local reference path are as follows: On the global smooth path, the straight-line distances between adjacent path nodes on the horizontal plane are accumulated in sequence to generate a cumulative arc length sequence; Perform segmentation on the cumulative arc length sequence to generate equidistant sampling arc lengths; For the sampling arc length, find the first path segment that is greater than or equal to the sampling arc length in the cumulative arc length sequence, and interpolate the starting point and end point of the path segment along a straight line according to the ratio of the sampling arc length to the starting distance of the path segment to obtain the sampling coordinates; Map the sampling coordinates to the grid of the semantically annotated raster map, and perform passability judgment on the path nodes based on the semantic category and voxel occupancy probability of the grid; All traversable nodes are connected in ascending order of sampled arc lengths to generate a local reference path.

6. The quadruped robot obstacle avoidance control method based on path planning according to claim 5, characterized in that: The multi-level obstacle avoidance strategy includes performing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery and sensor degradation for all nodes in the local reference path; In the execution process of the multi-level obstacle avoidance strategy, gait switching is completed before static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation; After gait switching is completed, the triggering conditions of static obstacle avoidance, dynamic obstacle avoidance, slip recovery and sensor degradation are monitored in parallel within the same obstacle avoidance execution cycle; When static obstacle avoidance is triggered, the local reference path is corrected for obstacle avoidance; When dynamic obstacle avoidance is triggered, adjust the quadruped robot's travel speed; When slip recovery is triggered, roll back the target landing point and insert the recovery node; When sensor degradation is triggered, the posture propulsion mode of the quadruped robot is switched.

7. The quadruped robot obstacle avoidance control method based on path planning according to claim 6, characterized in that: The generating of the foot-end trajectory sequence includes calculating the foot-end trajectory coordinates for the current traversable node and the next traversable node on the local reference path of the quadruped robot in combination with the swing height and swing period of the current gait switching, and generating the foot-end trajectory sequence according to the trajectory update period.

8. The quadruped robot obstacle avoidance control method based on path planning according to claim 7, characterized in that: The real-time monitoring of the foot-end trajectory tracking error and the grid occupancy status and the simultaneous monitoring of the triggering frequency of the multi-level obstacle avoidance strategy include real-time monitoring of the foot-end trajectory tracking error and the grid occupancy status and the simultaneous monitoring of the triggering frequency of the multi-level obstacle avoidance strategy, and counting the triggering frequency, which is recorded as the foot-end trajectory tracking error count, the grid occupancy event count, the static obstacle avoidance trigger count, the dynamic obstacle avoidance trigger count, the slip recovery trigger count and the sensor degradation trigger count.

9. The quadruped robot obstacle avoidance control method based on path planning according to claim 8, characterized in that: The execution of global smooth path replanning, local reference path replanning and foot end trajectory sequence regeneration according to the trigger frequency includes: When the grid occupancy event trigger count accounts for the highest proportion of all trigger events, global smooth path replanning is performed and all trigger counts are cleared; When the static obstacle avoidance trigger count accounts for the highest proportion of all trigger events, global smooth path replanning is performed and all trigger counts are cleared; When the dynamic obstacle avoidance trigger count accounts for the highest proportion of all trigger events, local reference path replanning is performed and all trigger counts are cleared; When any one of the foot-end trajectory tracking error count, the slip recovery trigger count, and the sensor degradation trigger count accounts for the highest proportion in all trigger events, the foot-end trajectory sequence is regenerated and all trigger counts are cleared.

10. A quadruped robot obstacle avoidance control system based on path planning, based on the quadruped robot obstacle avoidance control method based on path planning according to any one of claims 1 to 9, characterized in that: include, The map construction module is used to obtain environmental point cloud data, image data, and quadruped robot motion speed data. It generates semantic point cloud data by performing terrain semantic segmentation and semantic pixel back-projection on the image data, and merges it with the motion-compensated environmental point cloud data to construct a semantically annotated raster map. The smooth path generation module is used to perform global path search on the semantically annotated raster map and generate a global smooth path; The reference path generation module is used to extract path nodes on the global smooth path with a fixed step distance when the quadruped robot moves along the global smooth path. It also determines the terrain type and grid occupancy status at each path node based on the semantically annotated grid map to generate a local reference path. The multi-level obstacle avoidance execution module is used to execute the multi-level obstacle avoidance strategy on the local reference path and generate the foot end trajectory sequence; The path planning and adjustment module is used to perform foot-end trajectory tracking through joint drive and monitor the foot-end trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the triggering frequency of the multi-level obstacle avoidance strategy and performs global smooth path replanning, local reference path replanning and foot-end trajectory sequence regeneration according to the triggering frequency.

Images