A path planning-based obstacle avoidance control method and system for a quadruped robot

Abstract

The application discloses a kind of four-legged robot obstacle avoidance control method and system based on path planning, it is related to four-legged robot obstacle avoidance control technical field, including, acquisition environment point cloud data, image data and four-legged robot motion speed data, by to image data terrain semantic segmentation and semantic pixel back projection, generate semantic point cloud data, and with the environment point cloud data that is motion compensated is merged, constructs semantic labeling grid map;Global path search is carried out to semantic labeling grid map, generates global smooth path;Four-legged robot travels along global smooth path, with fixed step distance extracts path node on global smooth path, and according to semantic labeling grid map, the terrain class and grid occupancy state at each path node are judged, generate local reference path;Multi-level obstacle avoidance strategy is executed to local reference path, and foot end trajectory sequence is generated.

Description

Technical Field

[0001] This invention relates to the field of obstacle avoidance control technology for quadruped robots, and in particular to a method and system for obstacle avoidance control of quadruped robots based on path planning. Background Technology

[0002] In the field of autonomous obstacle avoidance on mobile platforms, quadruped robots commonly employ LiDAR or RGB-D sensors to acquire environmental point clouds or depth maps, constructing two-dimensional or three-dimensional occupancy grid maps, and generating smooth paths through global path search algorithms. Subsequently, most solutions utilize local obstacle avoidance algorithms such as dynamic windowing or temporal boundary methods to avoid static and dynamic obstacles around the path in real time. During execution, sensors such as joint encoders or IMUs estimate the robot's motion state, enabling foot trajectory tracking, and adjusting the travel speed based on preset speeds and obstacle detection results. This type of technology has been widely applied in complex indoor and outdoor environments, providing quadruped robots with basic path planning and obstacle avoidance capabilities.

[0003] However, conventional methods lack dynamic switching based on terrain type in gait strategies, typically using fixed gait patterns, making it difficult to balance energy consumption and stability. Furthermore, the monitoring of foot contact states and obstacle avoidance trigger frequencies is relatively coarse, failing to proactively perform global or local path replanning and foot trajectory regeneration based on errors or slip recovery. These shortcomings limit the stability and adaptability of quadruped robots during continuous movement in complex terrains and variable environments. Summary of the Invention

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

[0005] Therefore, this invention provides a path planning-based obstacle avoidance control method and system for quadruped robots, which solves the problems of lacking dynamic gait switching and lacking adaptive adjustment of path and trajectory based on foot contact state and obstacle avoidance triggering in complex terrain and variable environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a path planning-based obstacle avoidance control method for a quadruped robot, comprising,

[0008] The system acquires 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 motion-compensated environmental point cloud data to construct a semantically labeled raster map.

[0009] Perform a global path search on the semantically annotated raster map to generate a globally smooth path;

[0010] When the quadruped robot moves along the global smooth path, it extracts the path nodes on the global smooth path with a fixed step distance, and determines the terrain category and grid occupancy status at each path node based on the semantically labeled grid map to generate a local reference path.

[0011] A multi-level obstacle avoidance strategy is implemented on the local reference path, and a sequence of foot trajectory is generated;

[0012] Foot trajectory tracking is performed through joint drive, and foot trajectory tracking error and grid occupancy status are monitored in real time. At the same time, the trigger frequency of 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 trigger frequency.

[0013] As a preferred embodiment of the obstacle avoidance control method for quadruped robots based on path planning described in this 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 of the semantic labels and depth images to generate semantic point cloud data.

[0014] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the construction of the semantically labeled grid map includes performing motion compensation on environmental point cloud data based on the quadruped robot's motion speed data, and merging the semantic point cloud data with the motion-compensated environmental point cloud data to construct the semantically labeled grid map.

[0015] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the steps for generating a globally smooth path are as follows:

[0016] The semantically labeled raster map is filtered into voxels according to the height of the quadruped robot's foot, and the filtered voxels are projected onto a horizontal plane to construct a two-dimensional mesh.

[0017] Perform passage determination on the 2D mesh based on semantic labels and voxel occupancy probability;

[0018] Based on the passage determination results, a path search is performed on a two-dimensional grid, and a smoothing process is performed by eliminating redundant vertices through local sight distance to generate a globally smooth path.

[0019] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the steps for generating a local reference path are as follows:

[0020] On a globally smooth path, the straight-line distances between adjacent path nodes on the horizontal plane are accumulated sequentially to generate a cumulative arc length sequence.

[0021] Segment the cumulative arc length sequence to generate equidistant sampling arc lengths;

[0022] For the sampling arc length, the first path segment with a length greater than or equal to the sampling arc length is found in the cumulative arc length sequence. The sampling coordinates are obtained by interpolating the distance between the start and end points 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.

[0023] The sampled coordinates are mapped to the grid of the semantically labeled raster map, and the passage determination is performed on the path nodes according to the semantic category of the grid and the voxel occupancy probability.

[0024] Connect all passable nodes in ascending order of their sampled arc lengths to generate a local reference path.

[0025] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the multi-level obstacle avoidance strategy includes performing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation on all nodes in the local reference path.

[0026] 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.

[0027] After the gait switch 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;

[0028] When static obstacle avoidance is triggered, the local reference path is corrected for obstacle avoidance.

[0029] When dynamic obstacle avoidance is triggered, adjust the quadruped robot's walking speed;

[0030] When the slide recovery is triggered, the target landing point is reverted and a recovery node is inserted;

[0031] When sensor degradation is triggered, switch the quadruped robot's pose propulsion mode.

[0032] As a preferred embodiment of the obstacle avoidance control method for quadruped robots based on path planning described in this invention, the generation of the foot trajectory sequence includes calculating the foot trajectory coordinates for the current passable node and the next passable node on the local reference path of the quadruped robot, combined with the swing height and swing period of the current gait switch, and generating the foot trajectory sequence according to the trajectory update period.

[0033] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the real-time monitoring of foot trajectory tracking error and grid occupancy status, and the monitoring of the trigger frequency of multi-level obstacle avoidance strategies, includes real-time monitoring of foot trajectory tracking error and grid occupancy status, and monitoring of the trigger frequency of multi-level obstacle avoidance strategies, and counting the trigger frequencies, which are respectively recorded as foot trajectory tracking error count, grid occupancy event count, static obstacle avoidance trigger count, dynamic obstacle avoidance trigger count, slip recovery trigger count, and sensor degradation trigger count.

[0034] As a preferred embodiment of the path planning-based obstacle avoidance control method for quadruped robots described in this invention, the step of performing global smooth path replanning, local reference path replanning, and foot trajectory sequence regeneration according to the trigger frequency includes,

[0035] When the grid occupancy event trigger count has the highest proportion among all trigger events, global smooth path replanning is performed, and all trigger counts are cleared.

[0036] 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.

[0037] 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.

[0038] If any one of the foot trajectory tracking error count, slip recovery trigger count, and sensor degradation trigger count has the highest proportion among all trigger events, then the foot trajectory sequence is regenerated and all trigger counts are cleared.

[0039] Secondly, the present invention provides a path planning-based obstacle avoidance control system for a quadruped robot, comprising,

[0040] The map building module is used to acquire environmental point cloud data, image data, and quadruped robot motion speed data. By performing terrain semantic segmentation and semantic pixel back projection on the image data, semantic point cloud data is generated and merged with motion-compensated environmental point cloud data to construct a semantically labeled raster map.

[0041] The smooth path generation module is used to perform global path search on semantically annotated raster maps and generate global smooth paths;

[0042] 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, and to determine the terrain category and grid occupancy status of each path node based on the semantically labeled grid map to generate a local reference path.

[0043] The multi-level obstacle avoidance execution module is used to execute multi-level obstacle avoidance strategies on the local reference path and generate a sequence of foot trajectory.

[0044] The path planning and adjustment module is used to perform foot trajectory tracking through joint drive, and monitor the foot trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the trigger frequency of multi-level obstacle avoidance strategies, and performs global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration according to the trigger frequency.

[0045] The beneficial effects of this invention are as follows: By performing terrain semantic segmentation and semantic pixel back-projection on image data, and fusing motion-compensated environmental point clouds to construct a raster map with semantic annotations, a unified expression of geometric and semantic information is achieved, improving the accuracy of path generation. By implementing real-time obstacle monitoring, speed adjustment, and gait switching based on terrain category on local reference paths, and executing multi-level obstacle avoidance strategies, the quadruped robot achieves adaptive movement in complex terrain, enhancing the continuity and safety of its movement. Attached Figure Description

[0046] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart of a path planning-based obstacle avoidance control method for quadruped robots.

[0048] Figure 2 This is a schematic diagram of a path planning-based obstacle avoidance control system for a quadruped robot.

[0049] Figure 3 A flowchart for generating a global smooth path.

[0050] Figure 4 A flowchart generated for a local reference path. Detailed Implementation

[0051] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0052] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0053] Secondly, the term "one embodiment" or "embodiment" as used 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 different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0054] Reference Figures 1-4 This is one embodiment of the present invention, which provides a path planning-based obstacle avoidance control method for a quadruped robot, comprising the following steps:

[0055] S1. Acquire 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. Merge the semantic point cloud data with the motion-compensated environmental point cloud data to construct a semantically labeled raster map.

[0056] Furthermore, a LiDAR, an RGB-D sensor (a composite camera device capable of simultaneously acquiring color and depth images), and a velocity sensor are installed on the top, front, and main drive shaft of the quadruped robot to acquire environmental point cloud data, image data, and quadruped robot motion speed data. The specific steps are as follows:

[0057] By mounting the lidar on a rigid platform on the top of the quadruped robot, 360-degree unobstructed environmental point cloud data can be collected. At the same time, it can be connected to the trigger pulse signal output by an external clock to obtain drift-free environmental point cloud data.

[0058] The RGB-D sensor is mounted in the center of the front of the device to obtain a wide forward field of view, and color images and corresponding depth images are acquired synchronously through the same external clock.

[0059] A speed sensor is installed on the robot's main drive shaft. The movement speed data of the quadruped robot is collected in real time through the leg joint encoder, including the angular velocity of each joint of each leg, namely the angular velocity of the hip joint rolling axis, the hip joint pitch axis, and the knee joint pitch axis, which are read by the leg joint encoder, 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 speed sensor data with timestamps.

[0060] The external clock refers to the unified trigger pulse signal output by a dedicated clock distributor (such as a GPS synchronization clock), which is connected to the acquisition interface of each sensor via a cable to ensure time synchronization and unified timestamps for data acquisition from multiple sensors.

[0061] Furthermore, a semantic segmentation network model (such as DeepLabv3) is used to perform terrain semantic segmentation on the image data. Each pixel in the color image is classified into four semantic labels: ground, low obstacles, high obstacles, and traversable obstacles. The semantic labels are then back-projected onto the depth image to generate semantic point cloud data. Simultaneously, motion compensation is performed on the environmental point cloud data based on the quadruped robot's motion speed data. The semantic point cloud data and the motion-compensated environmental point cloud data are then merged and input into a 3D octree grid to construct a semantically labeled raster map. The specific steps are as follows:

[0062] Using multiple sets of checkerboard images and Zhang Zhengyou's calibration method, checkerboard images of known dimensions were captured. These images were obtained from different angles and positions to ensure rich and clear corner information. Image processing algorithms were used to extract the corners of each checkerboard image, obtaining their pixel coordinates. Optimized camera calibration algorithms, such as least squares, were employed to calculate the known physical dimensions of the checkerboard and the pixel coordinates of the corners, yielding the intrinsic parameters of the RGB-D sensor, including horizontal focal length, vertical focal length, and principal point coordinates. Simultaneously, image distortion parameters were calculated. This includes radial and tangential distortion of the camera lens. Distortion parameters cause the pixel coordinates of the image to shift, affecting the image's geometry. Based on the intrinsic parameters and distortion parameters of the RGB-D sensor, a remapping function is used to map the original image pixel coordinates to the corrected coordinates, generating a pixel correction mapping table. The pixel correction mapping table is then used to perform pixel-by-pixel mapping calculations on the color and depth images acquired by the RGB-D sensor, converting the original coordinates of each pixel into corrected coordinates to eliminate lens distortion and achieve pixel-level alignment between the color and depth images, resulting in corrected color and depth images.

[0063] The corrected color image is input into the pre-trained DeepLabv3 model. Based on a ResNet-101 backbone and a dilated spatial pyramid pooling decoder, the DeepLabv3 model uses deep learning methods to combine local image features with global contextual information for pixel-level classification. Specifically, the DeepLabv3 model uses a convolutional neural network to extract features from each pixel and calculate the class probability for each pixel. During pixel-level classification, the DeepLabv3 model expands the receptive field through dilated convolution operations, enabling it to simultaneously capture local details and global contextual information, thereby improving its ability to classify complex scenes. Finally, the DeepLabv3 model assigns four semantic labels to each pixel of the color image: ground, low obstacles, high obstacles, and traversable obstacles, and generates a semantic label map with the same resolution as the corrected color image.

[0064] The ground refers to obstacles or terrain surfaces, and the area where the average local height difference relative to the current support plane of the quadruped robot does not exceed one-twentieth of the length of the quadruped robot's legs. It is generally flat or slightly uneven terrain, suitable for walking directly without having to avoid obstacles or make special adjustments to the gait.

[0065] Low obstacles refer to obstacles or terrain features whose average local height difference with respect to the current support plane of the quadruped robot is greater than one-twentieth but not more than one-fifth of the length of the quadruped robot's legs, such as small stones or low steps.

[0066] High obstacles refer to obstacles or terrain surfaces whose local average height difference relative to the current support plane of the quadruped robot is greater than one-fifth but no more than one-third of the length of the quadruped robot's legs, such as large rocks or high steps.

[0067] Crossable obstacles refer to obstacles or terrain surfaces whose local average height difference relative to the current support plane of the quadruped robot is greater than one-third but not more than three-quarters of the length of the quadruped robot's legs, while the width of the obstacle does not exceed the maximum horizontal distance that the quadruped robot can reach during jumping. Examples include low fences and wide platform edges.

[0068] The current local mean of the support plane refers to the average height value obtained by performing local smoothing fitting calculations on the support plane with the current standing support leg landing point of the quadruped robot as a reference benchmark. 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.

[0069] The semantic label map is traversed pixel by pixel. When the semantic label of a colored image pixel in the semantic label map belongs to any of the categories of ground, low obstacle, high obstacle and traversable obstacle, the depth value of the corresponding colored image pixel in the corrected depth image is read. The coordinates and depth value of the colored image pixel are marked as semantic pixels, and a semantic pixel set is accumulated.

[0070] In the semantic pixel set, for each semantic pixel, its horizontal and vertical coordinates and corresponding depth value are first read. Then, combined with the horizontal focal length, vertical focal length and principal point coordinates obtained from the RGB-D sensor calibration, and according to the principle of view back projection, the horizontal component of the three-dimensional point is generated by combining the pixel horizontal coordinate, principal point horizontal coordinate, depth value and horizontal focal length. The vertical component of the three-dimensional 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 three-dimensional point of each semantic pixel in the RGB-D sensor coordinate system is obtained. The set of these three-dimensional points forms the semantic point cloud data in the RGB-D sensor coordinate system.

[0071] Using semantic point cloud data in the RGB-D sensor coordinate system as input, the semantic point cloud data in the quadruped robot body coordinate system is mapped from the RGB-D sensor coordinate system to the quadruped robot body coordinate system through the external parameters of the RGB-D sensor and the quadruped robot body coordinate system, namely the rotation matrix and translation vector describing the RGB-D sensor coordinate system relative to the quadruped robot body coordinate system.

[0072] Simultaneously, based on the twelve joint angular velocities collected by the velocity sensor and combined with the geometric parameters of the quadruped robot's legs, the angular velocities of each joint are converted into corresponding foot end linear velocities. Then, the supporting feet are identified, and their foot end linear velocities are extracted. The arithmetic mean of the foot end linear velocities in the forward direction of the robot body is used to obtain the forward linear velocity, and the arithmetic mean of the lateral linear velocity in the lateral direction of the robot body is used to obtain the yaw rate. Then, the difference in forward linear velocities between the two supporting feet is combined with the width of the robot body to obtain the yaw rate. To eliminate point cloud distortion, motion compensation is performed on each frame of environmental point cloud data based on uniform motion. The maximum forward velocity and maximum yaw rate of the quadruped robot on an unobstructed flat ground are used as the upper limit of motion compensation. Specifically, it is assumed that the maximum forward velocity and the maximum yaw rate remain constant within the laser frame sampling period. The pose increment is calculated by numerical integration and applied to the coordinates of each point in the original environmental point cloud data to eliminate the distortion caused by the quadruped robot's motion, thereby obtaining motion-compensated environmental point cloud data.

[0073] The maximum safe forward speed refers to the highest linear velocity at which a quadruped robot can maintain its center of mass within the supporting polygon and ensure stable foot contact force distribution while moving on a flat surface. When the forward speed exceeds the maximum safe forward speed, the quadruped robot may experience excessive center of mass shift, foot slippage, or joint overload during the transition between tri-legged and quad-legged support phases, leading to instability. The upper limit of the maximum safe forward speed is determined by factors such as the quadruped robot's center of gravity height, leg length and leg driving capability, as well as the coefficient of friction between the foot and the ground and the dynamic load distribution during phase transitions. It is typically determined through dynamic simulations and ground experiments comparing stability margins and foot force distribution at different speeds.

[0074] The maximum safe yaw rate refers to the highest angular velocity at which a quadruped robot can ensure that the contact force of its feet does not exceed the friction limit and that its center of mass projection does not deviate from the supporting polygon when turning laterally or rotating in place. When the yaw rate exceeds the maximum safe yaw rate, the quadruped robot's legs may slip due to insufficient friction during turning, or the supporting foot assembly may become unbalanced due to turning inertia, leading to 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 needs to be determined through turning dynamic tests and friction limit tests.

[0075] The laser frame sampling period is determined based on the rated scanning frequency of the lidar and the latency test of the end point cloud processing. It ensures that at least one to five point cloud samples are collected within each laser frame sampling period, while reserving a calculation buffer of no less than 10 milliseconds to achieve a balance between real-time performance and processing load.

[0076] Motion-compensated environmental point cloud data and semantic point cloud data in the quadruped robot's body coordinate system are inserted into a 3D octree grid. For each voxel, its log-probability is updated using the observed hit log-probability and observed loss log-probability, and the voxel occupancy probability is calculated by combining the prior log-probability. Then, the occupancy threshold is adaptively calculated through calibration experiments. Specifically, the calibration experiments first record the true state of each voxel in a scene containing rigid planes or pillars (true occupied areas) and unobstructed areas (true idle areas). Then, multiple frames of environmental point cloud data are continuously acquired, and the ratio of the number of times each truly occupied voxel is judged as occupied to its total number of detections is defined as the hit probability. The ratio of the number of times each truly idle voxel is incorrectly judged as idle to its total number of detections is defined as the loss probability. By combining the complementary terms of the hit probability and the loss probability and normalizing the combination result, the occupancy threshold is calculated, expressed as:

[0077] ;

[0078] in, Indicates the threshold occupied. Indicates the probability of hitting. Indicates the probability of loss;

[0079] The idle threshold is calculated by complementing the occupied threshold.

[0080] 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 marking the voxel occupancy status, the stored semantic labels are counted within each voxel, and the one with the highest count is used as the semantic label of the voxel. Finally, a semantically labeled raster map containing occupancy status and semantic labels for ground, low obstacles, high obstacles, and traversable obstacles is constructed.

[0081] It should be noted that the DeepLabv3 model in this method adopts the following training process: 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. A hollow spatial pyramid pooling module and an upsampling decoder are connected at the top of ResNet-101 to balance the global receptive field and boundary detail restoration. The DeepLabv3 model has four input channels, corresponding to the three-channel color image output by the RGB-D sensor and the one-channel depth image after linear normalization (in the example, the original depth value is linearly mapped from 0–5 meters to the 0–1 range). The training data comes from a 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 semantic labels: pixel-level ground, low obstacles, high obstacles, and traversable obstacles. The amount of training data ensures both scene diversity and training performance. To improve training efficiency and accelerate DeepLabv3 model convergence, 100 rounds of pre-training were performed on the Cityscapes public dataset. During pre-training, stochastic gradient descent was used to optimize the DeepLabv3 model parameters, and all training samples were traversed in each round. Then, a stochastic gradient descent optimizer with a learning rate of 0.001, momentum of 0.9, and weight decay of 0.0001 was used to fine-tune the self-built semantic segmentation dataset for 150 rounds. The fine-tuning batch size was set to 4 to adapt to mainstream GPU memory. The image enhancement strategies used during training included random horizontal flipping of the input image (flipping probability of 50%), random scaling (e.g., enlarging or shrinking to 0.5 to 2 times the original size), and random cropping to a fixed size (e.g., 512×512 pixels) to cover different shooting distances and viewpoints. The training objective was to minimize pixel-level cross-entropy loss, and finally achieve an average intersection-union ratio of 85% on the validation set.

[0082] The learning rate, in this example, ranges from 0.0001 to 0.01 and is determined based on a trade-off between convergence speed and gradient stability. Specifically, grid search was performed with learning rates of 0.0001, 0.001, 0.005, and 0.01 to monitor the rate of decrease of pixel-level cross-entropy loss and the change in gradient norm on the validation set. The results show that when the learning rate is between 0.0001 and 0.01, the loss curve does not exhibit significant oscillations and can converge to a stable level within a finite number of training rounds. When the learning rate is below 0.0001, convergence is too slow, and when it is above 0.01, training is prone to divergence.

[0083] Momentum, with an example value range of 0.8 to 0.95, is determined based on the requirements for oscillation suppression and convergence acceleration. Specifically, by comparing momentum with four candidate values ​​of 0.8, 0.9, 0.95, and 0.99, and observing the oscillation amplitude and iterative convergence speed during the training process, the optimal range of 0.8 to 0.95 was determined. Below 0.8, the convergence speed decreases, and above 0.95, gradient oscillations are aggravated.

[0084] Weight decay, with an example value range of 0.00001 to 0.001, is determined based on preventing overfitting and maintaining the capacity balance of the DeepLabv3 model. Specifically, it is verified by weight decaying at 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 a suitable range, which can effectively suppress overfitting without weakening the feature representation ability of the DeepLabv3 model due to excessive regularization.

[0085] S2. Perform a global path search on the semantically labeled raster map to generate a global smooth path.

[0086] Furthermore, the Theta* algorithm is used to perform path search at arbitrary angles on the semantically labeled raster map, and redundant vertices are removed by local view distance to achieve fast smoothing and generate a globally smooth path. The specific steps are as follows:

[0087] To limit the reachable range of the quadruped robot's feet, voxels within a vertical range of -0.05 meters to 0.15 meters are first reserved for the semantically labeled raster map. The -0.05-meter example is based on the reservation of slight ground depressions, and the 0.15-meter example is based on the maximum obstacle height requirement of the quadruped robot. The reserved voxels are then projected onto a horizontal plane, and a two-dimensional mesh is constructed with an example mesh spacing of 0.05 meters. The mesh spacing shown is derived from the matching requirements of the RGB-D sensor point cloud resolution and foot positioning accuracy.

[0088] For each 2D grid, a passage determination is performed. Specifically, for each grid, the semantic label and voxel occupancy probability are determined simultaneously. When the semantic label of the voxel corresponding to the grid is ground or a crossable obstacle and the voxel occupancy probability is lower than the occupancy threshold, the grid is marked as free. When the semantic label is a low obstacle or a 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.

[0089] After marking the 2D grid, the Theta* algorithm is used to perform path search at any angle for all grids marked as free or unknown. First, the starting grid is added to the open list and its cumulative cost is set to zero. Then, the following operations are repeated until the target grid is visited or the open list is empty. The grid with the smallest sum of cumulative cost and heuristic cost is selected from the open list as the current grid. The reachability of the eight adjacent grids of the current grid is checked one by one. First, it is determined whether the straight line connection between the current grid and the adjacent grid passes through any grid marked as obstruction. When the straight line connection does not pass through the obstruction grid, the sum of the cumulative cost of the current grid and the straight line distance between the center points of the two grids in the horizontal plane is used as the new migration cost of the adjacent grid. Otherwise, the sum of the cumulative cost of the current grid and the standard distance of the eight-connected (referring to the adjacency in the up, down, left, right and four diagonal directions) is used as the migration cost. Then, the horizontal straight line distance from the adjacent grid to the target grid is used as the heuristic cost. If the open list does not contain the adjacent grid or the new migration cost is less than the original cumulative cost of the adjacent grid, the cumulative cost and parent grid pointer of the adjacent grid are updated, and the adjacent grid is added to the open list.

[0090] After the target grid is visited, the initial path is formed by traversing back from the target grid to the starting grid along the parent grid pointer. In order to reduce the number of inflection points in the path, the initial path is sparsified according to the visibility rule of adjacent three grids. Specifically, for any three consecutive grids in the initial path, if the grids that pass through the first and last grids along the straight line are not marked as obstacles, the middle grid is deleted. After traversing the complete path, a global smooth path with the fewest inflection points and the smoothest curve is obtained.

[0091] It should be noted that Theta* algorithm is an arbitrary angle path search algorithm based on A* algorithm. Its main difference from A* lies in its use of line-of-sight judgment to assess path connectivity when expanding nodes, thereby generating paths with fewer vertices and closer to straight lines. Specifically, Theta* algorithm maintains the same open and closed list structure as A* algorithm, but when expanding each adjacent grid of the current grid, it first checks the straight-line connectivity between the parent grid and adjacent grids. If all intermediate grids between the parent grid and adjacent grids are marked as free or unknown, it directly attempts to relax the adjacent grid through the parent grid, calculating the true Euclidean distance from the parent grid to the adjacent grid and adding the cumulative cost of the parent grid as the new migration cost for the adjacent grid. Otherwise, it reverts to the traditional A* approach, using the adjacent distance from the current grid to the adjacent grid plus the cumulative cost of the current grid as the migration cost. After each migration cost calculation for adjacent grids, the Euclidean distance from the adjacent grid to the target grid is used as a heuristic estimate. The sum of the two values ​​serves as the grid priority. If an adjacent grid already exists in the open list and the new migration cost is smaller, the cumulative cost and parent pointer are updated; otherwise, the adjacent grid is inserted into the open list. Through the dynamic switching between the two sets of rules, "straight-line relaxation" and "eight-connected relaxation," the Theta* algorithm can significantly reduce path vertices while ensuring search integrity. Finally, the initial path formed by backtracking along the parent pointer after the target is visited is the optimized path with vertices connected at any angle.

[0092] S3. When the quadruped robot moves along the global smooth path, it extracts the path nodes on the global smooth path with a fixed step distance, and judges the terrain category and grid occupancy status of each path node according to the semantically labeled grid map to generate a local reference path.

[0093] Furthermore, starting from the starting point of the global smooth path, the straight-line distances between adjacent path points on the horizontal plane are calculated sequentially, and all straight-line distances are added sequentially 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 with an example step size of 0.2 meters as the node sampling interval. The first segment at the starting point of the global smooth path corresponds to a distance of 0.2 meters, and so on, generating a set of equidistant sampling arc lengths. The example step size of 0.2 meters is determined based on the field test results of the optimal gait stability and travel efficiency of the quadruped robot.

[0094] For each sampled arc length, the cumulative position that first reaches or exceeds the sampled arc length is found 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 sampled arc length and the starting cumulative arc length of the path segment. The sampled coordinates are uniformly calculated between the start point and the end point according to the distance ratio between the sampled arc length and the starting cumulative arc length of the path segment, thereby accurately reflecting the equidistant nodes on the global smooth path.

[0095] Furthermore, each sampled coordinate is projected onto the corresponding grid of the semantically labeled raster map, the semantic label and voxel occupancy probability of that grid are read, and the traversability of the path node is determined according to the following rules:

[0096] A node is considered passable when its voxel occupancy probability is below the idle threshold and its semantic label is ground or a crossable obstacle; a node is considered obstructive when its voxel occupancy probability is not below the occupancy threshold or its semantic label is a low obstacle or a high obstacle; all other cases are considered uncertain nodes.

[0097] The sampled coordinates of all nodes determined to be passable are connected sequentially in ascending order of sampling index to generate a local reference path; the sampled coordinates of nodes determined to be uncertain are reserved for subsequent processing of the multi-level obstacle avoidance strategy, and the resulting local reference path is then used for dynamic obstacle avoidance and foot trajectory generation.

[0098] S4. Implement a multi-level obstacle avoidance strategy on the local reference path and generate a foot trajectory sequence.

[0099] Furthermore, the multi-level obstacle avoidance strategy includes performing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation on all nodes in the local reference path, with the specific steps as follows:

[0100] Gait switching is always the first priority in the multi-level obstacle avoidance strategy. Before judging static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation, gait switching is first 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 in the semantically labeled raster map for each landing point. When the semantic label is ground, the fast gait is selected and the sample swing height is set to one-twentieth of the quadruped robot's leg length; when the semantic label is low obstacle, the slow gait is selected and the sample swing height is set to one-fifth of the quadruped robot's leg length; when the semantic label is high obstacle, the crawling gait is selected and the sample swing height is set to one-third of the quadruped robot's leg length; when the semantic label is traversable obstacle, the jumping gait is selected and the sample swing height is set to three-quarters of the quadruped robot's leg length.

[0101] In the example, the swing height of the fast walking gait is taken as one-twentieth of the quadruped robot's leg length, the slow walking gait as one-fifth of the quadruped robot's leg length, the crawling gait as one-third of the quadruped robot's leg length, and the jumping gait as three-quarters of the quadruped robot's leg length. These proportions are determined based on kinematic reachability and static stability margin analysis. Assuming the quadruped robot's leg length is 250 mm, to ensure the foot can traverse a maximum of approximately 40 mm of ground undulation without rubbing, the swing height should be set to one-twentieth of the quadruped robot's leg length. For low obstacles no more than one-fifth of the quadruped robot's leg length, a swing height of one-fifth of the quadruped robot's leg length allows for successful crossing without significant shrinkage of the three-legged support polygon. For high obstacles approximately one-third of the quadruped robot's leg length, a swing height of one-third of the quadruped robot's leg length ensures that the centroid projection still falls within the support area during the single-leg support phase. When the semantic label indicates an obstacle that 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 sufficient safety clearance between the foot and the ground during the swing and landing transitions. All of the above height ratios are obtained through kinematic derivation and stability margin analysis, requiring no additional field testing for verification.

[0102] After completing the gait switching for all landing points, the triggering conditions for 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, obstacle avoidance correction is performed on the local reference path. When dynamic obstacle avoidance is triggered, the quadruped robot's traveling speed 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 quadruped robot's pose propulsion mode is switched. All corrections are completed based on the gait switching results, and a foot trajectory sequence matching the current gait and swing height is uniformly generated before the end of the current obstacle avoidance execution cycle.

[0103] As the quadruped robot travels along the local reference path, with an example obstacle avoidance execution cycle of 10 milliseconds as the time interval, the voxel occupancy probability is read sequentially for all grids within an example 1-meter range ahead of the local reference path, including grids containing nodes marked as obstacles and those marked as uncertain. When the voxel occupancy probability of any grid is greater than or equal to the occupancy threshold, static obstacle avoidance is triggered. Starting from the triggered grid, the D*Lite algorithm is run within a local grid region with a radius of 1 meter and the triggered grid as the center to recalculate the avoidance corridor. The D*Lite algorithm maintains a priority queue and incremental cost differential propagation, recalculating the minimum path cost only for grid regions affected by the new obstacle, and dynamically updates the cost value of adjacent grids using heuristic distance, until the shortest avoidance path that bypasses the obstacle and connects to the region boundary is found within the local region. Finally, the corresponding path segment in the local reference path is replaced with the shortest avoidance path to ensure avoidance of the newly added static obstacle.

[0104] The example obstacle avoidance execution cycle is 10 milliseconds, which is determined based on the external clock trigger frequency of the speed sensor. This duration can ensure that the voxel occupancy probability acquisition and obstacle avoidance decision are completed within one cycle to meet the real-time response.

[0105] The example 1-meter range is based on the quadruped robot's step length of approximately 0.2 meters. It determines the minimum reserved distance required to complete obstacle detection and avoidance within four steps, providing sufficient space for path reconnection.

[0106] Within the same obstacle avoidance execution cycle, dynamic obstacle avoidance first sequentially reads the real-time position and velocity of each dynamic obstacle within a 1-meter range ahead of the local reference path, and calculates the safe obstacle avoidance distance by adding the obstacle avoidance safety margin, using the velocity sensor's sensing and decision-making delay time:

[0107] ;

[0108] in, Indicates the safe obstacle avoidance distance. Indicates the speed of dynamic obstacles. This represents the latency between speed sensor perception and decision-making. It is determined by testing the delay from sensor data acquisition to processing, and combining this with the information transmission delay during the decision-making process. Indicates the safety margin for obstacle avoidance;

[0109] When the horizontal distance between the quadruped robot's current position and the dynamic obstacle's position is less than the safe obstacle avoidance distance, a correction vector is superimposed on the preset reference velocity along the direction the quadruped robot is moving away from the obstacle. That is, the direction of the correction vector is opposite to the direction of the line connecting the obstacle. This results in a newly generated obstacle avoidance velocity vector, expressed as:

[0110] ;

[0111] in, This 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. This indicates the current position coordinates of the quadruped robot. This represents the preset reference speed, which is the target linear velocity of the quadruped robot traveling along the path on an unobstructed, flat surface. The example is 0.3 m / s, determined based on path tracking stability tests.

[0112] The newly generated obstacle avoidance velocity vector is applied to the quadruped robot's walking speed, thereby enabling real-time avoidance of dynamic obstacles;

[0113] When both static and dynamic obstacle 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.

[0114] Within the same obstacle avoidance execution cycle, the normal force of the foot is monitored by the foot force sensor of each leg of the quadruped robot. When the normal force of a certain leg is maintained at the typical level of the support phase and the duration is sufficient to confirm that the leg has completed the support phase switch, the leg is determined to be in the support state. If any condition is met at the same time in the support state, the normal force of the foot shows a significant downward trend or the horizontal velocity of the foot suddenly increases, the slip recovery is triggered. After the slip recovery is triggered, the target landing point of the slipping leg is immediately moved back 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 the stability of the movement.

[0115] The horizontal velocity of the foot is calculated based on the difference in horizontal displacement of the foot in the coordinate system of the quadruped robot and the sampling period.

[0116] Within the same obstacle avoidance execution cycle, the output frame rate of the RGB-D sensor is continuously monitored. When the RGB-D sensor continuously loses frames and does not receive new color or depth images, sensor degradation is triggered. After sensor degradation is triggered, the robot speed is estimated and propulsion is achieved solely based on leg odometry and laser data. According to the twelve joint angular velocities collected by the velocity sensor, combined with the quadruped robot's leg geometry parameters and the current gait cycle, the joint angular velocity of each supporting leg in the support phase is converted into the horizontal displacement increment of the foot end. Then, the displacement increments of all supporting legs are arithmetically averaged on the horizontal plane to obtain the estimated linear velocity of the robot body. The position increment is calculated by multiplying the laser frame sampling cycle by the estimated linear velocity of the robot body, and the quadruped robot's pose is advanced along the local reference path. When the RGB-D sensor recovers its frame rate, sensor degradation is exited, and the original multi-level obstacle avoidance strategy and foot trajectory generation process are restored.

[0117] Furthermore, for each swing of the quadruped robot's foot, based on the currently traversable node and the next traversable node on the local reference path, and combined with the swing height corresponding to the current gait and the example swing period, a smooth foot trajectory is generated using a cubic polynomial, and a normalized time ratio is defined as follows:

[0118] ;

[0119] ;

[0120] in, Indicates the normalized time proportion. This represents the oscillation period, with an example value ranging from 0.4 to 0.6 seconds, determined based on gait period and dynamic stability tests. Indicates the time at the foot. The three-dimensional trajectory coordinates, This indicates the coordinates of the current landing point of the swinging foot. This indicates the coordinates of the landing point of the swinging foot. Indicates the swing height under the corresponding gait. Represents a vertically upward unit vector;

[0121] During the trajectory update cycle (Example values ​​range from 0.01 to 0.05 seconds, determined based on travel control frequency and calculated load balancing) Within this range, the time interval will be... Starting from scratch The step size is incremented, and d(t) is calculated and collected sequentially to form a foot trajectory sequence, which is used by the lower-level driver to perform timing tracking.

[0122] S5. Perform foot trajectory tracking through joint drive, and monitor foot trajectory tracking error and grid occupancy status in real time. At the same time, monitor the trigger frequency of multi-level obstacle avoidance strategy, and perform global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration according to the trigger frequency.

[0123] Furthermore, within the same obstacle avoidance execution cycle, the foot trajectory sequence is converted into corresponding joint angle commands through inverse kinematics and sent sequentially to the joint actuators. Simultaneously, using real-time read joint angles and quadruped robot leg geometry parameters (such as thigh, calf, and knee lengths), the actual foot position of each swinging foot is calculated via inverse kinematics. Specifically, based on known joint angles, inverse kinematics calculates the transformation matrix of each joint's relative position, accumulating the positions of these joints layer by layer in the quadruped robot's body coordinate system to ultimately obtain the actual foot position of each swinging foot. The actual foot position of each swinging foot is then combined with the desired foot position (the desired foot position is the position of the foot at time...). The three-dimensional trajectory coordinates are used to calculate the foot trajectory tracking error, which is expressed as:

[0124] ;

[0125] in, This indicates the foot trajectory tracking error. This indicates the actual foot position of each swinging foot;

[0126] When the foot trajectory tracking error exceeds the error threshold, the foot trajectory tracking error count is incremented once. If the foot trajectory tracking error is less than or equal to the error threshold, no count is triggered. The error threshold, in this example, ranges from 0.01 to 0.05 meters, with a lower limit of 0.01 meters. This is based on the fact that typical foot positioning accuracy is approximately ±0.005 meters, and the minimum increment of control commands is approximately 0.003 to 0.007 meters. To avoid unnecessary error counting caused by measurement noise and minor command jitter, the lower limit of the error threshold is set to twice the noise level, effectively filtering out these interference factors and ensuring that error counting is triggered only when the actual deviation of the foot trajectory is large. The upper limit is used to ensure that the foot deviation does not exceed 1 / 4 of the stride distance, which can fully respond to trajectory deviation without frequently triggering replanning due to minor jitter.

[0127] Simultaneously, within the same obstacle avoidance execution cycle, the voxel occupancy probability of the grid where the next foot landing point is located is read for the local reference path. When the voxel occupancy probability is greater than or equal to the occupancy threshold, the grid occupancy event trigger count is incremented once.

[0128] Furthermore, the number of events triggered by multi-level obstacle avoidance strategies is monitored in parallel, including static obstacle avoidance trigger count, dynamic obstacle avoidance trigger count, slip recovery trigger count, and sensor degradation trigger count. All of these counts are accumulated when they are triggered in their respective processes.

[0129] After the same obstacle avoidance execution cycle ends, the proportion of each type of trigger count in all trigger events is evaluated sequentially to determine whether to update the path or trajectory. After any replanning or regeneration is completed, all trigger counts are reset to zero, according to the following rules:

[0130] If the grid occupancy event trigger count has the highest proportion among all trigger events, then perform global smooth path replanning and clear all trigger counts;

[0131] Otherwise, if the static obstacle avoidance trigger count accounts for the highest proportion of all trigger events, then global smooth path replanning is performed and all trigger counts are cleared.

[0132] Otherwise, if the dynamic obstacle avoidance trigger count accounts for the highest proportion of all trigger events, then local reference path replanning is performed and all trigger counts are cleared.

[0133] Otherwise, if any of the foot trajectory tracking error count, slip recovery trigger count, or sensor degradation trigger count has the highest proportion among all trigger events, then the foot trajectory sequence will be regenerated and all trigger counts will be cleared.

[0134] After any replanning or regeneration is completed, all trigger counts are reset to zero, and the original content is replaced with a newly generated global smooth path, local reference path, or foot trajectory sequence. The obstacle avoidance execution cycle enters the next execution loop, and obstacle judgment and foot trajectory tracking continue based on the updated path or trajectory.

[0135] This embodiment also provides a path planning-based obstacle avoidance control system for a quadruped robot, including:

[0136] The map building module is used to acquire environmental point cloud data, image data, and quadruped robot motion speed data. By performing terrain semantic segmentation and semantic pixel back projection on the image data, semantic point cloud data is generated and merged with motion-compensated environmental point cloud data to construct a semantically labeled raster map.

[0137] The smooth path generation module is used to perform global path search on semantically annotated raster maps and generate global smooth paths;

[0138] 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, and to determine the terrain category and grid occupancy status of each path node based on the semantically labeled grid map to generate a local reference path.

[0139] The multi-level obstacle avoidance execution module is used to execute multi-level obstacle avoidance strategies on the local reference path and generate a sequence of foot trajectory.

[0140] The path planning and adjustment module is used to perform foot trajectory tracking through joint drive, and monitor the foot trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the trigger frequency of multi-level obstacle avoidance strategies, and performs global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration according to the trigger frequency.

[0141] This embodiment also provides a computer device applicable to the obstacle avoidance control method for a quadruped robot 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 the computer-executable instructions to implement the obstacle avoidance control method for a quadruped robot based on path planning as proposed in the above embodiment.

[0142] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices 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 the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0143] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the obstacle avoidance control method for a quadruped robot based on path planning as proposed in the above embodiments. 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 Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0144] In summary, this invention improves the accuracy of path generation by performing terrain semantic segmentation and semantic pixel back-projection on image data, and fusing motion-compensated environmental point clouds to construct a semantically labeled raster map. This achieves a unified expression of geometric and semantic information. Furthermore, by implementing real-time obstacle monitoring, speed adjustment, and terrain-category-based gait switching on local reference paths, and executing multi-level obstacle avoidance strategies, the invention enables the quadruped robot to adapt to complex terrain, enhancing its mobility and safety.

[0145] 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 it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A path planning-based obstacle avoidance control method for a quadruped robot, characterized in that: include, The system acquires 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 motion-compensated environmental point cloud data to construct a semantically labeled raster map. Perform a global path search on the semantically annotated raster map to generate a globally smooth path; When the quadruped robot moves along the global smooth path, it extracts the path nodes on the global smooth path with a fixed step distance, and determines the terrain category and grid occupancy status at each path node based on the semantically labeled grid map to generate a local reference path. A multi-level obstacle avoidance strategy is implemented on the local reference path, and a sequence of foot trajectory is generated; Foot trajectory tracking is performed through joint drive, and foot trajectory tracking error and grid occupancy status are monitored in real time. At the same time, the trigger frequency of 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 trigger frequency. 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 image to generate semantic point cloud data. The construction of the semantically labeled raster map includes performing motion compensation on the environmental point cloud data based on the quadruped robot's motion speed data, and merging the semantic point cloud data with the motion-compensated environmental point cloud data to construct the semantically labeled raster map. The step of merging semantic point cloud data with motion-compensated environmental point cloud data includes inserting the motion-compensated environmental point cloud data and the semantic point cloud data in the quadruped robot body coordinate system into a three-dimensional octree grid, updating the log probability of each voxel with the observation hit log probability and the observation loss log probability, and calculating the voxel occupancy probability in combination with the prior log probability. The occupancy threshold is adaptively calculated through calibration experiments. Specifically, the calibration experiments record the real state of each voxel in a scene that includes real occupied areas and real idle areas. Multiple frames of environmental point cloud data are continuously acquired, and the ratio of the number of times each truly occupied voxel is identified as occupied to its total number of detections is defined as the hit probability. The ratio of the number of times each truly idle voxel is incorrectly identified as idle to its total number of detections is defined as the loss probability. By combining the complementary terms of the hit probability and the loss probability, and normalizing the combination result, the occupancy threshold is calculated, expressed as: ； in, Indicates the threshold occupied. Indicates the probability of a hit. Indicates the probability of loss; The idle threshold is calculated by complementing 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; otherwise, it remains unknown. After the voxel occupancy status is labeled, the stored semantic labels within each voxel are counted, and the one with the highest count is used as the semantic label of the voxel. Finally, a semantic labeled grid map containing occupancy status and semantic labels for ground, low obstacles, high obstacles and traversable obstacles is constructed.

2. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 1, characterized in that: The steps for generating a globally smooth path are as follows. The semantically labeled raster map is filtered into voxels according to the height of the quadruped robot's foot, and the filtered voxels are projected onto a horizontal plane to construct a two-dimensional mesh. Perform passage determination on the 2D mesh based on semantic labels and voxel occupancy probability; Based on the passage determination results, a path search is performed on a two-dimensional grid, and a smoothing process is performed by eliminating redundant vertices through local sight distance to generate a globally smooth path.

3. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 2, characterized in that: The steps for generating a local reference path are as follows: On a globally smooth path, the straight-line distances between adjacent path nodes on the horizontal plane are accumulated sequentially to generate a cumulative arc length sequence. Segment the cumulative arc length sequence to generate equidistant sampling arc lengths; For the sampling arc length, the first path segment with a length greater than or equal to the sampling arc length is found in the cumulative arc length sequence. The sampling coordinates are obtained by interpolating the distance between the start and end points 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. The sampled coordinates are mapped to the grid of the semantically labeled raster map, and the passage determination is performed on the path nodes according to the semantic category of the grid and the voxel occupancy probability. Connect all passable nodes in ascending order of their sampled arc lengths to generate a local reference path.

4. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 3, characterized in that: The multi-level obstacle avoidance strategy includes performing gait switching, static obstacle avoidance, dynamic obstacle avoidance, slip recovery, and sensor degradation on 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 the gait switch 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 walking speed; When the slide recovery is triggered, the target landing point is reverted and a recovery node is inserted; When sensor degradation is triggered, switch the quadruped robot's pose propulsion mode.

5. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 4, characterized in that: The process of generating the foot trajectory sequence involves calculating the foot trajectory coordinates for the current passable node and the next passable node on the local reference path of the quadruped robot, taking into account the swing height and swing period of the current gait switch, and generating the foot trajectory sequence according to the trajectory update cycle.

6. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 5, characterized in that: The real-time monitoring of foot trajectory tracking error and grid occupancy status, and the monitoring of the trigger frequency of multi-level obstacle avoidance strategies, include real-time monitoring of foot trajectory tracking error and grid occupancy status, and monitoring of the trigger frequency of multi-level obstacle avoidance strategies, and counting the trigger frequencies, which are respectively recorded as foot trajectory tracking error count, grid occupancy event count, static obstacle avoidance trigger count, dynamic obstacle avoidance trigger count, slip recovery trigger count, and sensor degradation trigger count.

7. The obstacle avoidance control method for a quadruped robot based on path planning as described in claim 6, characterized in that: The process of performing global smooth path replanning, local reference path replanning, and foot trajectory sequence regeneration based on trigger frequency includes... When the grid occupancy event trigger count has the highest proportion among 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. If any one of the foot trajectory tracking error count, slip recovery trigger count, and sensor degradation trigger count has the highest proportion among all trigger events, then the foot trajectory sequence is regenerated and all trigger counts are cleared.

8. A path planning-based obstacle avoidance control system for a quadruped robot, based on the path planning-based obstacle avoidance control method for a quadruped robot according to any one of claims 1 to 7, characterized in that: include, The map building module is used to acquire environmental point cloud data, image data, and quadruped robot motion speed data. By performing terrain semantic segmentation and semantic pixel back projection on the image data, semantic point cloud data is generated and merged with motion-compensated environmental point cloud data to construct a semantically labeled raster map. The smooth path generation module is used to perform global path search on semantically annotated raster maps and generate global smooth paths; 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, and to determine the terrain category and grid occupancy status of each path node based on the semantically labeled grid map to generate a local reference path. The multi-level obstacle avoidance execution module is used to execute multi-level obstacle avoidance strategies on the local reference path and generate a sequence of foot trajectory. The path planning and adjustment module is used to perform foot trajectory tracking through joint drive, and monitor the foot trajectory tracking error and grid occupancy status in real time. At the same time, it monitors the trigger frequency of multi-level obstacle avoidance strategies, and performs global smooth path replanning, local reference path replanning and foot trajectory sequence regeneration according to the trigger frequency.

Images