Robot gait adaptive adjustment method and system combined with terrain feature recognition

Abstract

The application discloses a robot gait adaptive adjustment method and system combined with terrain feature recognition, relates to the technical field of robot gait, and comprises the following steps: collecting terrain images in front of a robot travel path for deblurring processing, and extracting ground feature data through a lightweight neural network; matching the ground feature data with a preset gait control parameter database to obtain a gait control parameter set, and generating a control signal through a proportional-differential control method according to the gait control parameter set; correcting the control signal based on a reinforcement learning algorithm, and generating a target control instruction for driving a joint actuator; in the process of executing the target control instruction, state feedback data are collected in real time, and the feedback data are compared with the target control instruction currently executed and optimized to realize feedback adjustment; the problems of insufficient terrain recognition accuracy caused by image blurring, response lag caused by decoupling of perception and control, and lack of adaptive ability are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of robot gait technology, and more particularly, to a robot gait adaptive adjustment method and system combined with terrain feature recognition. BACKGROUND

[0002] With the continuous expansion of the application of robot technology, especially mobile robots and humanoid robots in complex environments, the requirements for the environmental adaptability of robots are increasing. In complex, unstructured terrains such as gravel, slopes, and mud, traditional fixed gait planning methods often fail to guarantee gait stability and motion efficiency due to the lack of environmental perception and adaptation mechanisms, which can easily cause problems such as falling, slipping, or excessive foot impact, severely limiting the practicality and reliability of robots.

[0003] Some existing methods attempt to obtain ground contact feedback through pre-installed sensors or adjust gait control parameters with the help of a rule base to improve adaptability. However, these methods generally have the following technical limitations:

[0004] In real environments, due to the prevalence of imaging problems such as motion blur and jitter blur, the quality of image feature extraction is reduced, affecting the accuracy and robustness of terrain recognition;

[0005] And existing control systems usually decouple perception and control, failing to form a closed-loop feedback control system based on ground feature driving, resulting in insufficient control strategy targeting, low response frequency, poor adaptability to sudden terrain changes, and difficulty in achieving efficient adjustment.

[0006] To address the above problems, the present application provides a solution. SUMMARY

[0007] To overcome the above-mentioned defects of the prior art, embodiments of the present application provide a robot gait adaptive adjustment method and system combined with terrain feature recognition, which integrates image deblurring processing, lightweight neural network terrain feature extraction, reinforcement learning driven control correction, and fuzzy logic fusion regulation mechanism to build a perception-control-feedback integrated closed-loop gait adjustment system, effectively solving the problems of insufficient terrain recognition accuracy due to image blur, response lag and lack of adaptability due to decoupling of perception and control.

[0008] To achieve the above-mentioned purposes, the present application provides the following technical solutions:

[0009] In a first aspect, the application provides a robot gait adaptive adjustment method combined with terrain feature recognition, which comprises: collecting terrain images in front of a robot travel path and performing deblurring processing, extracting ground feature data from the deblurred images through a lightweight neural network; matching the ground feature data with a preset gait control parameter database to obtain a gait control parameter set, and generating a control signal through a proportional-differential control method according to the gait control parameter set; correcting the control signal based on a reinforcement learning algorithm, and generating a target control instruction for driving a joint actuator; in the process of executing the target control instruction, collecting state feedback data in real time, and comparing and optimizing the feedback data with the currently executed target control instruction to realize feedback adjustment.

[0010] In one embodiment, the terrain images in front of the robot travel path are collected and deblurred, specifically: each frame of image in the terrain image is converted into a single-channel gray image, and a definition index is calculated; if the definition index is lower than a preset blur threshold, it is determined that the current image is a blurred image, and deblurring processing is performed; edge detection processing is performed on the blurred image to obtain an edge intensity map and a direction map; the direction map is divided into several direction intervals, the number of pixels in each direction is counted, and an edge direction distribution histogram is constructed; if the proportion of main direction pixels exceeds a preset direction concentration threshold, it is determined as linear motion blur; otherwise, it is non-linear jitter blur; deblurring operation is performed on the linear motion blur and the non-linear jitter blur.

[0011] In one embodiment, the deblurring operation is performed on the linear motion blur and the non-linear jitter blur, specifically: based on the linear motion blur, a two-dimensional Fourier transform is performed on the blurred image to obtain a blur direction and a distance, and the blur length is inversely calculated according to the distance; a linear motion blur kernel is constructed as a linear blur point spread function according to the blur direction and the blur length; according to the point spread function, a non-blind deconvolution algorithm is used to restore the blurred image to obtain a deblurred image; based on the non-linear jitter blur, the blurred image is set as an observed image, an estimated clear image and a non-linear blur kernel are initialized, and an optimization objective function is constructed; based on the optimization objective function, an alternating minimization strategy is used to optimize the image and the blur kernel until a preset iteration number is reached, and an optimized non-linear blur kernel is obtained; the non-linear blur kernel is used as a blur kernel for non-blind deconvolution to obtain a clear image.

[0012] In one of the embodiments, the ground feature data is extracted from the blurred image by a lightweight neural network, specifically: the blurred image is preprocessed to obtain a standardized image; the standardized image is taken as an input item to construct a lightweight convolutional neural network model for preliminary identification of the ground feature, and an output ground feature classification map is obtained; the confidence in the ground feature classification map is obtained, and based on a preset confidence threshold, an image region with low confidence in the ground feature classification map is extracted to construct a corresponding candidate region mask map; based on the candidate region mask map, a corresponding high-resolution image sub-block is extracted from the terrain image, and the sub-block is taken as an input to be transmitted to a preset fine identification model; the output of the fine identification model is fused with the ground feature classification map to generate a ground feature semantic segmentation map, which is converted into structured ground feature data.

[0013] In one of the embodiments, the ground feature data is matched with a preset gait control parameter database to obtain a gait control parameter set, and a control signal is generated by a proportional-differential control method based on the gait control parameter set, specifically: the ground feature data is linearly normalized to obtain a ground feature vector; the Euclidean distance between the current ground feature vector and the ground feature vector in the database is calculated based on a K-nearest neighbor algorithm, and the gait control parameter set corresponding to the minimum distance is selected as the output result; the state information of the current robot is obtained, and a control correction term is calculated based on the proportional-differential control method according to the error between the matched control parameter set and the state information; the control signal is generated based on the control correction term.

[0014] In one of the embodiments, the control signal is corrected based on a reinforcement learning algorithm, and a target control instruction for driving the joint actuator is generated, specifically: a reinforcement learning model is constructed, and the ground feature vector, the state information, and the control signal are taken as input states; an adjustment factor of the control signal is taken as an output action; a reward function is constructed, and the reward function includes a vibration spectrum index, a gait symmetry measure, and a contact surface stability index; based on the current input state, the adjustment factor of the control signal is output by the reinforcement learning model with the goal of maximizing the reward function; the adjustment factor and the control signal are fused to generate the target control instruction for driving the joint actuator based on a fuzzy logic reasoning method.

[0015] In one of the embodiments, the adjustment factor is fused with the control signal based on a fuzzy logic reasoning method to generate a target control instruction, specifically: taking the adjustment factor of the control signal output by the reinforcement learning model as an input variable of a fuzzy controller, constructing a corresponding membership function, and mapping the continuous numerical value of each adjustment factor to a fuzzy state; based on the fuzzy state, performing fuzzy logic reasoning using a pre-set experience rule base to generate a corresponding fuzzy control output; using the centroid method, converting the fuzzy control output into a numerical correction amount; fusing the numerical correction amount with the control signal to generate a target control instruction, the target control instruction including a target joint angle value, a target joint angular velocity value, a target joint torque value, and a target joint damping value; performing physical constraint verification on the target control instruction, and outputting the verified target control instruction for driving the joint actuator.

[0016] In one of the embodiments, during the execution of the target control instruction, state feedback data is collected in real time, and the feedback data is compared and optimized with the currently executed target control instruction to realize feedback adjustment, specifically: obtaining state feedback data in the process of executing the target control instruction by the joint actuator, the state feedback data including an actual joint angle value, an actual joint angular velocity value, an actual joint torque value, a foot-ground contact force value, and an acceleration value; performing numerical comparison between the state feedback data and the target control instruction to calculate errors respectively; based on the errors and the state feedback data, constructing a quantitative feedback evaluation index, the feedback evaluation index including a trajectory tracking error index, a gait stability index, and a contact impact index; forming a multi-dimensional state vector with the quantitative feedback evaluation index as input data and transmitting it to the reinforcement learning model to obtain an optimized adjustment factor for the next control cycle; using the optimized adjustment factor to generate a target control instruction for the next control cycle to realize closed-loop feedback adjustment.

[0017] In a second aspect, the application provides a robot gait adaptive adjustment method and system combined with terrain feature recognition, the system including:

[0018] An image recognition module for collecting terrain images in front of the robot's travel path and performing deblurring processing, and extracting ground feature data from the deblurring processed images through a lightweight neural network;

[0019] A control signal generation module for matching the ground feature data with a pre-set gait control parameter database to obtain a gait control parameter set, and generating a control signal through a proportional-differential control method according to the gait control parameter set;

[0020] A control signal correction module for correcting the control signal based on a reinforcement learning algorithm, and generating a target control instruction for driving the joint actuator;

[0021] A real-time feedback adjustment module is configured to collect state feedback data in real time during execution of the target control instruction, and compare and optimize the feedback data with the currently executed target control instruction, so as to realize feedback adjustment.

[0022] From the above technical solutions, the embodiments of the present application have the following advantages:

[0023] 1. By introducing a blur detection and classification mechanism in the image collection of the path in front of the robot, and combining the differentiated de-blurring processing strategies for linear and non-linear blur types, the image clarity and structural recognizability are significantly improved, solving the influence of image blur on terrain recognition accuracy during dynamic travel. At the same time, a lightweight convolutional neural network is used for preliminary ground feature extraction, combined with confidence analysis and region mask strategy, to guide the high-precision recognition model to finely perceive the low-confidence area, realizing high-precision recognition of key terrain areas under the condition of limited computing resources. Through multi-scale feature fusion and spatial pyramid pooling, the semantic modeling ability is enhanced, and finally the structured ground feature data is output, providing high-quality and quantifiable environmental perception support for subsequent path planning and gait adaptive control, significantly improving the adaptability and stability of the robot in complex unstructured terrain.

[0024] 2. By constructing a mapping database of ground features and gait control parameters, combining K-nearest neighbor matching and proportional-differential control to generate basic control signals, and introducing a reinforcement learning algorithm to realize adaptive optimization based on environmental feedback, the fine adjustment ability of the control signal is further improved by further integrating fuzzy logic reasoning, and a closed-loop feedback mechanism is constructed to correct the control parameters in real time, thereby forming a "perception-matching-control-optimization-feedback" integrated gait adjustment system. This scheme not only improves the stability and environmental adaptability of the robot in complex terrain, but also significantly enhances the robustness, coordination and energy efficiency of the control system through multi-dimensional reward function evaluation and dynamic strategy updating, with strong generalization ability and engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A robot gait adaptive adjustment method flowchart provided by the embodiments of the present application in combination with terrain feature recognition.

[0026] Figure 2 A robot gait adaptive adjustment system structure diagram provided by the embodiments of the present application in combination with terrain feature recognition.

[0027] Figure 3 An edge direction distribution histogram provided by the embodiments of the present application. DETAILED DESCRIPTION

[0028] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0029] Referring to Figure 1 As shown in the figure, the robot gait adaptive adjustment method combined with terrain feature recognition provided by the present application comprises the following steps:

[0030] S1, collect the terrain image in front of the robot travel path and perform deblurring processing, and extract ground feature data from the blurred image through a lightweight neural network.

[0031] Among them, through the visual sensor installed in the front end of the mobile platform, the visual sensor includes RGB camera, laser radar, depth camera or binocular stereo vision module, according to the robot travel speed dynamic adjustment collection frequency (such as 5~15 Hz) collection mobile platform in front of the path of the terrain image.

[0032] In this embodiment, the terrain image in front of the robot travel path is collected and deblurring processing is performed, specifically:

[0033] Each frame of image in the terrain image is converted into a single-channel gray image, and the definition index is calculated based on the image gradient variance method of Laplacian operator;

[0034] The definition index, the specific calculation formula is as follows:

[0035]

[0036] In the formula, The definition index is The value of the i-th pixel in the Laplacian response image is The pixel mean value is N, and the total number of pixels is N.

[0037] If the definition index is lower than the preset blur threshold, it is determined that the current image is a blurred image, and deblurring processing is performed, otherwise it is directly sent to the subsequent neural network for ground feature extraction;

[0038] The blurred image is subjected to edge detection processing, and the Sobel operator is used for horizontal direction and vertical direction gradient calculation to obtain the edge intensity map and the direction map.

[0039] For example Figure 3The direction map is divided into several direction intervals, the number of pixels in each direction is counted, and an edge direction distribution histogram is constructed.

[0040] If the proportion of the main direction pixels exceeds a preset direction concentration threshold, it is determined that the image has a main direction and is confirmed as linear motion blur; otherwise, it is a nonlinear jitter blur.

[0041] The linear motion blur and the nonlinear jitter blur are deblurred.

[0042] The linear motion blur refers to the uniform stretching of the pixels of an object in the image in a certain fixed direction due to the uniform linear motion of the image acquisition device during exposure, resulting in a trailing phenomenon. The nonlinear jitter blur refers to the image blur caused by high-frequency, irregular fine vibrations of the image acquisition device during exposure, and the motion trajectory does not have obvious directionality, showing nonlinear and non-uniform characteristics.

[0043] Further, the linear motion blur and the nonlinear jitter blur are deblurred, specifically:

[0044] Based on the linear motion blur, a two-dimensional Fourier transform is performed on the blurred image to obtain a blur direction and a distance between periodic grooves in the frequency spectrum, and the blur length is inversely deduced according to the distance, and the blur direction is a main direction angle;

[0045] A linear motion blur kernel is constructed as a linear blur point spread function according to the blur direction and the blur length;

[0046] According to the point spread function, a non-blind deconvolution algorithm is used to restore the blurred image to obtain a blurred image;

[0047] Based on the nonlinear jitter blur, the blurred image is set as an observed image, an estimated clear image and a nonlinear blur kernel are initialized, and an optimization objective function is constructed;

[0048] Based on the optimization objective function, an alternating minimization strategy is used to optimize the image and the blur kernel until a preset iteration number is reached, and a final optimized nonlinear blur kernel is obtained, the optimization including fixing the current blur kernel, optimizing the estimated image to approximate the observed image under the premise of preserving the edge, and fixing the current image estimate result, optimizing the blur kernel to more accurately characterize the blur characteristics in the current image;

[0049] The nonlinear blur kernel is used as a blur kernel for non-blind deconvolution to deblur the blurred image and obtain a clear image.

[0050] The optimization objective function has the following specific calculation formula:

[0051]

[0052] wherein, is an optimization objective function, is an observed image, is an estimated sharp image, is a nonlinear blur kernel, is an image gradient, is a regularization parameter, is a two-norm square, is a one-norm.

[0053] Wherein, the point spread function (PSF) is a function describing that a point light source is diffused into a certain shape pattern in the imaging process in an image acquisition system, used to represent the system blur characteristics, and is the core parameter of image blur modeling.

[0054] It should be noted that the deblurring processing can effectively identify and correct the image blur problem generated by the acquisition robot in the dynamic running process. Through blur detection and classification, different deblurring operations are matched for fine processing, which significantly improves the clarity and structural recognizability of the terrain image, lays a high-quality input foundation for subsequent ground feature extraction, and improves the overall recognition accuracy and robustness of the system.

[0055] Further, the ground feature data is extracted from the blurred image by a lightweight neural network, specifically:

[0056] The blurred image is preprocessed to obtain a standardized image, and the preprocessing includes image size normalization, contrast enhancement, histogram equalization and region clipping, so as to improve the feature extraction accuracy of the subsequent neural network;

[0057] A light-weight convolutional neural network model for preliminary identification of ground features is constructed based on a standardized image as an input item, and a ground feature classification map is output. The model adopts an encoding-decoding structure, and the overall structure includes an input preprocessing layer, a light-weight feature extraction backbone network, a feature enhancement module, a decoding module, and a classification output layer. The feature extraction backbone network is a light-weight convolutional neural network, and preferably adopts a MobileNetV3 or ShuffleNetV2 light-weight CNN structure. Each basic convolutional block in the network includes the following structural units: a depth separable convolution, a batch normalization, a Swish activation function, a channel attention mechanism, and a residual connection. In the backbone network structure, feature maps of different depths extract different scale semantic information, and are fused with the decoding path through a skip connection in the subsequent decoding stage to improve the edge structure and spatial positioning accuracy. The feature enhancement module preferably adopts a spatial pyramid pooling structure to supplement the global context perception ability for large-scale terrain structures. The decoding module gradually up-samples and restores low-resolution semantic feature maps, combines shallow features from the encoding stage to improve the detail restoration quality, and performs fusion processing through a light-weight convolutional module to output a prediction map consistent with the original image size. The output layer is a ground feature classification map per pixel, which adopts a Softmax activation function for multi-class probability output, and is used for subsequent cascade identification calls.

[0058] A confidence in the ground feature classification map is obtained, and based on a preset confidence threshold, an image region with low confidence in the ground feature classification map is extracted, and a corresponding candidate region mask map is constructed.

[0059] The last classification layer of the light-weight neural network usually adopts a Softmax activation function to convert the prediction result of each pixel into a probability distribution belonging to each class, and the maximum class probability value of each pixel is taken as the confidence of the pixel. Pixels with a confidence lower than the preset threshold are marked as image regions with low confidence.

[0060] Based on the candidate region mask map, a corresponding high-resolution image sub-block is extracted from the terrain image, and the sub-block is transmitted as an input to a preset fine identification model.

[0061] The fine identification model is a medium-depth high-precision neural network for further identifying the boundaries and fine-grained categories of ground objects in a local range. Preferably, it adopts a pruned and optimized EfficientNet, ResNet, or Swin-Transformer structure to control the computational complexity while ensuring accuracy.

[0062] The high-precision local identification result output by the fine identification model is fused with the ground feature classification map to generate a ground feature semantic segmentation map that takes into account global consistency and local accuracy.

[0063] The fusion process adopts a region-level updating strategy: the region specified by the mask map is covered by the refined identification model result, and the non-mask region remains unchanged in the prediction of the lightweight convolutional neural network model.

[0064] The semantic segmentation map is converted into structured ground feature data.

[0065] The ground feature data refers to a parameterized data set that can represent the physical form of the ground, the structure state of the ground, and its mechanical influence on the moving body, which is extracted based on the analysis of the visual image by the neural network model, and is mainly used to support the path planning, gait adaptive control, and stability optimization of robots and other moving bodies in natural or semi-structured environments. The ground feature data includes local slope, surface roughness, ground type, and obstacle boundary distance.

[0066] The candidate region mask map is a binary map consistent with the size of the original image, which is used to indicate the low-confidence pixel region in the ground feature classification map output by the lightweight neural network. Its function is to identify the region with high uncertainty in the prediction of the lightweight model as a candidate region for subsequent fine identification by the high-precision model, so as to realize local optimization of resources.

[0067] The high-resolution image sub-block refers to a local image block extracted from the original input image corresponding to the low-confidence region in the candidate region mask map, which retains its original resolution. These sub-blocks are used as inputs for the high-precision identification model to further extract high-order feature information such as ground object boundaries and fine-grained categories in a local range.

[0068] It should be noted that by introducing the dual-stage architecture of lightweight convolutional neural network and region-level fine identification model, high-precision perception of key regions is achieved while ensuring overall inference efficiency, overcoming the drawbacks of traditional single models in resource-constrained environments, such as insufficient accuracy or excessive computational burden. At the same time, the multi-scale feature fusion and spatial pyramid pooling mechanism are used to enhance the semantic modeling ability of complex ground structures, improving the recognition accuracy of edge contours and fine-grained terrain differences. The final output of structured ground feature data has higher semantic interpretability and parameterized expression ability, which can directly support path planning and dynamic control tasks of robot systems, significantly enhancing their environmental adaptability and stability in natural scenes.

[0069] S2, match the ground feature data with the preset gait control parameter database to obtain a set of gait control parameters, and generate a control signal according to the set of gait control parameters through a proportional-differential control method.

[0070] The gait control parameter database refers to a data structure of a pre-constructed gait control parameter set corresponding to a plurality of ground feature mappings, and is used to support fast searching of a gait control strategy set most matched with a current terrain in an actual walking process. The core function is to establish a mapping relationship from a ground feature space to a gait control strategy space, so that the robot can dynamically adjust the gait control mode according to the ground change, thereby improving the stability and environmental adaptability. The gait control parameter database is constructed by selecting representative ground attribute parameters as the constituent elements of the ground feature vector in combination with the target application scene of the robot (such as outdoor, factory, residence, etc.). Commonly include: slope, friction, hardness, unevenness. And combined with the sensing data of multiple different ground types collected in the simulation environment, multiple sets of ground feature vectors are obtained, and for each set of ground feature vectors, the parameters are designed and fine-tuned by expert experience and existing motion control data to generate the corresponding gait control parameter set.

[0071] In this embodiment, the ground feature data is matched with the preset gait control parameter database to obtain a gait control parameter set, and a control signal is generated according to the gait control parameter set by a proportional-differential control method, specifically:

[0072] The ground feature data is linearly normalized to obtain a ground feature vector representing the ground feature;

[0073] The Euclidean distance between the current ground feature vector and the ground feature vector in the database is calculated based on the K-nearest neighbor algorithm, and the gait control parameter set corresponding to the minimum distance is selected as the output result. The gait control parameter set includes stride length, step frequency, leg lifting height, angle trajectory array of each joint, and expected angular velocity array of each joint, expected torque value and expected damping value;

[0074] The state information of the current robot is obtained, including the current position angle and angular velocity of each joint, the attitude angle of the robot body, and the center of gravity position;

[0075] Based on the proportional-differential control method, the control correction term is calculated according to the error between the matched control parameter set and the state information, and the control signal for controlling the joint actuator is generated according to the control correction term. The control signal includes target angle, target angular velocity, target torque and target damping.

[0076] S3, the control signal is corrected based on the reinforcement learning algorithm, and the target control instruction for driving the joint actuator is generated.

[0077] The control signal is corrected by reinforcement learning, which can realize adaptive optimization based on environmental feedback and historical experience. Reinforcement learning can continuously adjust joint control parameters based on multi-dimensional sensor input by constructing a reward function oriented to actual movement performance, so that the control strategy has stronger robustness and generalization ability when facing different terrain disturbances, load changes or nonlinear disturbances. At the same time, the correction mechanism can significantly improve the stability, motion coordination and energy efficiency of the robot walking, thereby enhancing the intelligent level and adaptability of the overall gait control system.

[0078] In the embodiment, the control signal is corrected based on a reinforcement learning algorithm, and a target control instruction for driving the joint actuator is generated, specifically:

[0079] An reinforcement learning model is constructed, and the ground feature vector, state information and control signal are taken as input states.

[0080] The adjustment factors of the target angle, target angular velocity, target torque and target damping in the control signal are taken as output actions.

[0081] A reward function is constructed, and the reward function includes a vibration spectrum index, a gait symmetry measure and a contact surface stability index.

[0082] The adjustment factors of the control signal are outputted by the reinforcement learning model based on the current input state, with the goal of maximizing the reward function.

[0083] The adjustment factors and the control signal are fused based on a fuzzy logic reasoning method to generate a target control instruction for driving the joint actuator.

[0084] The reward function has the following specific calculation formula:

[0085]

[0086] In the formula, is the reward function, is the gait symmetry measure, is the contact surface stability index, is the vibration spectrum index, , , are weight coefficients, respectively.

[0087] The vibration spectrum index refers to a spectrum entropy index of a joint vibration signal of the robot. The higher the entropy value is, the more serious the control jitter is. The angle / acceleration data of the controlled joint is selected using a high-frequency IMU or a high-precision encoder, is sampled in a preset time window to obtain a discrete sequence, a fast Fourier transform (FFT) is performed on the discrete sequence to obtain an amplitude spectrum for obtaining a frequency distribution, the energy of each frequency point on the frequency distribution is normalized to form a probability distribution, and the Shannon entropy formula is applied to the probability distribution to obtain a spectrum entropy, and then the vibration spectrum index is obtained.

[0088] The gait symmetry metric refers to the relative deviation of the key gait parameters (such as stride length, ground contact duration, swing period) of the left / right leg. The key gait parameters of the left leg and the right leg are obtained, and the absolute value of the difference between the key gait parameters of the left leg and the right leg is divided by the sum of the key gait parameters of the left leg and the right leg, and each is added to construct a symmetry loss function , and the gait symmetry metric is obtained by subtracting 1- / 3. The closer to 1 indicates that the gait is more symmetrical.

[0089] The contact surface stability index refers to the actual contact area change rate of the robot foot and the ground (for example, whether the forefoot or the hindfoot is raised). An m*n pressure sensor is arranged on the foot, a pressure matrix between two points is collected, a contact point set greater than a preset pressure threshold in the pressure matrix is obtained, the maximum and minimum contact areas in the support period are taken to obtain the contact area change rate (i.e., the maximum and minimum contact areas divided by the maximum area), and the contact surface stability index is obtained by subtracting 1 from the contact area change rate. The larger the value is, the more stable the contact is.

[0090] It should be noted that the vibration spectrum index, the gait symmetry metric, and the contact surface stability index are introduced to construct the reward function in the reinforcement learning, which can realize multi-dimensional and fine evaluation of the robot gait control effect. Not only the overall stability and energy consumption are captured, but also the fine-grained dynamic performance is identified. The vibration spectrum index reflects the structural resonance and disturbance response, the gait symmetry metric reveals the motion coordination and execution consistency, and the contact surface stability embodies the reliability of the foot contact force and the ground interaction. By combining the three types of indexes, the problems of local instability, high energy consumption, and gait imbalance can be effectively avoided, thereby improving the convergence speed, generalization ability, and execution stability of the reinforcement learning strategy, and significantly enhancing the adaptability of the robot to complex terrains and task changes.

[0091] Further, the adjustment factor and the control signal are fused to generate a target control instruction for driving the joint actuator based on a fuzzy logic reasoning method. Specifically,

[0092] An adjustment factor of a control signal output by the reinforcement learning model is taken as an input variable of the fuzzy controller, a corresponding membership function is constructed to map continuous values of each adjustment factor to a fuzzy state, such as "low", "medium", and "high", and the membership function adopts a Gaussian-shaped function;

[0093] Based on the fuzzy state, a preset experience rule base is used to perform fuzzy logic reasoning to generate a corresponding fuzzy control output, the rule base contains a relationship between an adjustment factor state and a corresponding control signal adjustment suggestion, and an example rule includes: "if the target angle adjustment factor is high and the target joint torque adjustment factor is medium, then the target angle should be moderately increased and the torque should be slightly adjusted", which directly guides the fuzzy adjustment direction and amplitude of the control signal;

[0094] A centroid method is used to convert the fuzzy control output into a numerical correction amount;

[0095] The numerical correction amount is fused with the control signal to generate a target control instruction, and the fusion method includes weighted superposition or multiplication modulation, and the target control instruction includes a target joint angle value, a target joint angular velocity value, a target joint torque value, and a target joint damping value;

[0096] The target control instruction is subjected to physical constraint verification to ensure that it meets the safety range and dynamic constraints of the actuator;

[0097] The verified target control instruction is output to drive the joint actuator.

[0098] It should be noted that the physical constraint verification includes limiting the maximum change amplitude and rate of angle, angular velocity, torque, and damping to prevent system instability caused by sudden changes in the control signal.

[0099] S4, in the process of executing the target control instruction, real-time state feedback data is collected, and the feedback data is compared and optimized with the currently executed target control instruction to realize feedback adjustment.

[0100] In the embodiment, in the process of executing the target control instruction, real-time state feedback data is collected, and the feedback data is compared and optimized with the currently executed target control instruction to realize feedback adjustment, specifically:

[0101] State feedback data in the process of executing the target control instruction by the joint actuator is obtained, and the state feedback data includes an actual joint angle value, an actual joint angular velocity value, an actual joint torque value, a foot-ground contact force value, and an acceleration value;

[0102] Numerical comparisons are made between the state feedback data and target joint angle, target joint angular velocity, and target joint torque values set in the target control instruction to calculate errors respectively.

[0103] constructing a quantitative feedback evaluation index based on the error and state feedback data, the feedback evaluation index including a trajectory tracking error index, a gait stability index and a contact impact index;

[0104] The trajectory tracking error index is the sum of squares of the joint angle value errors, the gait stability index is calculated from the deviation of the trunk acceleration modulus from a set expected acceleration interval, for measuring the stability in the robot walking process, and the contact impact index is the peak value of the vertical acceleration signal when the foot first contacts the ground, directly obtained by the foot end acceleration sensor.

[0105] The quantitative feedback evaluation index is composed into a multi-dimensional state vector as input data transmitted to the reinforcement learning model to obtain the optimized adjustment factor of the next control cycle;

[0106] The optimized adjustment factor is used to generate the target control instruction of the next control cycle, so that the control parameters are closer to the current system execution state, and specifically includes correcting the target joint angle, target joint angular velocity and target joint torque value with the optimized adjustment factor, to realize closed-loop feedback regulation.

[0107] Referring to Figure 2 The robot gait adaptive adjustment system structure provided by the present application in combination with the terrain feature recognition includes an image recognition module, a control signal generation module, a control signal correction module and a real-time feedback regulation module, and there is a connection between the modules.

[0108] The image recognition module is used to collect the terrain image in front of the robot's travel path and perform deblurring processing, and extract ground feature data from the deblurred image through a lightweight neural network.

[0109] The control signal generation module is used to match the ground feature data with a preset gait control parameter database to obtain a gait control parameter set, and generate a control signal through a proportional-differential control method according to the gait control parameter set.

[0110] The control signal correction module is used to correct the control signal based on a reinforcement learning algorithm and generate a target control instruction for driving the joint actuator.

[0111] The real-time feedback regulation module is used to collect state feedback data in real time during the execution of the target control instruction, compare and optimize the feedback data with the currently executed target control instruction, and realize feedback regulation.

[0112] The above embodiments can be realized in whole or in part by software, hardware, firmware or any other combination. When realized by software, the above embodiments can be realized in whole or in part in the form of a computer program product.

[0113] Those skilled in the art can understand that the modules and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software manner depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0114] In addition, each functional module in each embodiment of the present application can be integrated in one processing module, or each module can exist physically alone, or two or more modules can be integrated in one module.

[0115] The above is merely specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0116] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A robot gait adaptive adjustment method combining terrain feature recognition, characterized in that, Includes the following steps: The terrain image ahead of the robot's path is collected and deblurred. Ground feature data is extracted from the blurred image using a lightweight neural network. Ground characteristic data is matched with a preset gait control parameter database to obtain a gait control parameter set, and a control signal is generated based on the gait control parameter set using a proportional-derivative control method; The control signal is corrected based on a reinforcement learning algorithm, and a target control command for driving the joint actuator is generated, specifically: Construct a reinforcement learning model, taking ground feature vectors, state information, and control signals as input states; The adjustment factor of the control signal is used as the output action; Construct a reward function, which includes a vibration spectrum index, a gait symmetry measure, and a contact surface stability index; With the goal of maximizing the reward function, an adjustment factor for the control signal output by the reinforcement learning model is generated based on the current input state. The adjustment factor and control signal are fused based on the fuzzy logic reasoning method to generate target control commands to drive the joint actuator; The fuzzy logic reasoning method fuses the adjustment factor and control signal to generate the target control command, specifically as follows: The adjustment factor of the control signal output by the reinforcement learning model is used as the input variable of the fuzzy controller to construct the corresponding membership function and map the continuous values ​​of each adjustment factor to the fuzzy state. Based on the fuzzy state, fuzzy logic reasoning is performed using a preset empirical rule base to generate corresponding fuzzy control outputs. The centroid method is used to convert the fuzzy control output into a numerical correction quantity; The numerical correction value is fused with the control signal to generate a target control command, which includes a target joint angle value, a target joint angular velocity value, a target joint torque value, and a target joint damping value. Perform physical constraint verification on the target control command, and output the verified target control command to drive the joint actuator; During the execution of the target control command, status feedback data is collected in real time, and the feedback data is compared and optimized with the currently executed target control command to achieve feedback adjustment.

2. The robot gait adaptive adjustment method combining terrain feature recognition according to claim 1, characterized in that, The process of collecting terrain images ahead of the robot's path and performing deblurring is as follows: Each frame of the terrain image is converted into a single-channel grayscale image, and a sharpness index is calculated. If the sharpness index is lower than the preset blur threshold, the current image is determined to be a blurry image and deblurring is performed. Edge detection processing is performed on the blurred image to obtain the edge intensity map and orientation map; Divide the directional map into several directional intervals, count the number of pixels in each direction, and construct an edge direction distribution histogram; If the proportion of pixels in the main direction exceeds the preset direction concentration threshold, it is judged as linear motion blur; Otherwise, it is a non-linear jitter blur; Perform deblurring operations on linear motion blur and nonlinear jitter blur.

3. The robot gait adaptive adjustment method combining terrain feature recognition according to claim 2, characterized in that, The deblurring operation for linear motion blur and nonlinear jitter blur specifically involves: A two-dimensional Fourier transform is performed on the blurred image based on linear motion blur to obtain the blur direction and spacing, and the blur length is inferred from the spacing. A linear motion fuzzy kernel is constructed based on the fuzzy direction and fuzzy length as a linear fuzzy point diffusion function; Based on the point spread function, a non-blind deconvolution algorithm is used to restore the blurred image and obtain the blurred image. Based on nonlinear jitter blur, the blurred image is set as the observed image, the estimated sharp image and the nonlinear blur kernel are initialized, and an optimization objective function is constructed; Based on the objective function, an alternating minimization strategy is used to optimize the image and the blur kernel until a preset number of iterations is reached, resulting in an optimized nonlinear blur kernel. The nonlinear blur kernel is used as the blur kernel for non-blind deconvolution to perform deblurring and obtain a clear image.

4. The robot gait adaptive adjustment method combining terrain feature recognition according to claim 1, characterized in that, The extraction of ground characteristic data from the blurred image using a lightweight neural network specifically involves: The blurred image is preprocessed to obtain a standardized image; Using standardized images as input, a lightweight convolutional neural network model is constructed for preliminary identification of ground features, and the output is a ground feature classification map. Obtain the confidence level in the ground feature classification map, and based on the preset confidence level threshold, extract the image regions with low confidence levels in the ground feature classification map to construct the corresponding candidate region mask map; Based on the candidate region mask map, the corresponding high-resolution image sub-blocks are extracted from the terrain image, and the sub-blocks are used as input to the preset refined recognition model; The output of the refined recognition model is fused with the ground feature classification map to generate a ground feature semantic segmentation map, which is then converted into structured ground feature data.

5. The robot gait adaptive adjustment method combining terrain feature recognition according to claim 1, characterized in that, The process involves matching ground characteristic data with a preset gait control parameter database to obtain a gait control parameter set, and then generating a control signal based on the gait control parameter set using a proportional-derivative control method. Specifically: The ground characteristic data is linearly normalized to obtain the ground feature vector; The Euclidean distance between the current ground feature vector and the ground feature vector in the database is calculated based on the K-nearest neighbor algorithm, and the set of gait control parameters corresponding to the minimum distance is selected as the output result. The current state information of the robot is obtained, and the control correction term is calculated based on the error between the matched set of control parameters and the state information using the proportional-derivative control method. Control signals are generated based on control correction terms.

6. The robot gait adaptive adjustment method combining terrain feature recognition according to claim 1, characterized in that, During the execution of the target control command, real-time status feedback data is collected, and the feedback data is compared and optimized with the currently executed target control command to achieve feedback adjustment. Specifically: Acquire state feedback data during the execution of target control commands by the joint actuator. The state feedback data includes actual joint angle value, actual joint angular velocity value, actual joint torque value, foot-ground contact force value, and acceleration value. The errors are calculated by comparing the status feedback data with the target control commands. Based on the error and state feedback data, a quantitative feedback evaluation index is constructed, which includes a trajectory tracking error index, a gait stability index, and a contact impact index. The quantitative feedback evaluation indicators are combined into a multidimensional state vector, which is then transmitted as input data to the reinforcement learning model to obtain the optimization adjustment factor for the next control cycle. The target control command for the next control cycle is generated by using the optimized adjustment factor, thereby realizing closed-loop feedback regulation.

7. A system using the robot gait adaptive adjustment method combining terrain feature recognition as described in any one of claims 1-6, characterized in that, It includes an image recognition module, a control signal generation module, a control signal correction module, and a real-time feedback adjustment module, and the modules are interconnected. The image recognition module is used to acquire terrain images ahead of the robot's path and perform deblurring processing. It then uses a lightweight neural network to extract ground feature data from the blurred images. The control signal generation module is used to match ground characteristic data with a preset gait control parameter database to obtain a gait control parameter set, and generate a control signal based on the gait control parameter set using a proportional-derivative control method. The control signal correction module is used to correct the control signal based on a reinforcement learning algorithm and generate target control commands to drive the joint actuator. The real-time feedback adjustment module is used to collect status feedback data in real time during the execution of the target control command, and compare and optimize the feedback data with the currently executed target control command to achieve feedback adjustment.

Images