A method for estimating foot-ground contact force and terrain classification for an ostrich-like robot

Abstract

The present invention discloses a method for estimating the foot-ground contact force and classifying terrain of an ostrich-like robot. The method comprises: constructing a foot-ground contact force mapping matrix of the ostrich-like robot, performing force-position hybrid control, obtaining the foot-ground contact force based on the matrix and the input torque of the leaf spring on the passive joint, and obtaining terrain classification information after determining the ground contact state; establishing a terrain classification fusion model based on a weighted voting mechanism, and training the model based on the terrain classification information; and performing terrain classification based on the terrain classification information and the trained model. The present invention utilizes the robot's joint angle information to estimate the foot-ground contact force without the need for additional sensors, achieving real-time and accurate foot-ground contact force estimation and contact state acquisition, effectively solving the problem of performance degradation or failure of the ostrich-like robot system caused by the fragility of the force sensor, simplifying the system configuration, and being able to complete the foot-ground contact force estimation and accurate terrain classification of the ostrich-like robot without relying on external sensors.

Description

Technical Field

[0001] The present invention relates to a method for estimating foot-ground contact force and classifying terrain, and relates to the field of perception of foot-type robots, in particular to a method for estimating foot-ground contact force and classifying terrain of an ostrich-like robot. Background Art

[0002] As a type of mobile robot, ostrich-like robots possess biologically similar maneuverability and excellent environmental adaptability, promising to replace humans in a wide range of workplaces. However, ostrich-like robots currently lack stable operation in complex terrain, as terrain changes can cause motion instability and even component damage. Their adaptability relies on acquiring foot-ground contact force and terrain information. Foot-ground contact force is crucial for environmental perception and compliant control, while terrain information is crucial for ostrich-like robots' adaptation. Compliant control methods such as terrain recognition and impedance control based on foot-ground contact force often achieve good control results. However, force / torque sensors are fragile and expensive, significantly impacting the robustness and cost-effectiveness of robotic systems. Furthermore, installing force / torque sensors on the foot increases the foot's mass and moment of inertia, compromising the robot's motion performance. Therefore, research on foot-ground contact force estimation and terrain classification methods that do not utilize force / torque sensors is an urgent issue. Summary of the Invention

[0003] In order to solve the problems existing in the background technology, the present invention provides a method for estimating the foot-ground contact force and classifying terrain of an ostrich-like robot.

[0004] The technical solution adopted in the present invention is:

[0005] The method for estimating the foot-ground contact force and classifying terrain of an ostrich-like robot of the present invention comprises:

[0006] In the first step, the foot-ground contact force mapping matrix is ​​constructed based on the kinematic model of the ostrich-like robot including joint leaf springs.

[0007] In the second step, the ostrich-like robot is subjected to force-position hybrid control. During the control process, the passive joints of the ostrich-like robot are subjected to torque balance analysis based on the foot-ground contact force mapping matrix to obtain an estimated foot-ground contact force. The ground contact state is determined based on the estimated foot-ground contact force, and terrain classification information including slope information and current characteristic information is obtained.

[0008] The third step is to establish a terrain classification fusion model based on the weighted voting mechanism, train the terrain classification fusion model through the current feature information in the terrain classification information, and obtain the trained terrain classification fusion model.

[0009] In the fourth step, when the ostrich-like robot jumps on the terrain to be detected, the same operation as in the second step is performed on the ostrich-like robot to obtain the classification information of the terrain to be detected, and the terrain classification is performed according to the classification information of the terrain to be detected and the trained terrain classification fusion model.

[0010] In the first step, the passive joints of the ostrich-like robot all use high-rigidity leaf springs as flexible components to provide input torque, and the active joints are driven by servo motors. The limb structure of the ostrich-like robot is a planar four-bar parallelogram mechanism with two passive joints. The deformation of the leaf springs of the passive joints is obtained by reading and calculating the joint encoders of the ostrich-like robot, and the input torque of the leaf springs on the passive joints is then obtained through the deformation and stiffness coefficient of the leaf springs.

[0011] In the first step, the foot-ground contact force mapping matrix is ​​as follows:

[0012]

[0013] Among them, F c Estimated foot-ground contact force of the ostrich-like robot; J P D is the mapping matrix of the foot-ground contact force to the passive joint; P is the mapping matrix from the joint encoder reading to the input torque of the leaf spring on the passive joint; Δθ is the difference between the joint encoder reading and the joint zero position; τ s is the input torque of the leaf spring on the passive joint; L3 and L4 are the lengths of the tibia and tarsus of the ostrich-like robot, respectively; s 123 and s 1234 are the first and second angular sines, s 123 =sin(θ1+θ2+θ3), s 1234 = sin(θ1+θ2+θ3+θ4), θ1, θ2, θ3, θ4 are the rotation angles of the hip joint, knee joint, first passive joint and second passive joint of the ostrich-like robot respectively, c 123 and c 1234 are the first and second angle cosines, c 123 =cos(θ1+θ2+θ3), c 1234 =cos(θ1+θ2+θ3+θ4).

[0014] In the second step, the ostrich-like robot is subjected to force-position hybrid control to control the ostrich-like robot to perform jumping movements along preset periodic trajectories on different terrains. Position control is performed on the ankle joint servo motor during the takeoff phase, and force control is performed during the touchdown phase. The output force is set to a small value to make the sole of the foot conform to the ground. For each jumping movement, when the estimated foot-ground contact force exceeds a preset touchdown detection threshold during the process of the ostrich-like robot jumping and landing, the current moment is used as the touchdown moment, the ankle joint angle read by the joint encoder at the touchdown moment is used as slope information, and the real-time current of the ankle joint motor is obtained. The first current extreme point of the ankle joint after the touchdown moment is used as the ankle joint current peak value, and the mean and variance of each ankle joint current peak value under each terrain are used as current characteristic information.

[0015] In the third step, the terrain classification fusion model includes a support vector machine (SVM) and a back propagation BP (Back Propagation) neural network optimized by a particle swarm optimization algorithm (PSO). The support vector machine (SVM) and the back propagation BP neural network optimized by a particle swarm optimization algorithm (PSO) are integrated using a weighted voting mechanism based on confidence rate. The mapping relationship between current feature information and classified terrain is constructed by the support vector machine (SVM) and the back propagation BP neural network optimized by a particle swarm optimization algorithm (PSO), thereby establishing a terrain classification fusion model. The back propagation BP neural network optimized by the particle swarm optimization algorithm (PSO) uses the weights and thresholds of the neural network as independent variables, and the mean square error of the network as the fitness function. The weights and thresholds of the network are optimized by the particle swarm optimization algorithm (PSO) to train the model.

[0016] The weighted voting mechanism based on confidence rate integrates the support vector machine (SVM) and the back propagation BP neural network optimized by the particle swarm optimization algorithm (PSO). The confidence level Y of the integrated terrain classification fusion model is as follows:

[0017]

[0018] in, and The confidence of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO is and It is the confidence of the output results of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO.

[0019] In the fourth step, slope information in the terrain classification information to be tested is used to identify slopes. The current characteristics in the terrain classification information to be tested are then input into the trained terrain classification fusion model to identify the current terrain type, completing the terrain classification. In specific implementations, other terrain types can include sandy and rocky terrain, flat land, and soft terrain.

[0020] The terrain classification system of the present invention comprises:

[0021] The data generation layer performs force-position hybrid control on the ostrich-like robot through a given trajectory, so that the ostrich-like robot generates data.

[0022] In the data acquisition layer, when the ostrich-like robot performs force-position hybrid control, the slope information is obtained through the readings of the joint encoder of the ostrich-like robot. At the same time, the current data of the joint driver of the ostrich-like robot is collected. The current feature information is obtained after feature extraction based on the foot-ground contact force estimated by the foot-ground contact force mapping matrix.

[0023] In the data model layer, the terrain classification model is trained according to the current characteristic information, and then the trained terrain classification fusion model is obtained through a weighted voting mechanism based on confidence rate.

[0024] The data analysis layer estimates slopes based on slope information and identifies other terrains based on current characteristic information and the trained terrain classification fusion model.

[0025] The electronic device of the present invention is characterized in that it comprises: a memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method as described above.

[0026] The computer-readable storage medium of the present invention stores program data thereon, and is characterized in that the program data implements the method described above when executed by a processor.

[0027] The present invention first establishes a mapping matrix of the foot-ground contact force to the passive joint, maps the foot-ground contact force to the passive joint, and then solves the passive joint input torque through the joint encoder reading in the parallelogram structure. The robot foot-ground contact force is obtained based on the torque balance equation of the two passive joints according to the passive joint input torque and the foot-ground contact force mapping torque to the passive joint. The foot-ground contact force is obtained based on the aforementioned foot-ground contact force estimation method for ground contact detection, and force-position hybrid control is performed based on the foot-ground contact state, so that the robot completes the set jumping gait on the terrain to be tested and obtains the joint current peak; data preprocessing is performed to obtain slope and current characteristic information; the slope is identified through the slope information; the mapping relationship between current characteristic information and terrain is constructed through the support vector machine (SVM) and the back propagation BP neural network optimized by the particle swarm optimization algorithm (PSO), and the identification of sandy and flat land is completed. The two terrain classification models are integrated by using a weighted voting mechanism based on confidence to obtain a terrain classification fusion model, which further improves the classification accuracy. The classification of flat land, sloped land, and sandy and rocky land is completed based on the robot's body information.

[0028] The beneficial effects of the present invention are:

[0029] This invention estimates foot-ground contact force by utilizing the robot's joint angle information, eliminating the need for additional sensors. This allows for real-time, accurate foot-ground contact force estimation and contact status acquisition, effectively resolving the issue of ostrich-like robot system performance degradation or failure caused by fragility of force sensors. This simplifies system configuration, reduces hardware costs, and improves the overall stability of the robot system. Without the need for additional sensors, the invention can classify flat land, sloping land, and sandy and rocky terrain. Verified terrain classification accuracy exceeds 90%. This provides an accurate and reliable terrain perception method that helps enhance the flexibility and robustness of ostrich-like robot motion. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a diagram of a single leg of the ostrich-like robot of the present invention, wherein: Figure 1 (a) is a structural diagram of a single leg of the ostrich-like robot of the present invention. Figure 1 (b) is a schematic diagram of the single-limb structure of the ostrich-like robot of the present invention;

[0031] Figure 2 is a diagram of the foot-ground contact force estimation framework of the present invention;

[0032] Figure 3 It is a framework diagram of the terrain classification system of the present invention;

[0033] Figure 4 is a flow chart of the jumping gait control of the present invention;

[0034] Figure 5 is a flow chart of current information preprocessing of the present invention;

[0035] Figure 6 : is the result of the foot-ground contact force estimation experiment of the present invention, Figure 6 (a) is a graph showing the experimental results of the present invention for estimating the collision force on flat ground. Figure 6 (b) is a graph showing the experimental results of the present invention for estimating the collision force on a slope. Figure 6 (c) is a graph showing the experimental results of the present invention for estimating the impact force on sandy ground. DETAILED DESCRIPTION

[0036] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the terms "first," "second," and the like in the specification and claims of the present invention, as well as in the accompanying drawings, are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, such that the embodiments of the present invention described herein can be practiced in sequences other than those illustrated or described herein.

[0037] like Figure 1 As shown in (a), the hip joints, knee joints, tibia joints, ankle joints, connecting rods, and leaf springs of the ostrich-like robot's limbs together form a rigid-flexible coupled closed link mechanism. To facilitate kinematic analysis, the model is simplified as follows: Figure 1 The four-bar linkage structure shown in (b) is shown in Figure 1, where points M and C are the centers of the hip joint, N is the center of the knee motor, P is the center of the first passive joint, and a leaf spring s1 is installed. A is the center of the second passive joint, and the installed leaf spring BQ is denoted as s2. F is the rotation center of the ankle joint. The MF line is perpendicular to the base, and the angle between it and MN is denoted as θ1. The angle between the MN extension line and NP is denoted as θ2, the angle between the NP extension line and PA is denoted as θ3, and the angle between the PA extension line and AF is denoted as θ4. The length of MN is denoted as L1, the length of NP is denoted as L2, the length of PA is denoted as L3, and the length of AF is denoted as L4. The origin of the base coordinate coincides with point M, the X0 axis coincides with the MF line and points to F, the Z0 axis is perpendicular to the connecting rod plane and points outward, and the coordinate axis position of the Y0 axis conforms to the right-hand system. The coordinates of each joint under the base coordinate are expressed as follows:

[0038]

[0039] For ease of description, the above formula has the following simplified formula: cos(θ i )=c i , cos(θ i +θ j )=c ij ,sin(θ i )=s1,sin(θi +θ j )=s ij , and so on.

[0040] like Figure 2 As shown, on this basis, the present invention performs foot-ground contact force estimation and terrain classification of the ostrich-like robot as follows:

[0041] Firstly, based on the kinematic model of an ostrich-like robot with joint leaf springs, a foot-ground contact force mapping matrix is ​​constructed. The passive joints of the ostrich-like robot all use high-stiffness leaf springs as flexible components to provide input torque, and the active joints are driven by servo motors. The limb structure of the ostrich-like robot is a planar four-bar parallelogram mechanism with two passive joints. The leaf spring deformation of the passive joint is obtained by reading the joint encoder of the ostrich-like robot, and the input torque of the leaf spring on the passive joint is obtained by the leaf spring deformation and stiffness coefficient.

[0042] A planar four-bar linkage has three types of input torque: the torque reflected from the foot-ground contact force on each joint, the input torque from the servo motor on the active joint, and the input torque from the leaf springs of the elastic elements on the passive joints. Since passive joints cannot transmit torque, the leaf spring of the first passive joint does not act on the leaf spring of the second passive joint. The leaf spring of the second passive joint acts on the second passive joint, and the force of its leaf spring link is directed toward the center of the hip joint. This force does not generate torque on the first passive joint, but is resisted by the mechanical structure itself. Therefore, the leaf spring of the second passive joint does not act on the leaf spring of the first passive joint. Therefore, the only torque balance to consider on the passive joints is the torque reflected from the foot-ground contact force on the passive joints and the input torque from the leaf springs of the passive joints.

[0043] For the working scenario of the ostrich-like robot, the deformation of the leaf spring during the working process of the ostrich-like robot and the input torque to the passive joint can be approximately linearly related. Therefore, the mapping matrix D from the joint encoder reading to the passive joint input torque can be obtained. P , the weighted least square method is used to identify the system and obtain the leaf spring stiffness. The input torque can be obtained from the stiffness coefficient and the leaf spring deformation.

[0044] The foot-ground contact force mapping matrix is ​​as follows:

[0045]

[0046] Among them, F c Estimated foot-ground contact force of the ostrich-like robot; J P D is the mapping matrix of the foot-ground contact force to the passive joint; P is the mapping matrix from the joint encoder reading to the input torque of the leaf spring on the passive joint; Δθ is the difference between the joint encoder reading and the joint zero position; τs is the input torque of the leaf spring on the passive joint; L3 and L4 are the lengths of the tibia and tarsus of the ostrich-like robot, respectively; s 123 and s 1234 are the first and second angular sines, s 123 =sin(θ1+θ2+θ3), s 1234 = sin(θ1+θ2+θ3+θ4), θ1, θ2, θ3, θ4 are the rotation angles of the hip joint, knee joint, first passive joint and second passive joint of the ostrich-like robot respectively, c 123 and c 1234 are the first and second angle cosines, c 123 =cos(θ1+θ2+θ3), c 1234 =cos(θ1+θ2+θ3+θ4).

[0047] Based on the kinematic model of the ostrich-like robot, the mapping matrix J of the ostrich-like robot's foot-ground contact force to the passive joint is constructed. P , considering the construction of the mapping matrix of the foot-ground contact force to the passive joint of the ostrich-like robot geometry:

[0048] First, the foot end position can be obtained as follows:

[0049]

[0050] in, is the transformation matrix from the passive joint to the end joint, are the transformation matrices from the passive joint P to the foot end coordinate system N to the passive joint P; x F and y F are the horizontal and vertical coordinates of the foot end in the passive joint coordinate system, x n and y n are the horizontal and vertical coordinates of the foot end in the end coordinate system respectively.

[0051] The mapping matrix of the foot-ground contact force of the ostrich-like robot to the passive joint is obtained by differentiating the passive joint transformation matrix. Based on the kinematics of the ostrich-like robot, the input torque of the passive joint is calculated by reading the single-turn absolute encoder of the joint in the parallelogram structure. A torque balance analysis is performed on the passive joint. Considering the mapping torque of the foot-ground collision force on each joint and the input torque of the elastic element leaf spring on the passive joint, the foot-ground contact force is mapped to the passive joint through the mapping matrix. The torque balance analysis is performed with the passive joint input torque to obtain the torque balance equation. Input torque τ of leaf spring on passive joint sThe mapping matrix from the passive joint encoder readings to the passive joint input torque of the ostrich-like robot can be obtained. Therefore, the foot-ground contact force, i.e., the foot-ground contact force mapping matrix, can be obtained through the robot body sensor information.

[0052] The mapping matrix from joint encoder readings to passive joint input torque is as follows:

[0053]

[0054] Among them, K P and K A are the leaf spring stiffness at the passive joint point P and point A respectively.

[0055] The passive joint input torque is as follows:

[0056]

[0057] Where K1 and K2 are the stiffnesses of leaf spring 1 and leaf spring 2, respectively; Δθ2, Δθ3, and Δθ4 are the encoder readings of joints 2, 3, and 4 relative to the zero position, respectively.

[0058] Then, the ostrich-like robot was subjected to force-position hybrid control. During the control process, the passive joints of the ostrich-like robot were subjected to torque balance analysis based on the foot-ground contact force mapping matrix to obtain an estimated foot-ground contact force. The ground contact state was determined based on the estimated foot-ground contact force, and terrain classification information, including slope information and current characteristics, was obtained. The ostrich-like robot was controlled by force-position hybrid control to perform jumping movements on different terrains along a preset periodic trajectory. The ankle joint servo motor was position-controlled during the takeoff phase and force-controlled during the touchdown phase, with the output force set to a small value to ensure that the sole of the foot was in contact with the ground. For each jump, when the estimated foot-ground contact force exceeded the preset touchdown detection threshold during the jump and landing process, the current moment was regarded as the touchdown moment. Based on the foot-ground contact force range between the robot's suspension and touchdown periods during the experiment, 1000N was used as the touchdown threshold. When the foot-ground contact force exceeded the touchdown threshold, it was considered to have touched the ground. The ankle joint angle read by the joint encoder at the moment of contact is used as the slope information, and the real-time current of the ankle joint motor is obtained at the same time. The first current extreme point of the ankle joint after the moment of contact is used as the ankle joint current peak value, and the mean and variance of each ankle joint current peak value under each terrain are used as current characteristic information. Figure 4 As shown, when the foot-ground contact force F c If the force is greater than the set threshold F1, the robot is in the ground contact phase, and the ankle joint motor uses the force control mode and sets the output force to a smaller value to make the sole of the foot fit the ground; when the foot contact force F cIf the value is less than the set threshold F2, it is in the air phase, and the ankle joint motor uses the position control mode in the air phase. At the same time, the slope and current characteristic information are collected. Figure 5 The collision moment is recorded as the index value, and the index value is used to record the first current extreme point after the collision moment from the ankle joint motor current as the ankle joint current peak value of each collision, and the mean and variance of the ankle joint current peak value are used as current characteristic information.

[0059] Based on the foot-ground contact state, force-position hybrid control is performed, allowing the robot to complete a programmed periodic jumping gait on the terrain to be identified, while simultaneously collecting ankle joint encoder and joint current information. Based on the foot-ground contact state, these two types of proprioceptive information, joint motor current and encoder readings, are preprocessed to obtain slope information and current characteristics.

[0060] A terrain classification fusion model based on a weighted voting mechanism was then established. The model was trained using the current signature information from the terrain classification data to obtain a trained terrain classification fusion model. After multiple experiments, the current signature information was labeled according to terrain type to form a training dataset: sandy terrain was assigned a value of 1, and non-sandy terrain was assigned a value of 0. This training dataset was then fed into an SVM model and a BP neural network improved by PSO for training. The PSO-improved BP neural network was a three-layer, 8×8×1, fully connected BP neural network with two inputs and one output. The weights and thresholds of the BP neural network were optimized using the PSO algorithm. Based on the terrain classification accuracy obtained from the experiments, the confidence rate of the terrain classification model was calculated. Using a weighted voting mechanism based on the confidence rate, a terrain classification fusion model was obtained that integrated the SVM terrain classification model with the PSO-optimized BP neural network. The terrain classification fusion model uses a support vector machine (SVM) and a back-propagation BP neural network optimized by a particle swarm optimization (PSO) algorithm to construct a mapping relationship between current characteristic information and classified terrain. The PSO-optimized back-propagation BP neural network uses the neural network weights and thresholds as independent variables and the network's mean squared error as the fitness function. The PSO-optimized back-propagation BP neural network optimizes the network weights and thresholds for model training. The confidence level Y of the integrated terrain classification fusion model is as follows:

[0061]

[0062] in, and The confidence of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO is and It is the confidence of the output results of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO.

[0063] Finally, when the ostrich-like robot jumps on the terrain to be tested, the same operations as in step 2 are performed on it to obtain the terrain classification information to be tested. Slopes are identified using the slope information in the terrain classification information to be tested. The current characteristic information in the terrain classification information to be tested is then input into the trained terrain classification fusion model to identify the current terrain category and complete the terrain classification. In specific implementations, other terrains can include gravel, flat, and soft terrain. The current characteristic information is input into the terrain classification fusion model to complete the classification of gravel terrain: when the model output is greater than or equal to 0.5, the output result is gravel terrain; otherwise, it is non-sand and gravel terrain. At this point, the ostrich-like robot has completed the classification of three typical terrain types: flat, sloping, and gravel, based on the ontological information.

[0064] like Figure 3 As shown, the present invention also designs a terrain classification system, including a data generation layer, a data acquisition layer, a data model layer, and a data analysis layer. The data generation layer performs force-position hybrid control on the ostrich-like robot using a given trajectory, thereby enabling the ostrich-like robot to generate data. The given trajectory is used to obtain joint trajectories for the ostrich-like robot's motion control through inverse kinematics. When the ostrich-like robot touches the ground, force control and position control modes are switched to reduce the impact of foot-ground collision on the robot's joints. When the ostrich-like robot performs force-position hybrid control, the data acquisition layer obtains slope information through the ostrich-like robot's joint encoder readings. At the same time, it collects current data from the ostrich-like robot's joint actuators and extracts current feature information based on the foot-ground contact force estimated by the foot-ground contact force mapping matrix. The data model layer trains the terrain classification model based on the current feature information and then obtains a trained terrain classification fusion model through a weighted voting mechanism based on confidence rate. The data analysis layer estimates slopes based on the slope information and identifies other terrains based on the current feature information and the trained terrain classification fusion model.

[0065] The present invention first establishes a mapping matrix of the foot-ground contact force to the passive joint, maps the foot-ground contact force to the torque of the passive joint, calculates the deformation of the passive joint flexible component through the encoder reading, and then obtains the passive joint input torque, and performs torque balance analysis on the input torque at the passive joint to obtain the foot-ground contact force. Touchdown detection is performed based on the obtained foot-ground contact force. The robot is subjected to force-position hybrid control based on the foot-ground contact state, so that it completes the set jumping gait on the terrain to be tested, and the robot's joint current and encoder reading are obtained. Flat land and slope land are classified according to the ankle joint encoder reading at the moment of touchdown, and the mean and variance of the ankle joint current peak are obtained as current feature information based on the foot-ground contact state. The classification of slope land and sandy land is completed through the terrain classification fusion model. The present invention can complete the foot-ground contact force estimation and terrain classification of the ostrich-like robot without relying on external sensors.

[0066] In a specific embodiment of the present invention, a limb prototype platform was used to conduct foot-ground contact force estimation and terrain recognition experiments on three types of terrain: flat land, sloping land, and sandy land. The experiments are as follows:

[0067] 1) Control the robot to complete the set jumping gait on three types of terrain, and record the terrain type and experimental data. The experimental data for contact force estimation are as follows Figure 6 (a) Figure 6 (b) and Figure 6 As shown in (c), the estimated contact force quickly reaches over 1000 N after ground contact. Therefore, this contact force estimation method can accurately detect ground contact.

[0068] 2) Conduct multiple experiments on three different terrain types, controlling the robot to perform a set jumping gait and recording the terrain type and experimental data. Divide the experimental data into a training dataset and a test dataset. After training and testing the terrain classification model, conduct a terrain recognition experiment. The experimental process and results are shown below:

[0069] 1. First, train the terrain classification model. The support vector machine (SVM) classifier parameters are: Support Vector Machine (SVM) Type: Classification Support Vector Machine (SVM); Kernel Function Type: Linear Kernel Function; Penalty Parameter: 1; Whether to use heuristics: Enabled; Termination Criterion: 0.001. Label the current characteristic information according to the terrain category: 1 for sandy terrain and 0 for non-sandy terrain. After training the terrain classification model with a training dataset consisting of multiple sets of labeled current characteristic information, the identification of sandy and flat land can be completed. The 150 sets of data were divided into a training dataset and a test dataset in a ratio of 4:1, and the accuracy rates on the two datasets were 100% and 93.3%, respectively. The three-layer 8×8×1 fully connected BP neural network terrain classification model optimized by PSO was divided into a training dataset and a test dataset in a ratio of 4:1 for training and testing, and the accuracy rates on the two datasets were 100% and 96.7%, respectively. The terrain classification fusion model had an accuracy of 100% in the test dataset, and the output of the terrain classification fusion model was:

[0070]

[0071] in, and is the confidence of the support vector machine (SVM) model and the back propagation BP neural network optimized by the particle swarm optimization algorithm (PSO). Here, the accuracy of the two models on the experimental test set is used as the model confidence. and It is the confidence of the output results of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO.

[0072] 2. After model training is completed, six groups of terrain classification experiments are conducted on three types of terrain: slope, flat land, and sandy land, and the experimental results are analyzed:

[0073] First, the slope terrain is identified, and the slope terrain identification threshold is set to 15°. The experimental data and results are shown in Table 1. It can be seen that in all six groups of experiments, the slope terrain was correctly identified.

[0074] Table 1 Summary of experimental data on slope identification

[0075]

[0076] Next, we categorize the terrain into sandy and flat terrain. Based on the foot-ground contact state, we calculate the mean and variance of the ankle joint motor current peak value during each contact. The experimental data for the flat terrain are shown in Table 2:

[0077] Table 2 Summary of flat land experimental data

[0078]

[0079] The experimental data of sand and gravel underground are shown in Table 3:

[0080] Table 3 Summary of experimental data on sandy and rocky land

[0081]

[0082] The model input parameters are input into the trained terrain classification model. From Table 2, Table 3 and the results of the test data set, it can be seen that all three models have completed the terrain classification, and the accuracy of the fusion model result is higher than that of the terrain classification model using the support vector machine (SVM) and the BP neural network based on PSO optimization alone.

[0083] In summary, the method of the present invention can accurately estimate foot-ground contact force and classify common terrains such as slopes, flat land, and sandy and rocky terrain. Furthermore, the contact force estimation and terrain classification of this method are performed online, meeting the real-time control requirements of the ostrich-like robot. Then, contact force estimation and terrain recognition experiments were conducted on an experimental platform. The contact force estimation experiment verified the effectiveness of force estimation and the accuracy of ground contact detection based on force estimation. The terrain classification experiment correctly classified the terrain in six experiments conducted on three types of terrain, verifying the accuracy of the proposed terrain classification method. The two terrain classification models were integrated using a confidence-based weighted voting mechanism to obtain a terrain classification fusion model, further improving classification accuracy. This demonstrates that the proposed contact force estimation and terrain classification method for the ostrich-like robot has high accuracy and reliability and can be applied to the robot's perception system to estimate foot-ground contact force and classify terrain.

[0084] The above content is only the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for estimating foot-ground contact force and terrain classification for an ostrich-like robot, characterized in that: include: The first step is to construct the foot-ground contact force mapping matrix based on the kinematic model of the ostrich-like robot including joint leaf springs; The second step is to perform force-position hybrid control on the ostrich-like robot. During the control process, the passive joints of the ostrich-like robot are subjected to torque balance analysis based on the foot-ground contact force mapping matrix to obtain an estimated foot-ground contact force. The ground contact state is determined based on the estimated foot-ground contact force, and terrain classification information including slope information and current characteristics is obtained. The third step is to establish a terrain classification fusion model based on a weighted voting mechanism. The terrain classification fusion model is trained using the current feature information in the terrain classification information to obtain a trained terrain classification fusion model. In the fourth step, when the ostrich-like robot jumps on the terrain to be detected, the same operation as in the second step is performed on the ostrich-like robot to obtain the classification information of the terrain to be detected, and the terrain classification is performed based on the classification information of the terrain to be detected and the trained terrain classification fusion model; In the first step, the passive joints of the ostrich-like robot all use leaf springs as flexible components to provide input torque, and the active joints are driven by servo motors. The ostrich-like robot's limb structure is a planar four-bar parallelogram mechanism with two passive joints. The leaf spring deformation of the passive joint is obtained by reading the joint encoder of the ostrich-like robot, and the input torque of the leaf spring on the passive joint is obtained by the leaf spring deformation and stiffness coefficient. In the first step, the foot-ground contact force mapping matrix is ​​as follows: in, Estimated foot-ground contact force for the ostrich-like robot; is the mapping matrix of the foot-ground contact force to the passive joints; It is the mapping matrix from the joint encoder reading to the input torque of the leaf spring on the passive joint; is the difference between the joint encoder reading and the joint zero position; is the input torque of the leaf spring on the passive joint; and are the lengths of the tibia and tarsus of the ostrich-like robot, respectively; and are the first and second angle sines, , , 、 、 、 They are the rotation angles of the hip joint, knee joint, first passive joint and second passive joint of the ostrich-like robot. and are the first and second angle cosines, , ; In the second step, the ostrich-like robot is subjected to force-position hybrid control, and is controlled to perform jumping movements along preset periodic trajectories on different terrains. For each jumping movement, when the estimated foot-ground contact force exceeds a preset touchdown detection threshold during the ostrich-like robot's jump and landing, the current moment is used as the touchdown moment, and the ankle joint angle read by the joint encoder at the touchdown moment is used as slope information. At the same time, the real-time current of the ankle joint motor is obtained, and the first current extreme point of the ankle joint after the touchdown moment is used as the ankle joint current peak value. The mean and variance of each ankle joint current peak value under each terrain are used as current characteristic information.

2. The method for estimating foot-ground contact force and classifying terrain of an ostrich-like robot according to claim 1, characterized in that: include: In the third step, the terrain classification fusion model includes a support vector machine (SVM) and a back-propagation BP neural network optimized based on a particle swarm optimization algorithm (PSO). The support vector machine (SVM) and the back-propagation BP neural network optimized based on a particle swarm optimization algorithm (PSO) are integrated using a weighted voting mechanism based on confidence rate. The mapping relationship between current feature information and classified terrain is constructed through the support vector machine (SVM) and the back-propagation BP neural network optimized based on a particle swarm optimization algorithm (PSO), thereby establishing a terrain classification fusion model.

3. The method for estimating foot-ground contact force and classifying terrain of an ostrich-like robot according to claim 2, wherein: include: The weighted voting mechanism based on confidence rate integrates the support vector machine SVM and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO. The confidence of the integrated terrain classification fusion model is as follows: in, and The confidence of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO is and It is the confidence of the output results of the support vector machine SVM model and the back propagation BP neural network optimized by the particle swarm optimization algorithm PSO.

4. The method for estimating foot-ground contact force and classifying terrain of an ostrich-like robot according to claim 1, wherein: include: In the fourth step, the slope is identified by the slope information in the terrain classification information to be detected, and the current characteristic information in the terrain classification information to be detected is input into the trained terrain classification fusion model to identify the category of the current terrain and complete the terrain classification.

5. A terrain classification system suitable for the method according to any one of claims 1 to 4, characterized in that: include: The data generation layer performs force-position hybrid control on the ostrich-like robot through a given trajectory, so that the ostrich-like robot generates data; The data acquisition layer obtains slope information through the joint encoder readings of the ostrich-like robot when the robot performs force-position hybrid control. At the same time, it collects current data from the joint actuators of the ostrich-like robot. The current feature information is obtained by extracting the foot-ground contact force estimated by the foot-ground contact force mapping matrix. The data model layer trains the terrain classification model based on the current characteristic information, and then obtains the trained terrain classification fusion model through a weighted voting mechanism based on confidence rate; The data analysis layer estimates slopes based on slope information and identifies other terrains based on current characteristic information and the trained terrain classification fusion model.

6. An electronic device, characterized in that: include: A memory and a processor coupled to each other, wherein the memory stores program data, and the processor calls the program data to execute the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having program data stored thereon, characterized in that: When the program data is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images