Foot type robot terrain recognition method and equipment based on three-dimensional force perception
By employing 3D force sensing and wavelet transform techniques, the problem of terrain recognition in complex environments using visual and traditional tactile sensing methods has been solved, achieving high-precision and low-cost terrain recognition and enhancing the robot's autonomy and task execution capabilities in changing environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-10
Smart Images
Figure CN121637121A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of legged robot perception technology, specifically to a method and device for terrain recognition of legged robots based on three-dimensional force perception. Background Technology
[0002] Legged robots, as an important mobile platform, have demonstrated excellent adaptability in various environments. Terrain perception and classification are among the key technologies for autonomous navigation and environmental adaptation in legged robots. Accurate terrain perception helps robots make decisions in complex, unknown, or dynamically changing environments, thereby improving their autonomy and task execution capabilities. Through terrain classification, robots can identify different ground types, such as grass, sand, rocks, or concrete surfaces, which is crucial for robots to perform tasks such as exploration, rescue, and patrol in outdoor environments. Effective terrain classification not only improves the robot's motion stability but also enables it to plan paths effectively, avoiding traps or injuries caused by uneven ground or obstacles.
[0003] Traditional methods for terrain recognition in legged robots largely rely on visual perception systems. These methods typically employ techniques such as image processing, deep learning, or convolutional neural networks to analyze ground information. However, these processing methods require high-performance platforms to ensure real-time performance. Furthermore, visual perception systems significantly degrade in low-light environments, under changing viewing angles, and with occlusion, impacting the robot's terrain recognition accuracy. The dependence of visual systems on ambient lighting conditions results in significant variations in their adaptability to different environments.
[0004] Unlike visual perception, tactile perception-based terrain recognition methods determine terrain characteristics by sensing changes in the contact state between the robot and the ground. Existing tactile perception technologies mostly employ force sensors, contact sensors, or vibration sensors to capture signal responses generated by the contact surface for terrain recognition. However, current tactile perception methods primarily focus on time-domain analysis, neglecting the frequency-domain characteristics of the signal. This results in an inability to accurately capture subtle mechanical differences caused by different terrain surfaces, thus affecting the accuracy and robustness of terrain recognition. Summary of the Invention
[0005] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and device for terrain recognition of legged robots based on three-dimensional force perception.
[0006] To achieve the above objectives, according to one aspect of the present invention, a method for terrain recognition of a legged robot based on three-dimensional force perception is provided, comprising the following steps: S1: By integrating a force sensing module into the foot of the legged robot, three-dimensional force signals of the interaction between the robot's foot and the ground are obtained; S2: Perform filtering preprocessing on the force signal collected in step S1, extract the time-frequency domain features of the signal, and assign weight coefficients to each extracted feature component. S3: Based on the signal features processed in step S2, perform terrain recognition using a pre-trained classification algorithm and output the recognition results. Preferably, the method for extracting time-frequency domain features in step S2 is discrete wavelet transform or continuous wavelet transform. Specifically, this involves setting a window of a fixed size. N And at each time step Δ t The content of the updated window is used to extract signal features using discrete wavelet transform or continuous wavelet transform.
[0007] The method for extracting signal features using discrete wavelet transform is as follows: for a length of... N input signal F The discrete wavelet transform is performed using the following expression to extract the low-frequency characteristic approximation coefficients and high-frequency characteristic detail coefficients of the signal. The transform formula is as follows:
[0008] in, A j [ k [Is the] number j Layer, First k Approximation coefficients, It is the first j Layer, First k A detailed coefficient, j Indicates the number of decomposition layers. It is the first j Layer, First k A scaling function, ψ j,k It is the first j Layer, First k Wavelet functions, n Indicates the index of the signal. F [ n ] indicates the input signal F The Middle n Values.
[0009] The method for extracting signal features using continuous wavelet transform is as follows: Morlet wavelet or Daubechies wavelet is used as the wavelet basis function. The signal is analyzed by wavelets obtained through dilation or scaling at different scales and translation. The transform formula is as follows:
[0010] in, W F ( s , t ) represents the wavelet transform coefficients.F ( t ) is the input signal. s This is a scale parameter used to control frequency resolution. t The translation parameters are used to control the time resolution. ψ It is a wavelet function. ψ * yes ψ The complex conjugate function; For scale parameters s Set and determine the maximum scale parameter s max and minimum scale parameter s min This is to improve real-time extraction efficiency.
[0011] Preferably, in step S3, the classification algorithm is trained offline and the terrain classification model parameters are determined through optimization calculation. The optimization calculation is based on the K-nearest neighbor algorithm, decision tree, support vector machine, random forest, or neural network classifier to establish a computational model; specifically, it includes the following steps: S31: Perform motion experiments in various terrain environments and collect three-dimensional force data of the terrain recognition device throughout the entire motion cycle; S32: Divide the signals collected in step S31 into periods according to the actual situation, and manually label the signals of each period with real terrain labels; S33: Extract time-frequency domain features from the signal of each cycle in step S32; S34: The time-frequency domain features extracted in step S33 are distinguished into feature components originating from shear force in the X direction, shear force in the Y direction, and normal pressure in the Z direction according to their physical meaning, and different weight coefficients are assigned to each feature to adjust the contribution of force signals in different directions to terrain classification. S35: Based on the weighted features processed in step S34 as the model input, and combined with the real terrain labels marked in step S32, a classification algorithm is used for training to obtain the trained terrain classification algorithm parameters.
[0012] Preferably, in step S31, the motion experiment involves installing the terrain recognition device on the foot of a legged robot or the end of a multi-dimensional mobile platform, and controlling the robot or mobile platform to conduct motion experiments under different speeds, accelerations, and terrain conditions.
[0013] Preferably, in step S32, in the legged robot, the period division is to divide each step of motion into one period, or to take a period of length... L The window is divided into a cycle.
[0014] Preferably, the method for extracting the time-frequency domain signal in step S33 is discrete wavelet transform or continuous wavelet transform, and the extracted time-frequency domain features include wavelet transform coefficients and power spectrum.
[0015] Preferably, the weighted fused feature vector in step S34 is represented as follows:
[0016] in, F x This is a subset of shear force features in the X direction. F y This is a subset of shear force features in the Y direction. F z Z-direction normal pressure feature subset ,a、b、c These are the weighting coefficients of each component, and satisfy... α + β + c =1.
[0017] To achieve the above objectives, according to another aspect of the present invention, a footed robot terrain recognition device based on three-dimensional force perception is provided, employing the aforementioned footed robot terrain recognition method, including a robot foot module, wherein the robot foot module sequentially includes a connection module, a circuit module, a force perception module, and a load-bearing module; the overall configuration of the robot foot module is one of spherical, square, cylindrical, prismatic, conical, or biomimetic shapes; The bearing module is in contact with the environment, and its surface is provided with microstructures. When it comes into contact with different terrain surfaces, it can cause different degrees of mechanical response due to the different terrain textures, and transmit the acquired mechanical response to the force sensing module. The force sensing module is used to sense the mechanical response transmitted by the load-bearing module and convert it into a piezoresistive, piezoelectric, or electromagnetic physical quantity change signal. The circuit module includes a microcontroller and a processing circuit, which are used to receive signals transmitted by the force sensing module, calculate them into three-dimensional forces, and transmit and save the data. The connection module is used to connect other modules and has a reserved interface for connecting the robot foot module to the legged robot.
[0018] In summary, the technical solutions conceived by this invention have the following beneficial effects compared with the prior art: This invention provides a terrain recognition method for legged robots based on three-dimensional force perception. Based on tactile perception, it acquires the three-dimensional contact force between the robot and the ground, enabling stable terrain recognition in low-light environments, thus overcoming the perception bottleneck of vision systems in low-light or extreme environments. This allows the robot to maintain high recognition accuracy and adaptability in changing and complex environments. By integrating a simple force perception module, the hardware cost and energy consumption of the robot system are significantly reduced. By employing an offline-trained classification algorithm combined with real-time tactile perception, the robot can quickly and accurately determine the terrain in real-time environments, enhancing its practicality and efficiency in complex tasks.
[0019] Furthermore, this invention employs wavelet transform for time-frequency domain analysis of force signals, which, compared to traditional time-domain analysis methods, can extract terrain features more comprehensively. By extracting the time-frequency domain features of the signal, it can capture the subtle mechanical differences caused by different terrain surfaces, thereby helping to improve the accuracy of terrain identification. Attached Figure Description
[0020] Figure 1 This is a flowchart of a terrain recognition method for legged robots based on three-dimensional force perception.
[0021] Figure 2 This is an architecture diagram of a footed robot terrain recognition device based on three-dimensional force perception.
[0022] Figure 3 This is a schematic diagram of a multi-terrain motion experiment method based on a three-dimensional mobile platform.
[0023] Figure 4 This is a schematic diagram of three-dimensional force time-domain signals under different terrains.
[0024] Figure 5 This is a comparison chart of the power spectrum of signals after wavelet transform under different terrain conditions.
[0025] Figure 6 This is a comparison of the signal power spectrum after wavelet transform of PVC terrain under dry and wet conditions. Detailed Implementation
[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0027] Example 1 Please see Figure 1 and Figure 2The present invention provides a method for terrain recognition of a legged robot based on three-dimensional force perception, comprising the following steps: S1: By integrating a force sensing module into the foot of the legged robot, three-dimensional force signals of the interaction between the robot's foot and the ground are obtained; S2: Perform filtering preprocessing on the force signal collected in step S1, extract the time-frequency domain features of the signal, and assign weight coefficients to each extracted feature component. S3: Based on the signal features processed in step S2, perform terrain recognition using a pre-trained classification algorithm and output the recognition results. The method for extracting time-frequency domain features in step S2 is either discrete wavelet transform or continuous wavelet transform. Specifically, this involves setting a window of a fixed size. N And at each time step Δ t The content of the updated window is used to extract signal features using discrete wavelet transform or continuous wavelet transform.
[0028] The method for extracting signal features using discrete wavelet transform is as follows: for a length of... N input signal F The discrete wavelet transform is performed using the following expression to extract the low-frequency characteristic approximation coefficients and high-frequency characteristic detail coefficients of the signal. The transform formula is as follows:
[0029] in, A j [ k [Is the] number j Layer, First k Approximation coefficients, It is the first j Layer, First k A detailed coefficient, j Indicates the number of decomposition layers. It is the first j Layer, First k A scaling function, ψ j,k It is the first j Layer, First k Wavelet functions, n Indicates the index of the signal. F [ n ] indicates the input signal F The Middle n Values.
[0030] The method for extracting signal features using continuous wavelet transform is as follows: Morlet wavelet or Daubechies wavelet is used as the wavelet basis function. The signal is analyzed by wavelets obtained through dilation or scaling at different scales and translation. The transform formula is as follows:
[0031] in, W F ( s , t ) represents the wavelet transform coefficients. F ( t ) is the input signal. s This is a scale parameter used to control frequency resolution. t The translation parameters are used to control the time resolution. ψ It is a wavelet function. ψ * yes ψ The complex conjugate function; For scale parameters s Set and determine the maximum scale parameter s max and minimum scale parameter s min This is to improve real-time extraction efficiency.
[0032] In step S3, the classification algorithm is trained offline and the terrain classification model parameters are determined through optimization calculation. The optimization calculation is based on the K-nearest neighbor algorithm, decision tree, support vector machine, random forest, or neural network classifier to establish a computational model; specifically, it includes the following steps. S31: Perform motion experiments in various terrain environments and collect three-dimensional force data of the terrain recognition device throughout the entire motion cycle; S32: Divide the signals collected in step S31 into periods according to the actual situation, and manually label the signals of each period with real terrain labels; S33: Extract time-frequency domain features from the signal of each cycle in step S32; S34: The time-frequency domain features extracted in step S33 are distinguished into feature components originating from shear force in the X direction, shear force in the Y direction, and normal pressure in the Z direction according to their physical meaning, and different weight coefficients are assigned to each feature to adjust the contribution of force signals in different directions to terrain classification. S35: Based on the weighted features processed in step S34 as the model input, and combined with the real terrain labels marked in step S32, a classification algorithm is used for training to obtain the trained terrain classification algorithm parameters.
[0033] In step S31, the motion experiment involves installing the terrain recognition device on the foot of a legged robot or the end of a multi-dimensional mobile platform, and controlling the robot or mobile platform to conduct motion experiments under different speeds, accelerations, and terrain conditions.
[0034] In step S32, in a legged robot, period division means dividing each step of motion into one period, or taking a period of length... L The window is divided into a cycle. For custom mobile platforms, the cycle division can be adjusted according to the platform's motion characteristics, and the cycle length can be customized according to the platform's step size and motion mode.
[0035] Preferably, the method for extracting the time-frequency domain signal in step S33 is discrete wavelet transform or continuous wavelet transform, and the extracted time-frequency domain features include wavelet transform coefficients and power spectrum.
[0036] The weighted fused feature vector in step S34 is represented as follows:
[0037] in, F x This is a subset of shear force features in the X direction. F y This is a subset of shear force features in the Y direction. F z Z-direction normal pressure feature subset ,a、b、c These are the weighting coefficients of each component, and satisfy... α + β + c =1.
[0038] By setting different weight values, the role of force signals in specific directions in classification can be flexibly enhanced. To emphasize the perception of tangential force, a higher weight can be set. α and β Value; if the discriminative significance of a force signal in a certain direction is weak in a specific terrain, its weight can also be set low. In particular, when α =0 and β When = 0, the system degenerates into a traditional method that relies solely on positive pressure characteristics; however, this invention utilizes non-zero = 0. α and β Assigning values enables explicit modeling of friction and slip characteristics in different directions, thereby improving the discrimination ability of classification algorithms under complex terrain.
[0039] Example 2 The present invention also provides a footed robot terrain recognition device based on three-dimensional force perception, which adopts the footed robot terrain recognition method in Embodiment 1, including a robot foot module. The robot foot module includes a connection module, a circuit module, a force perception module and a load-bearing module in sequence. The overall configuration of the robot foot module is one of the following: spherical, square, cylindrical, prismatic, conical or biomimetic. The load-bearing module is in contact with the environment and has microstructures on its surface. When it comes into contact with different terrain surfaces, it can cause different degrees of mechanical response due to different terrain textures, such as different vibration frequencies and degrees of deformation. The acquired mechanical response is then transmitted to the force sensing module. The force sensing module is used to sense the mechanical response transmitted by the load-bearing module and convert it into a piezoresistive, piezoelectric, or electromagnetic physical quantity change signal. The circuit module includes a microcontroller and a processing circuit, which are used to receive signals transmitted by the force sensing module, calculate them into three-dimensional forces, and transmit and save the data. The connection module is used to connect other modules and has a reserved interface for connecting the robot foot module to the legged robot.
[0040] The classification algorithm is integrated into the microcontroller of the circuit module and then connected to the main control module of the legged robot; alternatively, it can be directly deployed in the main control module of the robot.
[0041] The following presents a multi-terrain motion experiment method based on a three-dimensional mobile platform. Figure 3 A schematic diagram of the method is provided. The experiment includes the following steps: (1) Install the terrain recognition device at the end of the three-dimensional mobile platform.
[0042] (2) Fix a terrain material onto the surface of a three-dimensional moving platform.
[0043] (3) Control the three-dimensional moving platform to make the terrain recognition device contact the terrain material and maintain a certain pressure.
[0044] (4) Control the three-dimensional moving platform to move forward at a constant speed of 2 mm / s for 50 mm along the direction shown in the figure. During the experiment, the circuit module reads and saves the motion experiment data.
[0045] (5) Change the terrain material and repeat steps (3)-(4).
[0046] Secondly, the multi-terrain motion experiment method described above was used to collect motion experiment data of the terrain recognition device in four types of terrain: wood, aluminum alloy, glass, and PVC. For example... Figure 4 The diagram shows the three-dimensional force time-domain signals acquired by the terrain recognition device under different terrain conditions in this experiment. (Controlling the normal force acting on the ground) F z When the frictional force is a constant value, the frictional force in the direction of travel is... F x The numerical values differ, according to the formula for sliding friction. This study preliminarily verified that different materials have different coefficients of friction.
[0047] The following steps are taken to perform time-frequency analysis of three-dimensional force signals on different terrain materials using continuous wavelet transform: (1) Perform signal preprocessing to remove outliers from the data and subtract the signal mean.
[0048] (2) Controlling the calculation scale s The range, determining the maximum scale s max and minimum scale s min .
[0049] (3) Perform continuous wavelet transform on the data, where the wavelet function uses the Morlet wavelet, and obtain the wavelet transform coefficients. W(s,τ) This indicates that the signal is at a certain scale. s and time location t The value below.
[0050] The wavelet scale affects signal analysis by compressing or expanding the wavelet function. A larger scale results in a wavelet covering a longer time interval (lower frequency components), while a smaller scale results in a narrower wavelet (higher frequency components). Morlet wavelets are widely used in time-frequency analysis, especially for non-stationary signals, exhibiting good localization properties in both the time and frequency domains.
[0051] Based on the obtained wavelet transform coefficients, the friction force in the forward direction is extracted. F x Calculate the power spectrum of the signal based on its characteristics. For example... Figure 5 The image shows a comparison of the signal power spectrum after wavelet transform under different terrains. The results indicate that after removing the signal mean, the signal under the aluminum alloy material terrain exhibits a relatively stable frequency variation, with a generally smaller power spectrum amplitude. Wood, glass, and PVC show similar time-frequency distribution characteristics over one cycle, but the power spectrum modulus amplitude of wood is much smaller than that of glass and PVC. PVC, however, has a larger modulus amplitude in the high-frequency components (3-5Hz) compared to glass.
[0052] The above results demonstrate that the terrain recognition device interacts with different terrains, and the characteristics exhibit significant differences after time-frequency analysis. These differences can serve as the basic input for the next step of terrain classification.
[0053] Finally, using the multi-terrain motion experiment method described above, motion data of the terrain recognition device were collected under both dry and wet terrain conditions in PVC terrain. The same feature extraction method as described above was employed to obtain the wavelet transform coefficients and other features of the three-dimensional force signal under these experimental conditions.
[0054] like Figure 6The image shows a comparison of the signal power spectrum after wavelet transform of PVC terrain under dry and wet conditions. The overall power amplitude distribution in the time-frequency diagram is similar under both terrain conditions, but the amplitude across the entire frequency band is significantly higher under the wet terrain. This result indicates that the tangential force on the robot's foot changes considerably under wet conditions, leading to an increase in the energy of the high-frequency components (2-5Hz) in the signal. Furthermore, this result further verifies the feasibility of using terrain recognition equipment to identify the same material under different degrees of dryness and wetness.
[0055] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for terrain recognition of a legged robot based on three-dimensional force sensing, characterized by, The method comprises the following steps: S1: acquiring three-dimensional force signals of the interaction between the robot foot and the contacted ground through a force sensing module integrated at the foot end of the foot robot; S2: filtering and preprocessing the force signals collected in step S1, extracting time-frequency domain features of the signals, and assigning weight coefficients to each feature component; S3: based on the signal features processed in step S2, using a pre-trained classification algorithm for terrain recognition, and outputting the recognition result.
2. The method of claim 1, wherein, The time-frequency domain feature extraction method in step S2 is discrete wavelet transform or continuous wavelet transform, and the specific method is as follows: a fixed size window is set N , and the window content is updated at each time step Δ t , and the signal features are extracted by using discrete wavelet transform or continuous wavelet transform.
3. The method of claim 2, wherein the method is based on three-dimensional force sensing. The method for extracting signal features by the discrete wavelet transform is: for an input signal with a length of N F , the following expression is used for the discrete wavelet transform to extract the low-frequency band feature approximation coefficient and the high-frequency band feature detail coefficient, and the transform formula is: wherein A j [ k ] is the first j layer, the first k approximation coefficient, is the first j layer, the first k detail coefficient, j denotes the number of decomposition layers, is the first j layer, the first k scale function, The method for extracting signal features by continuous wavelet transform is: using Morlet wavelet or Daubechies wavelet as wavelet base function, analyzing the signal by subwave obtained by expansion or stretching of different scales and translation, and the transform formula is: j,k is the first j layer, the first k wavelet function, n denotes the index of the signal, F [ n ] denotes the input signal F in the first n value; The method for extracting signal features by continuous wavelet transform is: using Morlet wavelet or Daubechies wavelet as wavelet base function, analyzing the signal by subwave obtained by expansion or stretching of different scales and translation, and the transform formula is: in, W F ( s , The classification algorithm in step S3 is trained offline and determines the terrain classification model parameters through optimization calculation, and the optimization calculation is based on K nearest neighbor algorithm, decision tree, support vector machine, random forest or neural network classifier to establish a calculation model; it specifically includes the following steps: ) represents the wavelet transform coefficients. F ( t ) is the input signal. s This is a scale parameter used to control frequency resolution. S31: performing motion experiment under multiple terrain environments, and collecting three-dimensional force data of the terrain recognition device in the whole motion cycle; The translation parameters are used to control the time resolution. S32: dividing the signals collected in step S31 into cycles according to the actual situation, and manually labeling each cycle signal to label the real terrain label; It is a wavelet function. S33: extracting time-frequency domain features of each cycle signal in step S32; * yes S34: according to the physical meaning, the time-frequency domain features extracted in step S33 are divided into feature components derived from X direction shear force, Y direction shear force and Z direction normal pressure, and different weight coefficients are assigned to each direction feature to adjust the contribution degree of different direction force signals to terrain classification; The complex conjugate function; For the scale parameter s , the maximum scale parameter s max and the minimum scale parameter s min are set to improve the real-time extraction efficiency.
4. The method of claim 1, wherein the method is based on three-dimensional force sensing. S35: based on the weighted features processed in step S34, the real terrain label labeled in step S32 is combined, and the classification algorithm is trained to obtain the trained terrain classification algorithm parameters. In step S31, the motion experiment is to install the terrain recognition device at the foot end of the foot robot or the end of the multi-dimensional moving platform, and to realize the motion experiment under different speed, acceleration and terrain conditions by controlling the robot or the moving platform. In step S33, the method for extracting time-frequency domain signal is discrete wavelet transform or continuous wavelet transform, and the extracted time-frequency domain features include wavelet transform coefficients and power spectrum. In step S34, the feature vector after weighted fusion is represented as: , α, β, γ In step S34, the feature vector after weighted fusion is represented as:
5. The method of claim 4, wherein the method is based on three-dimensional force sensing. The robot foot module comprises a connecting module, a circuit module, a force sensing module and a bearing module.
6. The method of claim 4, wherein the method is based on three-dimensional force sensing. In the foot robot in step S32, the cycle division is to divide the motion of each step into a cycle, or to divide a window with a length of L into a cycle.
7. The method of claim 4, wherein the method is based on three-dimensional force sensing. The robot foot module has one of the following overall configurations: spherical, square, cylindrical, prismatic, conical or bionic; 8. The method of claim 4, wherein the method is based on three-dimensional force sensing. The bearing module is in contact with the environment, and the surface thereof is provided with a microstructure, which can cause different degrees of mechanical response when contacting different terrain surfaces, and transmit the acquired mechanical response to the force sensing module; wherein, F x is a subset of features for X-direction shear force, F y is a subset of features for Y-direction shear force, F z is a subset of features for Z-direction normal force The force sensing module is used to sense the mechanical response transmitted by the bearing module and convert it into a piezoresistive, piezoelectric or electromagnetic physical quantity change signal; are weight coefficients for each component, and satisfy α + β + The circuit module comprises a single-chip microcomputer and a processing circuit, which is used to receive the signal transmitted by the force sensing module, calculate three-dimensional force, and transmit and save data. = 1.
9. A three-dimensional force-sensing based legged robot terrain recognition apparatus employing the legged robot terrain recognition method according to any one of claims 1 to 8, characterized by, 10. The three-dimensional force sensing based legged robot terrain identification device of claim 9, wherein, The connecting module is used for connecting other modules and reserving an interface for connecting the robot foot module with the foot-type robot.