A method and system for terrain perception and motion state recognition of quadruped robots based on multidimensional accelerometer sensors
By installing multidimensional accelerometers on quadruped robots and combining them with hierarchical classifiers, the problems of sensor damage and signal interference were solved, enabling high-precision recognition of complex environmental terrain and motion states, and improving the autonomous navigation and gait control capabilities of quadruped robots.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-02-09
- Publication Date
- 2026-06-30
AI Technical Summary
Existing quadruped robots struggle to achieve high-precision terrain recognition and motion state decoupling in complex, unstructured environments. Their sensors are easily damaged, and their signals are susceptible to interference, failing to meet the real-time perception requirements of high-performance control algorithms.
Employing a six-dimensional piezoelectric accelerometer and four three-dimensional piezoelectric accelerometers, combined with wavelet denoising and low-pass filtering for signal processing, terrain perception and motion state recognition are achieved through a two-level hierarchical classifier. Multi-dimensional feature vectors are constructed, and terrain category, gait pattern, and motion state are output.
Significantly reduces hardware costs, improves durability and maintenance-free capability, enables high-precision terrain and motion state recognition in complex environments, ensures accurate perception under harsh conditions, and provides rapid-response autonomous navigation and gait regulation.
Smart Images
Figure CN122308353A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent quadruped robot technology, and in particular to a method and system for terrain perception and motion state recognition of quadruped robots based on multidimensional acceleration sensors. Background Technology
[0002] Quadruped robots, with their exceptional mobility, have demonstrated irreplaceable application value in fields such as disaster search and rescue, military reconnaissance, planetary exploration, and power line inspection. To maintain stability in complex unstructured environments such as grasslands, gravel, mud, and slippery surfaces, quadruped robots need precise terrain recognition, slip detection, and gait phase segmentation capabilities, which heavily rely on their ability to perceive ground contact information. Current technologies primarily employ methods to achieve these perception functions, such as installing six-dimensional force sensors on the feet, integrating force / torque sensors into a single leg structure, or mounting a single inertial measurement unit (IMU) in the torso. However, these solutions suffer from significant engineering application drawbacks:
[0003] A. Foot sensors are in direct contact with the ground for a long time, making them extremely susceptible to impact from sand and gravel, immersion in water, and corrosion from mud, which can damage the sensors or cause reading drift, severely limiting the reliability of quadruped robots in harsh outdoor environments.
[0004] B. High-precision force sensors are expensive, and the limited space at the foot end makes wiring complex and signal transmission susceptible to interference. They also increase the inertia at the end of the legs, affecting the dynamic performance of the quadruped robot.
[0005] C. The torso IMU is far from the ground contact point, and the signal attenuates and becomes noisy after passing through the leg structure. This means that this solution can usually only distinguish macroscopic terrain features such as flat ground and slopes, and it is difficult to finely distinguish specific terrain textures such as asphalt roads, grass, and gravel roads. Relying solely on torso inertial data makes it difficult to decouple complex motion features. In particular, when quadruped robots switch between multiple motion states such as forward movement, lateral movement, and turning in place, it is difficult to achieve high-precision gait phase segmentation and terrain classification simultaneously, which cannot meet the real-time perception requirements of high-performance control algorithms.
[0006] Therefore, there is an urgent need to develop a method and system for terrain perception and motion state recognition of quadruped robots based on multidimensional accelerometer sensors. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for terrain perception and motion state recognition of quadruped robots based on multi-dimensional accelerometer sensors, so as to solve the problems existing in the prior art.
[0008] The technical solution adopted to achieve the purpose of this invention is as follows: a method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer sensor, comprising the following steps:
[0009] S1) Signal Acquisition. Acceleration sequence signals of the quadruped robot's torso and legs under different motion trajectories are synchronously acquired at a set frequency using a six-dimensional piezoelectric accelerometer and four three-dimensional piezoelectric accelerometers. The six-dimensional piezoelectric accelerometer is mounted on the back of the quadruped robot's torso. The four three-dimensional piezoelectric accelerometers are respectively mounted at the hip joints of each leg.
[0010] S2) Signal preprocessing. Wavelet denoising and low-pass filtering are performed on all acquired channel signals to remove high-frequency noise.
[0011] S3) Feature Extraction. The preprocessed signal is segmented using a sliding time window, and the temporal features within each time window are extracted to construct a multidimensional feature vector.
[0012] S4) Terrain perception. The multidimensional feature vector is input into the trained first-level classifier to identify the terrain category where the quadruped robot is currently located.
[0013] S5) Decoupling of motion state. Based on the terrain category output in step S4), index and call the corresponding second-level classifier. Input the multidimensional feature vector into the second-level classifier, and output the current gait pattern, motion state, and / or gait phase.
[0014] S6) Status Output. Based on the terrain type, gait pattern, motion state, and gait phase, generate and output the status flag of the quadruped robot.
[0015] Furthermore, the measurement center of the six-dimensional piezoelectric accelerometer coincides with or has a preset compensation offset from the center of mass of the quadruped robot's torso. The Z-axis of the three-dimensional piezoelectric accelerometer is perpendicular to the ground when the quadruped robot is standing.
[0016] Furthermore, in step S3), the extracted time-domain features include at least peak-to-peak value and standard deviation.
[0017] Furthermore, time-domain features also include one or more of absolute mean, skewness, and kurtosis.
[0018] Furthermore, in step S3), the length of the sliding time window includes at least two gait cycles.
[0019] Furthermore, the first-level classifier and the second-level classifier are constructed using any one of feedforward neural networks, support vector machines, or random forests.
[0020] Furthermore, in step 8), the motion state output by the second-level classifier includes forward, backward, omnidirectional translation, and turning in place in the Walk gait. In the Trot gait, it includes at least forward and backward. The gait phase includes at least a support phase and a swing phase.
[0021] The present invention also discloses a quadruped robot terrain and motion state recognition system without foot force sensors, comprising: a sensor module, a data acquisition and signal conditioning module, and a processor module.
[0022] The sensor module includes a six-dimensional piezoelectric accelerometer mounted on the back of the quadruped robot's torso, and four three-dimensional piezoelectric accelerometers mounted on the hip joints of the four legs.
[0023] The data acquisition and signal conditioning module is used to synchronously acquire the channel signals output by the sensor module.
[0024] The processor module is configured to receive the collected signals and execute steps S2) to S6) of the method described above, and output the terrain category and motion status results.
[0025] Furthermore, the six-dimensional piezoelectric accelerometer is mounted at the center of the circular mounting area at the front end of the quadruped robot's back load structure. The three-dimensional piezoelectric accelerometer is mounted at the reference point of the hip joint screw recess.
[0026] The present invention also discloses a quadruped robot, including a quadruped robot terrain and motion state recognition system without foot end force sensors as described above.
[0027] The technical effects of this invention are beyond doubt:
[0028] A. Significantly reduces hardware costs and improves the durability and maintenance-free capability of quadruped robots in harsh environments such as mud and water wading;
[0029] B. The innovative use of sensor array layout combined with a two-level hierarchical classification architecture effectively solves the problem of the coupling and difficulty in distinguishing terrain and motion state characteristics in complex environments;
[0030] C. Based entirely on high-frequency acceleration signals, it is unaffected by lighting conditions, smoke, or rain, and can still accurately perceive even at night or in extreme scenarios where visual sensors fail. The system can simultaneously output terrain type, motion trend, and gait phase in real time, providing extremely fast response speed for the quadruped robot's autonomous navigation and gait control. Attached Figure Description
[0031] Figure 1 This is a diagram of the system hardware framework.
[0032] Figure 2This is a schematic diagram showing the sensor installation location;
[0033] Figure 3 Here is a flowchart of the hierarchical classification algorithm;
[0034] Figure 4 For terrain identification, a confusion matrix diagram is generated.
[0035] Figure 5 This is a ranking map of terrain feature importance based on MLP.
[0036] Figure 6 This is a confusion matrix diagram for identification before and after;
[0037] Figure 7 This is a ranking diagram of feature importance before and after MLP.
[0038] Figure 8 A confusion matrix diagram for left and right turns;
[0039] Figure 9 This is a ranking diagram of the importance of left and right turn features based on MLP.
[0040] Figure 10 This is a confusion matrix diagram that has been shifted left and right.
[0041] Figure 11 This is a ranking diagram of the importance of left and right translation features based on MLP.
[0042] The diagram shows: a six-dimensional piezoelectric accelerometer 1, a three-dimensional piezoelectric accelerometer 2, a data acquisition and signal conditioning module 3, and a processor module 4. Detailed Implementation
[0043] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0044] Example 1:
[0045] This embodiment aims to solve the problem of decoupling multimodal motion states of quadruped robots in unstructured environments. See also... Figure 3 This embodiment provides a method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer sensor, including the following steps:
[0046] S1) Signal Acquisition. Acceleration sequence signals of the quadruped robot's torso and legs under different motion trajectories are synchronously acquired at a set frequency using a six-dimensional piezoelectric accelerometer 1 and four three-dimensional piezoelectric accelerometers 2. The six-dimensional piezoelectric accelerometer 1 is mounted on the back of the quadruped robot's torso. The six-dimensional piezoelectric accelerometer 1 is used to measure the three-axis acceleration and three-axis angular acceleration of the torso in real time. The four three-dimensional piezoelectric accelerometers 2 are respectively mounted at the hip joints of each leg. The three-dimensional piezoelectric accelerometers 2 are used to measure the three-axis acceleration at the hip joints of each leg in real time.
[0047] S2) Signal preprocessing. Wavelet denoising and low-pass filtering are performed on all acquired channel signals to remove high-frequency noise.
[0048] S3) Feature Extraction. The preprocessed signal is segmented using a sliding time window, and the temporal features within each time window are extracted to construct a multidimensional feature vector.
[0049] S4) Terrain perception. The multidimensional feature vector is input into the trained first-level classifier to identify the terrain category where the quadruped robot is currently located.
[0050] S5) Decoupling of motion state. Based on the terrain category output in step S4), index and call the corresponding second-level classifier. Input the multidimensional feature vector into the second-level classifier, and output the current gait pattern, motion state, and / or gait phase.
[0051] S6) Status Output. Based on the terrain type, gait pattern, motion state, and gait phase, generate and output the status flag of the quadruped robot.
[0052] This embodiment deploys multi-axis accelerometers on the body and feet of a quadruped robot to collect high-frequency inertial sequence signals in real time under different motion trajectories. The collected raw data is first preprocessed and subjected to temporal feature mapping to construct a multi-dimensional feature vector library containing terrain interaction impact features, periodic gait features, and spatial pose evolution features. This embodiment utilizes heterogeneous information fusion from 3D and 6-axis accelerometers. A two-level hierarchical classification architecture is used to decouple the output of omnidirectional motion trends, gait phase, and adaptability to complex terrain. This embodiment solves the problems of coupled motion states and low feature recognition accuracy in quadruped robots in complex unstructured environments. Experiments demonstrate that this method has high recognition accuracy and real-time response speed under various terrain and gait intertwining conditions, significantly improving the environmental adaptability and intelligent gait control level of quadruped robots during autonomous navigation.
[0053] Example 2:
[0054] The main content of this embodiment is the same as that of Embodiment 1, except that the measurement center of the six-dimensional piezoelectric accelerometer 1 coincides with or has a preset compensation offset with the center of mass of the quadruped robot's torso. The Z-axis of the three-dimensional piezoelectric accelerometer 2 is perpendicular to the ground when the quadruped robot is standing.
[0055] Example 3:
[0056] The main content of this embodiment is the same as that of Embodiment 1 or 2, wherein, in step S3), the extracted time-domain features include at least peak-to-peak value and standard deviation. In actual production, the time-domain features also include one or more of absolute mean, skewness, and kurtosis.
[0057] Example 4:
[0058] The main content of this embodiment is the same as any one of embodiments 1 to 3, wherein, in step S3), the length of the sliding time window includes at least two gait cycles.
[0059] Example 5:
[0060] This embodiment is essentially the same as any one of embodiments 1-4, except that the first-level classifier and the second-level classifier are constructed using any one of feedforward neural networks, support vector machines, or random forests. The selection of the second-level classifier is based on the real-time indexing of the terrain semantic labels output by the first-level classifier. The processor module is integrated into a NIPXIe acquisition system or an embedded industrial computer, running a LabVIEW real-time acquisition program and a pre-trained classification model in MATLAB or Python.
[0061] This embodiment constructs a hierarchical classification architecture based on a multilayer perceptron. High-order statistical moment features of the preprocessed signal are extracted to construct a heterogeneous feature matrix. First-level identification: The feature matrix is input into the first-level terrain perception model, and the terrain probability distribution is output through a Softmax layer to identify the current terrain (e.g., asphalt road, grass, or gravel road). Second-level decoupling: Based on the identified terrain context, the system automatically indexes and attaches the corresponding action recognition sub-model. Utilizing the kinetic energy evolution features within the gait cycle, the specific motion state (e.g., forward, backward, left / right turning, or lateral translation) is decoupled and output. Simultaneously with the above state recognition, real-time impact detection is performed on the Z-axis signals of the four hip 3D sensors. When the peak value of the detected Z-axis signal exceeds a set threshold, it is determined as the ground contact moment; when the signal rapidly decreases after the peak and falls below the threshold, it is determined as the ground lift moment. This step achieves accurate segmentation of the four-leg support phase and swing phase, and experiments verify that its segmentation error is less than 20ms.
[0062] Appendix Figures 4-11This data provides comprehensive validation of the performance of the hierarchical classification architecture described above, primarily characterizing the model's recognition accuracy and feature interpretability. Specifically, it can be divided into four groups, each containing a confusion matrix and a feature importance ranking diagram:
[0063] Figure 4 , Figure 5 Perform terrain recognition verification. Figure 4 The representation model's accuracy in distinguishing different terrains (such as asphalt, grass, and gravel). The darker the color of the matrix's diagonal and the larger the value, the fewer the false positives. Figure 5 The study identifies which sensor features play a key role in terrain recognition tasks, demonstrating that the model classifies based on physical laws rather than noise.
[0064] Figure 6 , Figure 7 Verify the motion before and after. Figure 6 , Figure 7 The classification accuracy of the representation model in distinguishing between forward and backward motion, and the corresponding core contributing features.
[0065] Figure 8 , Figure 9 Perform steering motion verification. Figure 8 , Figure 9 The accuracy of the model in recognizing left and right turns and the weights of key features are represented.
[0066] Figure 10 , Figure 11 Perform lateral translation verification. Figure 10 , Figure 11 The accuracy of the representation model in recognizing the robot's left and right lateral movements and the corresponding ranking of feature importance.
[0067] Example 6:
[0068] The main content of this embodiment is the same as any one of embodiments 1 to 5. In step 8), the motion state output by the second-level classifier includes forward, backward, omnidirectional translation, and turning in place in the Walk gait. In the Trot gait, it includes at least forward and backward.
[0069] The gait phase includes at least a support phase and a swing phase. The gait phase is determined by one of the following methods:
[0070] a) Perform threshold detection or zero-crossing point detection on the Z-axis signal of the hip joint three-dimensional accelerometer;
[0071] b) Detect the moment when the Z-axis signal peak occurs.
[0072] Example 7:
[0073] This embodiment is essentially the same as any one of embodiments 1-6, except that all channel signals are preprocessed sequentially with wavelet noise reduction and low-pass filtering with a cutoff frequency not exceeding 50Hz. The window length is 800-1500ms, and the overlap rate between adjacent windows is 70%-90%. The first-stage classifier is capable of distinguishing at least two of the following: asphalt road surface, grass, and gravel road surface. All sensors are fixed with shock-absorbing pads, and the signal lines are protected by shielded cables or flexible conduits.
[0074] Using an NI PXIe-1083 acquisition card and two PST-SC18 constant current sources, raw acceleration signals from a total of 20 channels of the aforementioned sensors were simultaneously acquired at a sampling frequency of 1000Hz. Two-step preprocessing was performed on the acquired raw signals: first, wavelet thresholding with a 4-level decomposition using a db8 wavelet basis was applied to remove background noise; then, a fourth-order Butterworth low-pass filter with a cutoff frequency of 40Hz was used to eliminate high-frequency jitter interference, resulting in a clean inertial signal sequence.
[0075] Example 8:
[0076] See Figure 1 and Figure 2 This embodiment provides a terrain and motion state recognition system for a quadruped robot without foot force sensors, including: a sensor module, a data acquisition and signal conditioning module 3, and a processor module 4.
[0077] The sensor module includes a six-dimensional piezoelectric accelerometer 1 mounted on the back of the quadruped robot's torso, and four three-dimensional piezoelectric accelerometers 2 mounted on the hip joints of the four legs.
[0078] The data acquisition and signal conditioning module 3 is used to synchronously acquire 20 channels of signals output by the sensor module at a sampling frequency of not less than 500Hz.
[0079] The processor module 4 is configured to receive the collected signals and execute the method steps S2) to S6) as described in any one of Embodiments 1 to 7, and output the terrain category and motion status results.
[0080] Example 9:
[0081] The main content of this embodiment is the same as that of embodiment 8. The six-dimensional piezoelectric accelerometer 1 is installed at the center of the circular mounting area at the front end of the load structure on the back of the quadruped robot. This center is 100-120mm laterally spaced from the three-dimensional sensor on the right hip. The three-dimensional piezoelectric accelerometer 2 is installed at the reference point of the hip joint screw recess, with a coordinate offset of (13±5mm, -15±5mm).
[0082] Example 10:
[0083] This embodiment provides a quadruped robot, including a quadruped robot terrain and motion state recognition system without foot-end force sensors as described in Embodiment 8 or 9.
Claims
1. A method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional acceleration sensor, characterized in that, Includes the following steps: S1) Signal acquisition; using a six-dimensional piezoelectric accelerometer (1) and four three-dimensional piezoelectric accelerometers (2) at a set frequency to synchronously acquire acceleration sequence signals of the torso and each leg of the quadruped robot under different motion trajectories; wherein, the six-dimensional piezoelectric accelerometer (1) is installed on the back of the quadruped robot torso; the four three-dimensional piezoelectric accelerometers (2) are respectively installed on the hip joints of each leg; S2) Signal preprocessing: Wavelet denoising and low-pass filtering are performed on all acquired channel signals to remove high-frequency noise; S3) Feature extraction: The preprocessed signal is segmented using a sliding time window, and the temporal features within each time window are extracted to construct a multidimensional feature vector; S4) Terrain perception; The multi-dimensional feature vector is input into the trained first-level classifier to identify the terrain category where the quadruped robot is currently located; S5) Decoupling of motion state; Based on the terrain category output in step S4), index and call the corresponding second-level classifier; Input the multidimensional feature vector into the second-level classifier, and output the current gait pattern, motion state and / or gait phase; S6) Status Output: Based on the terrain type, gait pattern, motion state, and gait phase, generate and output the status flag of the quadruped robot. 2.The method of claim 1, wherein the method further comprises: The measurement center of the six-dimensional piezoelectric accelerometer (1) coincides with the center of mass of the quadruped robot's torso or has a preset compensation offset; the Z-axis of the three-dimensional piezoelectric accelerometer (2) is perpendicular to the ground when the quadruped robot is standing.
3. The method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer according to claim 1, characterized in that: In step S3), the extracted time-domain features include at least peak-to-peak value and standard deviation.
4. The method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer according to claim 3, characterized in that: Temporal features also include one or more of absolute mean, skewness, and kurtosis.
5. The method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer sensor according to claim 1, characterized in that: In step S3), the length of the sliding time window includes at least two gait cycles.
6. The method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer sensor according to claim 1, characterized in that: The first-level classifier and the second-level classifier are constructed using any one of the following: feedforward neural network, support vector machine, or random forest.
7. The method for terrain perception and motion state recognition of a quadruped robot based on a multi-dimensional accelerometer according to claim 1, characterized in that: In step 8), the motion state output by the second-level classifier includes forward, backward, omnidirectional translation and turning in place in the Walk gait; and at least forward and backward in the Trot gait; the gait phase includes at least the support phase and the swing phase.
8. A terrain and motion state recognition system for a quadruped robot without foot-end force sensors, characterized in that, include: Sensor module, data acquisition and signal conditioning module (3) and processor module (4); The sensor module includes a six-dimensional piezoelectric accelerometer (1) installed on the back of the quadruped robot's torso, and four three-dimensional piezoelectric accelerometers (2) installed on the hip joints of the four legs respectively. The data acquisition and signal conditioning module (3) is used to synchronously acquire the channel signals output by the sensor module; The processor module (4) is configured to receive the collected signals and execute the method steps S2) to S6) as described in any one of claims 1 to 7, and output the terrain category and motion status results.
9. A quadruped robot terrain and motion state recognition system without foot-end force sensors according to claim 8, characterized in that: The six-dimensional piezoelectric accelerometer (1) is installed at the center of the circular mounting area at the front end of the load structure on the back of the quadruped robot; the three-dimensional piezoelectric accelerometer (2) is installed at the reference point of the hip joint screw recess.
10. A quadruped robot, characterized in that: This includes a quadruped robot terrain and motion state recognition system without foot-end force sensors as described in claim 8 or 9.