A method, system, device and medium for perception coordination application of a quadruped robot

Abstract

The application provides a kind of four-legged robot perception coordination application method, system, equipment and medium, method includes: with clock synchronization signal as general reference to the clock alignment of multi-modal sensor is carried out, obtains multi-modal perception data;With the high-frequency pose data of extraction to the spatial displacement integral compensation of olfactory sensor is carried out, scalar chemical data is mapped into spatial concentration gradient vector;To the perception data containing the vector is carried out cross-modal feature fusion, generates the perception state vector containing gas parameter feature and foot end slip feature;Based on the vector to the robot is carried out motion planning and decision, according to concentration change trend adjusts the anti-skid control sensitivity for foot end slip feature;And based on the trend determines potential pollution area, according to the relative spatial distance of robot and the area generates gait adjustment instruction.The technical scheme provided by the application can solve the problem of data space-time dislocation, realize pre-anti-skid prediction, and improve the motion stability.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and in particular to a method, system, device and medium for the perception and coordination application of a quadruped robot. Background Technology

[0002] With the increasing application of quadruped robots in complex and unstructured scenarios such as industrial inspection and emergency rescue, the workload of quadruped robots in these scenarios continues to rise, and their application coverage is constantly expanding. How to collect, process, and analyze multimodal data such as the geometric space, physical contact, and environmental chemical media around the quadruped robot has become a core requirement for ensuring the stability of quadruped robot movement and improving its environmental adaptability.

[0003] In existing technologies, the environmental perception and decision-making control of quadruped robots are typically achieved through a multi-sensor parallel sampling system based on independent clock sources. For example, the vision system independently outputs the spatial obstacle distribution at a fixed frame rate, the foot force sensors output changes in contact force according to their own operating frequency, and chemical sensors such as olfaction sensors are usually only used as independent safety monitoring and alarm devices. The main control unit receives the above independent data streams respectively and, based on the preset single-modal judgment thresholds of each sensor, sequentially outputs the corresponding gait control commands (such as adjusting posture after force detection of slippage, or triggering an audible and visual alarm only after olfaction exceeds the threshold).

[0004] However, in practical applications, existing technologies based on independent clock sampling and single-modal threshold triggering suffer from problems such as spatiotemporal mapping misalignment of multimodal data and inconsistent environmental state representation when dealing with multimodal perception scenarios involving high-frequency dynamic interactions and low-frequency environmental media changes, and the fusion of multi-source heterogeneous data with asymmetrical sampling frequency distribution and comprehensive environmental judgment. At the same time, traditional single-modal force feedback control often has lag, usually triggering adjustments only after physical slippage occurs. Furthermore, existing multi-sensor systems lack a deep coupling mechanism for cross-modal spatiotemporal features, making it impossible to detect and intervene in potential dynamic risks in advance in complex and unknown environments (such as areas with hazardous media leaks), thus leading to potential risks of motion instability and decision failure for quadruped robots. Summary of the Invention

[0005] In view of this, this application provides a method, system, device and medium for the perception and coordination application of quadruped robots to solve the above problems.

[0006] Firstly, a perceptual coordination application method for quadruped robots is provided, applicable to quadruped robots equipped with edge computing units and multimodal sensors, including spatial vision sensors, inertial measurement units, foot force sensors, and olfactory sensors. The method includes:

[0007] The clock synchronization signal is acquired and used as a universal time reference to perform clock alignment on the multimodal sensor. Based on the aligned time reference, multimodal sensing data with the corresponding sampling frequency is acquired.

[0008] Using the trigger time of the clock synchronization signal as the reference time, high-frequency pose estimation data output by the inertial measurement unit and spatial vision sensor is extracted from the multimodal sensing data. The high-frequency pose estimation data is then used to perform integral compensation on the spatial displacement of the olfactory sensor within the low-frequency sampling period, and the scalar chemical data output by the olfactory sensor is mapped into a spatial concentration gradient vector.

[0009] Multimodal sensing data containing spatial concentration gradient vectors are input into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector. The sensing state vector contains gas parameter features generated by the transformation of spatial concentration gradient vectors and foot slip features extracted from sensing data based on foot force sensor.

[0010] Motion planning and decision-making for the quadruped robot are performed based on the perception state vector, and the anti-slip control sensitivity is adjusted according to the concentration change trend in the gas parameter characteristics for foot slippage characteristics.

[0011] Potentially polluted areas are identified based on concentration change trends, and support phase gait adjustment control commands are generated based on the relative spatial distance between the quadruped robot and the potential polluted areas.

[0012] The above technical solution aligns multimodal sensors by using a clock synchronization signal as a universal reference and integrates low-frequency olfactory data using high-frequency pose data to generate a spatial concentration gradient vector. Combined with the perception state vector generated by cross-modal feature fusion, the quadruped robot can adjust its anti-slip control sensitivity in advance based on the gas concentration change trend and adjust its gait according to the relative distance of potential pollution areas. This solves the problem of spatiotemporal misalignment of multi-source heterogeneous data and the risk of hysteresis and physical slippage that exists in traditional single-modal force-sensing anti-slip control.

[0013] Optionally, the spatial displacement of the olfactory sensor during the low-frequency sampling period is integrally compensated using high-frequency pose estimation data, mapping the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector, specifically including:

[0014] Extracting multidimensional pose coordinate sequences from high-frequency pose estimation data;

[0015] Determine the motion trajectory curve within the low-frequency sampling period based on the multi-dimensional pose coordinate sequence;

[0016] The spatial displacement increment matrix is ​​obtained by numerically calculating the motion trajectory curve using a preset numerical integration algorithm.

[0017] The numerical changes of scalar chemical data within a low-frequency sampling period are extracted, and the derivative relationship between the numerical changes and the spatial displacement increment matrix is ​​calculated to generate a spatial concentration gradient vector.

[0018] The above technical solution determines the motion trajectory curve by extracting high-frequency multidimensional pose coordinate sequences and uses a preset numerical integration algorithm to calculate the spatial displacement increment matrix. Then, it calculates the derivative relationship between the matrix and the numerical change of scalar chemical data, which accurately transforms the low-frequency scalar chemical signal that originally had no spatial direction attribute into a concentration gradient vector with physical spatial directionality. This provides a quantifiable geometric mapping basis for subsequent space early warning and navigation control.

[0019] Optionally, a multi-dimensional pose coordinate sequence can be extracted based on high-frequency pose estimation data, specifically including:

[0020] Extract the inertial coordinate sequence of the corresponding inertial measurement unit and the visual odometry coordinate sequence of the corresponding spatial vision sensor from the high-frequency pose estimation data.

[0021] The characteristic residual values ​​of the inertial extrapolation coordinate sequence and the visual odometry coordinate sequence are calculated by using a preset Kalman filter model.

[0022] Based on the characteristic residual values, pose drift correction is performed on the inertial estimation coordinate sequence to obtain a smooth pose coordinate sequence;

[0023] The smooth pose coordinate sequence is determined as a multidimensional pose coordinate sequence.

[0024] The above technical solution extracts the inertial estimation coordinate sequence and the visual odometry coordinate sequence, and uses the Kalman filter model to calculate the characteristic residual values ​​of the two to perform pose drift correction on the inertial estimation coordinate sequence. This effectively eliminates the cumulative pose error of the quadruped robot in the complex motion process, obtains a high-precision smooth pose coordinate sequence, and ensures the accuracy of low-frequency olfactory spatial displacement integral compensation calculation.

[0025] Optionally, multimodal sensing data containing spatial concentration gradient vectors is input into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector, specifically including:

[0026] Multimodal sensing data containing spatial concentration gradient vectors is input into a pre-defined long short-term memory network model to perform time series feature alignment and obtain an aligned sensing data matrix.

[0027] Extract the chemical feature submatrix of the corresponding olfactory sensor and the mechanical feature submatrix of the corresponding foot force sensor from the alignment perception data matrix;

[0028] The cross-correlation weights between the chemical feature submatrix and the mechanical feature submatrix are calculated using a pre-defined cross-modal attention mechanism.

[0029] Weighted feature extraction is performed on the aligned sensing data matrix based on cross-correlation weights to obtain cross-modal fusion features, and the cross-modal fusion features are determined as gas parameter features;

[0030] The gas parameter features and foot slip features are concatenated to generate a perception state vector.

[0031] The above technical solution performs time series feature alignment on multimodal sensing data through a long short-term memory network model, and uses a cross-modal attention mechanism to calculate the cross-correlation weights of the chemical feature submatrix and the mechanical feature submatrix. The extracted cross-modal fusion features are identified as gas parameter features and are then spliced ​​with the foot slip features calculated at the physical level. This achieves the decoupling and deep integration of black-box data features and white-box physical features, avoiding the representation confusion caused by the forced fusion of features from different modalities.

[0032] Optionally, the steps for extracting foot slippage features from the perceived state vector specifically include:

[0033] Extract multi-axis contact force data from the sensor data of the foot force sensor;

[0034] The multi-axis contact force data is denoised using a preset sliding window filtering algorithm to obtain smooth contact force data.

[0035] Extract the horizontal tangential force and vertical normal force from the smooth contact force data, and calculate the ratio of the horizontal tangential force to the vertical normal force to obtain the real-time friction coefficient estimate.

[0036] The difference between the preset critical slip threshold and the real-time friction coefficient estimate is calculated to obtain the foot slip characteristics.

[0037] The above technical solution extracts multi-axis contact force data from the foot force sensor and performs sliding window noise reduction processing. It calculates the real-time friction coefficient estimate by using the ratio of the horizontal tangential force component to the vertical normal force component, and then obtains the difference between the preset critical slip threshold and the estimate. This provides a quantitative assessment basis for the physical slip state of the foot with low computational complexity and in accordance with objective mechanical laws.

[0038] Optionally, the sensitivity of anti-slip control for foot slippage characteristics can be adjusted based on the concentration change trend in the gas parameter characteristics, specifically including:

[0039] Extract the magnitude of the spatial concentration gradient vector from the gas parameter characteristics, calculate the time derivative of the magnitude within a continuous preset time window, and determine the time derivative as the concentration change trend;

[0040] The determination result is obtained by comparing the concentration change trend with the preset concentration alarm threshold range, and the environmental pollution level is determined based on the determination result.

[0041] Extract the corresponding preset sensitivity compensation coefficient based on the environmental pollution level;

[0042] Add the preset sensitivity compensation coefficient to the preset anti-slip sensitivity base value to obtain the adjusted anti-slip control sensitivity.

[0043] The above technical solution extracts the magnitude of the spatial concentration gradient vector and calculates its time derivative within a continuous preset time window as the concentration change trend. This reduces the dimensionality of the multidimensional vector into a pure scalar value that is easy to compare. The value is then compared with a preset concentration alarm threshold range to classify the environmental pollution level. By superimposing the corresponding sensitivity compensation coefficient, the stepwise quantitative adjustment of the anti-slip control sensitivity is achieved.

[0044] Optionally, the concentration change trend is compared with the preset concentration alarm threshold range to obtain the judgment result, and the environmental pollution level is determined based on the judgment result, specifically including:

[0045] Extract the positive and negative sign characteristics and absolute values ​​corresponding to the concentration change trend;

[0046] When the positive and negative sign characteristics are positive and the absolute value is greater than or equal to the preset first derivative threshold, the environmental pollution level is mapped to a high-risk leakage pollution level.

[0047] When the positive and negative sign characteristics are positive and the absolute value is less than the first derivative threshold and greater than or equal to the preset second derivative threshold, the environmental pollution level is mapped to the potential diffusion pollution level.

[0048] When the positive or negative sign characteristic is negative or the absolute value is less than the second derivative threshold, the environmental pollution level is mapped to a safe attenuation pollution level. Here, the first derivative threshold is greater than the second derivative threshold, and the first derivative threshold and the second derivative threshold constitute the judgment boundary of the preset concentration alarm threshold range.

[0049] The above technical solution extracts the positive and negative sign characteristics and absolute values ​​of the concentration change trend, and compares them with the first and second derivative thresholds with clearly defined magnitudes. This constructs a mutually exclusive conditional judgment interval that covers three states: high-risk leakage pollution, potential diffusion pollution, and safe decay pollution. This provides a clear and comprehensive underlying code logic judgment boundary for the system to determine the environmental pollution level.

[0050] Secondly, a perception and coordination application system for a quadruped robot is provided, applicable to a quadruped robot equipped with an edge computing unit and multimodal sensors. The multimodal sensors include a spatial vision sensor, an inertial measurement unit, a foot force sensor, and an olfactory sensor. The system includes:

[0051] The timing synchronization module is used to acquire the clock synchronization signal and use the clock synchronization signal as a universal time reference to perform clock alignment on the multimodal sensor. Based on the aligned time reference, it acquires multimodal sensing data with the corresponding sampling frequency.

[0052] The spatial mapping module is used to extract high-frequency pose estimation data from the multimodal sensing data, based on the trigger time of the clock synchronization signal, and to perform integral compensation on the spatial displacement of the olfactory sensor within the low-frequency sampling period using the high-frequency pose estimation data, thereby mapping the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector.

[0053] The feature fusion module is used to input multimodal sensing data containing spatial concentration gradient vectors into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector. The sensing state vector contains gas parameter features generated by the transformation of spatial concentration gradient vectors and foot slip features extracted from sensing data based on foot force sensor.

[0054] The anti-slip adjustment module is used to perform motion planning and decision-making for the quadruped robot based on the perceived state vector, and to adjust the anti-slip control sensitivity for foot slip characteristics according to the concentration change trend in the gas parameter characteristics.

[0055] The gait control module is used to determine potential contamination areas based on concentration change trends and to generate support phase gait adjustment control commands based on the relative spatial distance between the quadruped robot and the potential contamination areas.

[0056] Thirdly, an electronic device is provided, including a processor, a memory, a user interface, and a network interface, wherein the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any of the above.

[0057] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing instructions that, when executed, perform the method as described in any of the preceding claims.

[0058] In summary, implementing one or more technical solutions provided in this application has at least the following technical effects or advantages:

[0059] By constructing a dual-dimensional decision-making architecture that integrates "chemical olfactory prior judgment" and "physical force-sensory contact feedback," this application overcomes the technical shortcomings of traditional quadruped robot systems, where chemical perception is only used for independent safety alarms and motion control relies on passive force-sensory compensation in isolation. This application transforms the macroscopic, low-frequency insight into the evolution of environmental chemical media into preventative intervention commands targeting microscopic, high-frequency dynamic slippage, shifting the robot's control logic from post-event feedback regulation to pre-event feedforward control. This cross-modal fusion scheme significantly reduces the probability of robot instability and falls in complex, unstructured scenarios filled with unknown lubricating or corrosive leaking liquids, such as chemical inspections and disaster relief, improving the operational continuity of quadruped robots under extreme conditions. Attached Figure Description

[0060] Figure 1 This is a flowchart illustrating a perceptual coordination application method for a quadruped robot disclosed in this application;

[0061] Figure 2 This is a schematic diagram of a perceptual coordination application system for a quadruped robot disclosed in this application;

[0062] Figure 3 This is a schematic diagram of the structure of an electronic device disclosed in this application.

[0063] Explanation of reference numerals in the attached figures: 201, Timing synchronization module; 202, Spatial mapping module; 203, Feature fusion module; 204, Anti-slip adjustment module; 205, Gait control module; 301, Processor; 302, Communication bus; 303, User interface; 304, Network interface; 305, Memory. Detailed Implementation

[0064] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0065] In the description of the embodiments of this application, the words "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design that is described as "for example" or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design options. Rather, the use of the words "for example" or "for instance" is intended to present the relevant concepts in a specific manner.

[0066] In the description of the embodiments of this application, the term "multiple" means two or more. For example, multiple systems means two or more systems, and multiple screen terminals means two or more screen terminals. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0067] Figure 1 This is a flowchart illustrating a perceptual coordination application method for a quadruped robot according to an embodiment of this application. This method can be implemented using a computer program, a microcontroller, or run on a perceptual coordination application system for a quadruped robot. The computer program can be integrated into the application or run as a standalone tool application. The specific steps of the perceptual coordination application method for a quadruped robot are described in detail below.

[0068] S101: Acquire a clock synchronization signal and use the clock synchronization signal as a universal time reference to perform clock alignment on the multimodal sensor. Based on the aligned time reference, acquire multimodal sensing data with the corresponding sampling frequency.

[0069] In the embodiments of this application, the clock synchronization signal refers to a global level pulse or network timestamp instruction used to coordinate the time cycle of each independently operating hardware module at the underlying level. For example, a hard trigger pulse signal that is sent cyclically at a fixed high frequency (such as 100Hz) on the underlying communication bus of the main control board. Each device immediately resets its internal counter or marks it to the same physical time the moment it receives the pulse.

[0070] Specifically, the system first acquires the globally distributed clock synchronization signal and uses this signal as the absolute time scale (i.e., a universal time reference) that is followed across hardware platforms. For multimodal sensors (a collection of hardware devices with different physical detection principles, including spatial vision, inertial, force, and chemical olfaction sensors), since each relies on an independent built-in clock source and has a different inherent operating rhythm, direct reading would cause timing mismatches in the data stream that exceed the control cycle tolerance. Therefore, it is necessary to use the aforementioned universal time reference to perform clock alignment on the multimodal sensors. This is achieved by forcibly eliminating the phase difference and time offset between the internal clocks of each heterogeneous sensor through hardware pulse hard triggering or software timestamp interpolation compensation. After completing the strict alignment of the global timeline, based on the aligned time reference, the system periodically extracts real-time measurement values ​​according to the inherently different sampling frequencies of each sensor, thereby acquiring multimodal sensing data with a unified and accurate timestamp label.

[0071] S102: Using the trigger time of the clock synchronization signal as the reference time, extract the high-frequency pose estimation data output by the inertial measurement unit and the spatial vision sensor from the multimodal sensing data, and use the high-frequency pose estimation data to perform integral compensation on the spatial displacement of the olfactory sensor within the low-frequency sampling period, and map the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector.

[0072] For example, because olfactory sensors require physical contact and reaction processes to perceive chemical molecules, their sampling frequency is typically much lower than the robot's motion control frequency. This results in the chemical concentration change measured within a single low-frequency sampling cycle when a quadruped robot is moving dynamically being distributed along a segment of the robot's trajectory. The core function of this step is to utilize the high-frequency refresh rate of the vision and inertial units (on the order of hundreds of hertz) to accurately reconstruct the direction and distance the robot has moved in three-dimensional space during this relatively long odor collection period through high-density data interpolation. By mathematically mapping this precise physical displacement to the pure numerical concentration increment, the originally directionless numerical value representing only a scalar increase or decrease in concentration is given a real physical spatial orientation, transforming it into a spatial concentration gradient vector that can directly point to the location of the pollution source.

[0073] In one possible implementation, high-frequency pose estimation data is used to perform integral compensation on the spatial displacement of the olfactory sensor during the low-frequency sampling period, mapping the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector. Specifically, this includes: extracting a multidimensional pose coordinate sequence based on the high-frequency pose estimation data; determining the motion trajectory curve during the low-frequency sampling period based on the multidimensional pose coordinate sequence; performing numerical calculations on the motion trajectory curve using a preset numerical integration algorithm to obtain a spatial displacement increment matrix; extracting the numerical change of the scalar chemical data during the low-frequency sampling period, and calculating the derivative relationship between the numerical change and the spatial displacement increment matrix to generate a spatial concentration gradient vector.

[0074] In the embodiments of this application, the spatial concentration gradient vector refers to a multidimensional physical quantity that simultaneously possesses spatial physical direction attributes and concentration change magnitude attributes. For example, it is a mathematical vector used to represent the specific geometric direction in which the concentration of a certain leaked chemical gas in the environment increases the fastest in a three-dimensional spatial coordinate system and the corresponding degree of drastic change.

[0075] Specifically, the system first extracts a multi-dimensional pose coordinate sequence based on high-frequency pose estimation data (referring to a set of discrete data points containing three-dimensional position coordinates and Euler angles recorded during continuous robot movement in physical space); based on the above multi-dimensional pose coordinate sequence, the system determines the motion trajectory curve of the olfactory sensor within a relatively long low-frequency sampling period (referring to the inherently long signal acquisition time span of chemical sensors due to the slow physical contact reaction of gas molecules); and then uses a preset numerical integration algorithm (referring to mathematical tools such as the trapezoidal rule or Runge-Kutta method used to calculate the cumulative quantity of a continuous function in space) to numerically calculate the motion trajectory curve. The calculation precisely reconstructs the cumulative distance of each spatial element within the low-frequency period, yielding a spatial displacement increment matrix (a mathematical matrix representing the sum of actual displacement vectors along each coordinate axis). Simultaneously, it extracts the numerical change of scalar chemical data (i.e., the pure numerical difference between the current and previous concentration values) within the aforementioned low-frequency sampling period, which lacks any directional indication. It then calculates the derivative relationship between this numerical change and the spatial displacement increment matrix (i.e., the partial derivative of the concentration scalar with respect to the three-dimensional spatial displacement matrix). By establishing a partial differential mapping between the rate of change of chemical concentration and the cumulative displacement in physical space, a spatial concentration gradient vector is generated. Specifically, the mathematical representation involves: extracting the spatial displacement coordinate sequence and the corresponding scalar chemical concentration set within the current moment and multiple consecutive historical low-frequency sampling periods; constructing an overdetermined linear equation system with spatial displacement components of each dimension as independent variables and the corresponding chemical concentration values ​​as dependent variables; and then using the least squares method to optimally solve the overdetermined linear equation system to obtain the three-dimensional spatial concentration gradient vector that minimizes the overall squared error. The direction of the resultant vector points to the spatial physical orientation where the concentration increase is steepest. , and These represent the spatial concentration partial derivative components of the three-dimensional spatial concentration gradient vector along the three orthogonal directions of the X-axis, Y-axis, and Z-axis in the local Cartesian coordinate system.

[0076] In one possible implementation, a multidimensional pose coordinate sequence is extracted based on high-frequency pose estimation data. Specifically, this includes: extracting the inertial estimation coordinate sequence corresponding to the inertial measurement unit and the visual odometry coordinate sequence corresponding to the spatial vision sensor from the high-frequency pose estimation data; calculating the feature residual values ​​of the inertial estimation coordinate sequence and the visual odometry coordinate sequence using a preset Kalman filter model; performing pose drift correction on the inertial estimation coordinate sequence based on the feature residual values ​​to obtain a smooth pose coordinate sequence; and determining the smooth pose coordinate sequence as a multidimensional pose coordinate sequence.

[0077] In the embodiments of this application, the preset Kalman filter model refers to a statistical algorithm framework that uses the linear system state equation, combined with system input and multiple observation data, to make the optimal recursive estimation of the dynamic state of the system. For example, it is a mathematical filtering model that is pre-written in the robot navigation code and is specifically used to fuse inertial data and visual data to filter environmental noise and predict the state closest to the real physical trajectory.

[0078] Specifically, the system first decomposes the previously acquired high-frequency pose estimation data, extracting the inertial estimation coordinate sequence corresponding to the inertial measurement unit (the set of coordinate trajectories independently calculated by internal gyroscopes and accelerometers through high-frequency numerical integration) and the visual odometry coordinate sequence corresponding to the spatial vision sensor (the set of coordinate trajectories calculated by extracting static feature points in continuous environmental image frames and matching relative motion). These two data sources, with consistent spatiotemporal references but complementary error characteristics, are input. Using the state observation equation within a preset Kalman filter model, the low-frequency, high-precision characteristics of the visual data are used to constrain the high-frequency divergence characteristics of the inertial data, thus calculating the inertial estimation coordinate sequence and the visual odometry coordinate sequence. The characteristic residual values ​​of the two sequences at the same time in the spatial geometric estimate (used to represent the difference data of inherent bias and cumulative error between observations from different sensors); based on the calculated characteristic residual values ​​and the corresponding Kalman optimal gain coefficient, pose drift correction is performed on the inertial extrapolation coordinate sequence which has a tendency to diverge over time, eliminating the trajectory deviation caused by integral accumulation, and thus obtaining a smooth pose coordinate sequence with high frequency continuity and minimal absolute error (used to represent the high-order smooth three-dimensional coordinate set after removing random noise and long-term drift distortion); the smooth pose coordinate sequence output by the filter correction is directly determined as the required multi-dimensional pose coordinate sequence, thereby completing the high-fidelity extraction of heterogeneous sensor pose data.

[0079] S103: Input multimodal sensing data containing spatial concentration gradient vector into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector. The sensing state vector contains gas parameter features generated by the transformation of spatial concentration gradient vector and foot slip features extracted from sensing data based on foot force sensor.

[0080] For example, when a quadruped robot performs inspection tasks in a complex chemical plant area, the information received by its edge computing unit is completely decoupled at the underlying physical properties: one end is the spatial distribution pattern of the chemical medium captured by the olfactory sensor, and the other end is the physical contact force parameters fed back in real time by the foot force sensor. The process of cross-modal feature fusion performed by the edge computing unit is essentially to forcibly align and encapsulate these two heterogeneous physical quantities that span chemical and mechanical dimensions into the same mathematical feature space. The resulting perception state vector is equivalent to providing the robot's motion decision center with a set of comprehensive state observation matrices from a global perspective: this matrix includes both the prior environmental prediction weights for the unknown chemical pollution status of the road surface ahead and the ontological physical observation values ​​of the robot's current actual foot friction state, thus laying a unified data foundation for the system to subsequently formulate control strategies that take into account both safety avoidance and stable walking.

[0081] In one possible implementation, multimodal sensing data containing spatial concentration gradient vectors is input to an edge computing unit to perform cross-modal feature fusion and generate a sensing state vector. Specifically, this includes: inputting multimodal sensing data containing spatial concentration gradient vectors into a preset long short-term memory network model to perform time-series feature alignment to obtain an aligned sensing data matrix; extracting chemical feature sub-matrices corresponding to olfactory sensors and mechanical feature sub-matrices corresponding to foot force sensors from the aligned sensing data matrix; calculating the cross-correlation weights of the chemical feature sub-matrices and mechanical feature sub-matrices through a preset cross-modal attention mechanism; performing weighted feature extraction on the aligned sensing data matrix based on the cross-correlation weights to obtain cross-modal fusion features, and determining the cross-modal fusion features as gas parameter features; and concatenating the gas parameter features with foot slip features to generate a sensing state vector.

[0082] In this application embodiment, the preset long short-term memory network model refers to an artificial neural network architecture that is good at processing time series data and can effectively solve the problem of long sequence time-series dependencies. For example, it is a deep learning computing model that is pre-deployed in the edge device of a quadruped robot and is specifically used to extract the implicit time evolution rules and spatiotemporal mapping relationships in the historical data of continuously acquired multimodal sensors.

[0083] Specifically, the system first inputs the pre-processed multimodal sensing data, including spatial concentration gradient vectors, into a pre-defined long short-term memory network model. It then utilizes mechanisms such as forget gates and input gates within this model to extract the implicit dynamic changes in the time dimension of the data streams from various heterogeneous sensors. This is to perform time series feature alignment (eliminating data lag and misalignment caused by inconsistent sampling frequencies of different sensors, achieving absolute synchronization in the time step dimension within the feature space of the neural network). The system then outputs a standardized and reconstructed aligned sensing data matrix (representing a two-dimensional numerical array of tightly aligned and dimensionally uniform modal features along the time axis). Following a pre-defined network channel index, the system retrieves data from this matrix... The chemical feature sub-matrix (a local data block representing the spatial distribution and evolution of gas concentration) of the olfactory sensor and the mechanical feature sub-matrix (a local data block representing the evolution of plantar contact reaction force and friction) of the foot force sensor are extracted from the aligned sensing data matrix. These two heterogeneous sub-matrices are then fed into a pre-defined cross-modal attention mechanism (a computational operator model that dynamically evaluates the importance of mutual influence between different modal data based on matrix inner product and normalization operations) to calculate the potential prior mapping relationship between low-frequency chemical medium changes and high-frequency physical slippage, thereby accurately calculating the cross-correlation weights of the chemical and mechanical feature sub-matrices. In the underlying computational logic of the pre-defined cross-modal attention mechanism, to achieve chemical-mechanical coupling, the mechanical feature sub-matrix is ​​linearly transformed and mapped to a query vector matrix Q, and the chemical feature sub-matrix containing prior environmental prediction information is mapped to a bond vector matrix K and a value vector matrix V through different weight matrices. The weighted feature extraction process for the cross-correlation weights follows the scaling dot product attention formula:

[0084]

[0085] in, The feature dimension of the key vector; Key vector matrix The transpose of the matrix; This is the normalization exponential function, used to transform the scaled matrix inner product result into a probability weight distribution that sums to 1; Let V be the feature output function of the attention mechanism, representing the cross-modal fusion feature matrix obtained by performing a weighted summation on the value vector matrix V based on the probability weight distribution.

[0086] Through this operator, the system can adaptively increase the weight of chemical prior feature channels with drastic changing trends in the mechanics-physical feature fusion calculation, thereby achieving deep mathematical decoupling and fusion of underlying heterogeneous features.

[0087] Furthermore, based on the calculated cross-correlation weights, the original aligned sensing data matrix is ​​re-enhanced with weighted feature extraction, focusing on amplifying the activation values ​​of chemical prior data channels that are strongly correlated with physical slip, thereby obtaining highly condensed cross-modal fusion features. These attention-weighted cross-modal fusion features are then directly identified as gas parameter features. The gas parameter features extracted and weighted by the deep network are combined with the foot slip features (used to represent the numerical representation of the current transient physical friction margin of the foot) independently calculated by the underlying physical model, according to a preset vector dimension order, and then subjected to dimensionality reduction and feature concatenation operations. This outputs a sensing state vector containing dual coupling information of high-frequency physics and low-frequency chemistry, providing a unified data base for subsequent motion decisions.

[0088] In one possible implementation, the step of extracting foot slippage features from the perceived state vector specifically includes: extracting multi-axis contact force data from the perceived data of the foot force sensor; denoising the multi-axis contact force data using a preset sliding window filtering algorithm to obtain smooth contact force data; extracting the horizontal tangential force component and the vertical normal force component from the smooth contact force data, and calculating the ratio of the horizontal tangential force component to the vertical normal force component to obtain a real-time friction coefficient estimate; and calculating the difference between a preset critical slippage threshold and the real-time friction coefficient estimate to obtain the foot slippage features.

[0089] In this embodiment, the preset sliding window filtering algorithm refers to a local smoothing processing mechanism based on time series. It is a numerical processing model that calculates the average or weighted value of the data within the window to filter out high-frequency burst noise signals by setting a data window of fixed length and continuously sliding and updating it on the time axis as new sampled data arrives. For example, the mean filtering code with a length of 5 consecutive sampling periods is preset in the underlying motion controller to eliminate instantaneous force peak interference caused by robot mechanical vibration or small gravel on the ground.

[0090] Specifically, the system first parses the data stream at the hardware level, extracting multi-axis contact force data from the foot force sensor (referring to the original discrete numerical set of physical force feedback in each direction in three-dimensional space obtained in real time by the three-axis or six-axis torque sensors installed on the foot end of the quadruped robot); inputting the above-mentioned original force sequence containing sudden environmental noise interference, and performing noise reduction processing on the continuously input multi-axis contact force data through a preset sliding window filtering algorithm, that is, using a sliding data window to remove abnormal force abrupt isolation points and performing mean or median smoothing calculation, thereby obtaining smooth contact force data that eliminates high-frequency mechanical jitter interference (used to represent a continuous and reliable force numerical sequence that can stably reflect the real physical contact state of the foot); based on the encoder angle data in the motors of each leg joint, The transformation matrix of the foot kinematic coordinate system obtained by solving the spatial forward kinematics is used to decouple and extract the horizontal tangential component (the magnitude of the lateral force parallel to the ground contact surface that propels the robot forward or causes slippage) and the vertical normal component (the magnitude of the positive force perpendicular to the contact surface that reflects the robot's supporting pressure) from the smooth contact force data. Based on Coulomb's fundamental law of friction, the ratio of the horizontal tangential component to the vertical normal component is calculated, thus obtaining a real-time friction coefficient estimate (a dynamic quantitative evaluation index characterizing the actual physical anti-slip capability between the foot and the ground contact surface). A preset critical slip threshold (the maximum safe friction coefficient reference limit that can be tolerated before physical slippage occurs on the corresponding medium surface) is extracted based on calibration experiments or material properties. Specifically, it is assumed that the smooth contact force data is decomposed into two mutually orthogonal tangential components along the horizontal plane in the robot's foot coordinate system. and and the normal component perpendicular to the ground. Real-time friction coefficient estimate The specific calculation model is as follows:

[0091]

[0092] Furthermore, let the pre-calibrated critical slip threshold of the corresponding safe medium surface be... Then the characteristics of foot slippage The quantitative numerical expression is as follows .when When the value approaches zero or is negative, it indicates an extremely high risk of physical slippage. The difference between the preset critical slippage threshold and the currently calculated real-time friction coefficient estimate is used to accurately assess the critical safety margin between the current foot force state and the actual physical slippage, and outputs the foot slippage characteristics that quantitatively characterize the degree of slippage risk in the underlying physical mechanics.

[0093] S104: Based on the perception state vector, the quadruped robot performs motion planning and decision-making, and adjusts the anti-slip control sensitivity for foot slip characteristics according to the concentration change trend in the gas parameter characteristics.

[0094] For example, in the actual inspection operation of a quadruped robot, if the integrated perception state vector indicates that the current physical surface is flat, but the olfactory channel detects a sharp positive increase in the derivative of the odor concentration of lubricating fluid or chemical solvent ahead, the system determines that the robot is approaching a potentially low-friction hazard area. At this time, the edge computing unit not only performs deceleration or replanning motion decisions on the macroscopic path, but also lowers the friction difference threshold for foot anti-slip determination in the underlying motor control. This means that the gain coefficient of the underlying impedance controller is pre-configured to a high-sensitivity response range. When the foot subsequently experiences millimeter-level initial physical slippage, the controller responds and outputs joint torque compensation and posture correction actions that reach the preset peak torque limit, thereby transforming the traditional feedback compensation control into a feedforward impedance adjustment strategy based on prior prediction of chemical odor.

[0095] In one possible implementation, the anti-slip control sensitivity targeting foot slippage characteristics is adjusted based on the concentration change trend in the gas parameter characteristics. Specifically, this includes: extracting the magnitude of the spatial concentration gradient vector in the gas parameter characteristics; calculating the time derivative of the magnitude within a continuous preset time window; and determining the time derivative as the concentration change trend; comparing the concentration change trend with a preset concentration alarm threshold range to obtain a judgment result; determining the environmental pollution level based on the judgment result; extracting a corresponding preset sensitivity compensation coefficient based on the environmental pollution level; and adding the preset sensitivity compensation coefficient to a preset basic value for anti-slip sensitivity to obtain the adjusted anti-slip control sensitivity.

[0096] In this embodiment, the anti-slip control sensitivity refers to the comprehensive quantitative parameter of the threshold threshold and torque compensation gain of the underlying motion controller in response to unexpected physical slippage of the foot. For example, when the robot walks on a flat and dry road, the sensitivity is kept at a low default value. However, when the robot predicts the presence of dangerous media such as lubricating fluid ahead by smell, the sensitivity is increased in advance, so that the controller can immediately trigger a powerful anti-slip posture recovery action with a larger joint torque at the moment when a tiny slip occurs on the sole of the foot.

[0097] Specifically, the system first separates the spatial concentration gradient vector with three-dimensional spatial orientation from the gas parameter features output through cross-modal fusion. Then, it eliminates the directional attribute through mathematical modulo operations, extracting a pure scalar value, i.e., the modulus, that only represents the absolute drasticness of concentration changes. To accurately capture the dynamic characteristics of this scalar's evolution over time, a continuous preset time window is selected (e.g., a fixed-length time period counting backwards from the current moment). A numerical differentiation algorithm is used to calculate the time derivative of the modulus within this continuous preset time window. This time derivative, representing the rate of increase or decrease in gas concentration over time, is directly determined as the concentration change trend for feedforward intervention. A pre-defined preset concentration alarm threshold interval (usually composed of multiple numerical step boundaries) is retrieved from internal memory, and the real-time calculated concentration change trend is rigorously compared with the value of this interval. The system determines which sub-interval a robot falls into and directly maps this result to determine the environmental pollution level (e.g., high-risk, potential, or safe). Based on this determined environmental pollution level, it queries and extracts the strictly corresponding preset sensitivity compensation coefficient (i.e., the additional control gain value required to cope with this specific hazard level) from the pre-configured control mapping table. It then obtains the preset anti-slip sensitivity baseline value for the robot's default operation in a normal safe environment and linearly adds the extracted preset sensitivity compensation coefficient to this baseline value. Through this feedforward numerical compensation method, it obtains the adjusted anti-slip control sensitivity, which serves as the latest input constraint for the underlying impedance controller. This allows for the dynamic tightening of the physical anti-slip warning threshold for high-frequency foot slippage characteristics using the prior evolution trend of chemical concentration. Specifically, the adjusted anti-slip control sensitivity is inversely mapped to foot slippage characteristics in the underlying control framework. The dynamic trigger threshold. When the system determines the dynamically calculated threshold... When the current foot contact state falls below the dynamic trigger threshold, it indicates that the current foot contact state has breached the safety redundancy boundary set by the chemical prior information, and the system immediately triggers the underlying anti-slip torque compensation and attitude correction actions.

[0098] In one possible implementation, a judgment result is obtained by comparing the concentration change trend with a preset concentration alarm threshold range. The environmental pollution level is determined based on the judgment result. Specifically, this includes: extracting the positive and negative sign features and absolute values ​​corresponding to the concentration change trend; when the positive and negative sign features are positive and the absolute value is greater than or equal to a preset first derivative threshold, the environmental pollution level is mapped to a high-risk leakage pollution level; when the positive and negative sign features are positive and the absolute value is less than the first derivative threshold but greater than or equal to a preset second derivative threshold, the environmental pollution level is mapped to a potential diffusion pollution level; when the positive and negative sign features are negative or the absolute value is less than the second derivative threshold, the environmental pollution level is mapped to a safe decay pollution level. The first derivative threshold is greater than the second derivative threshold, and the first derivative threshold and the second derivative threshold constitute the judgment boundary of the preset concentration alarm threshold range.

[0099] In this embodiment of the application, the preset concentration alarm threshold range refers to a numerical judgment range composed of multiple concentration change rate limits, used to measure the drastic change in the concentration of environmental chemical media and trigger the corresponding level of anti-slip response. For example, in the underlying navigation algorithm code, a set of logical judgment boundaries is pre-set by two key rate nodes of 50ppm / s and 20ppm / s, used to strictly define whether the gas is currently in a state of rapid leakage, slow diffusion or normal dissipation.

[0100] Specifically, the system first performs mathematical attribute decomposition on the input concentration change trend, extracting the positive and negative sign features (directional indicators representing whether the gas concentration is increasing or decreasing over time) and the absolute value (a non-negative scalar parameter reflecting the rate of concentration change after stripping away directional attributes). The extracted features are then placed into multi-level conditional branches for rigorous logical comparison. When the positive and negative sign features are positive (representing a continuous increase in the concentration of harmful or lubricating gases in the surrounding area) and the absolute value is greater than or equal to a preset first derivative threshold (a pre-defined highest-level rate threshold representing an extremely dangerous rapid concentration spike), the highest-level warning logic is triggered, mapping the environmental pollution level (a qualitative classification label representing the degree of danger of chemical media in the current physical space and the corresponding urgency of anti-slip measures) to a high-risk leakage pollution level. Conversely, when the positive and negative sign features are also positive, but the absolute value fails to reach the highest danger level (i.e., is less than the first derivative threshold) and is simultaneously greater than or equal to the preset first derivative threshold, the system performs a higher-level warning. When the second derivative threshold (a pre-calibrated threshold representing a relatively slow rate of gas spread) is reached, it indicates the presence of a hazardous medium ahead, but the concentration peak has not yet been reached. In this case, the pollution level is downgraded, mapping the environmental pollution level to the potential diffusion pollution level. When the positive or negative sign is negative (representing a decrease in gas concentration and the hazard source moving away) or, regardless of the sign direction, the absolute value is strictly less than the second derivative threshold (meaning the concentration change is extremely weak, falling within the sensor's normal noise floor or safe environmental fluctuation range), the motion warning restriction is lifted, mapping the environmental pollution level to the safe attenuation pollution level. In the entire judgment mechanism, the first derivative threshold is mathematically set to be strictly greater than the second derivative threshold. It is precisely by using the first and second derivative thresholds that a judgment boundary is constructed that is mutually exclusive and covers all possible preset concentration alarm threshold ranges. This allows for a complete comparison of the concentration change trend with the preset concentration alarm threshold range to obtain the judgment result, and accurately determines the environmental pollution level based on the judgment result.

[0101] S105: Determine potential pollution areas based on concentration change trends, and generate support phase gait adjustment control commands based on the relative spatial distance between the quadruped robot and the potential pollution areas.

[0102] In this embodiment, the support phase gait adjustment control command refers to the underlying control signal issued to the legs (i.e., the support phase) of the quadruped robot that maintain contact with the ground during the walking cycle, bear the weight of the robot, and provide physical driving force. This signal is used to change the landing trajectory, joint impedance, or stepping frequency in advance. For example, when it is predicted that there is a large area of ​​lubricating oil leakage on the ground ahead, the specific execution code sent to the joint motors of each leg currently stepping on the ground is to "shorten the stride length, increase the control gain, and switch to high-frequency small step mode".

[0103] Specifically, the system first uses the current pose coordinates determined by high-frequency pose estimation data as the reference position. Along the physical direction of the spatial concentration gradient vector, it uses preset physical space mapping coefficients or nonlinear transformation functions to convert the previously calculated concentration change trend (used to represent the derivative or intensity of the chemical gas concentration scalar's evolution over time within a preset time window) into a warning length in physical space. It then constructs a geometric distribution surface with this warning length as the boundary in the local navigation cost map and identifies this virtual geometric surface as a potential pollution area (referring to a specific plot in the physical environment that, although not actually stepped on or touched by feet, is highly likely to contain hazardous media or pose a risk of physical slippage based on prior chemical characteristics).

[0104] The underlying execution logic for constructing the geometric distribution surface in the local navigation cost map specifically includes: converting the direction of the spatial concentration gradient vector and the warning length from a polar coordinate system with the robot body as the origin. Transform the projection onto the Cartesian coordinate system of the local 2D raster map. In the middle, with the projection endpoint as the geometric center, a preset two-dimensional Gaussian decay function or morphological dilation operator is used to assign gradient increment values ​​to the cost of each discrete grid cell covering the geometric surface. The grid region with the higher the cost value, the greater the penalty weight it is selected when the underlying local path planner (such as the dynamic window method) is executed, thereby forcing the robot to plan a smooth bypass trajectory to avoid the contaminated area in advance.

[0105] Furthermore, the center point position of the quadruped robot is tracked in real time in the global coordinate system, and the shortest straight line length between this position and the mapped boundary of the potential contamination area is calculated using a spatial geometry algorithm to obtain the relative spatial distance (used to represent the physical buffer margin between the robot's current safe position and the edge of the hazard source in front). This relative spatial distance is used as the trigger condition for feedforward control. When this distance continuously shortens and becomes less than the preset safety braking threshold, the regular periodic walking plan is interrupted, and the foot landing point and force parameters are replanned according to the severity of the contamination. The leg movements that are in the ground contact and force exertion phase are reconstructed to generate support phase gait adjustment control commands, thereby ensuring that the robot outputs the target anti-slip posture control parameters before contacting the hazardous medium.

[0106] At the physical drive execution level, the support phase gait adjustment control command specifically acts on the underlying compliant impedance control model of the leg joint motor. The force equation of this model is expressed as:

[0107]

[0108] in, This represents the final control force (or torque) output to the leg joint. This represents the system's preset virtual mass (or inertia) coefficient matrix; and These represent the desired velocity and the actual velocity at the foot (i.e., the first time derivative of the displacement), respectively. and These represent the desired acceleration and the actual acceleration (i.e., the second time derivative of the displacement) at the foot, respectively. and These represent the desired and actual displacements of the foot, respectively. When a risk of slippage is confirmed ahead, the control command not only changes the macroscopic landing point but also dynamically adjusts the impedance parameters at the microscopic level: significantly increasing the virtual damping coefficient. To expand the power limit of the instantaneous sudden change in kinetic energy dissipated by the system during slippage; at the same time, to appropriately reduce the virtual stiffness coefficient. This allows the foot to exhibit compliance with preset contact force constraints when contacting an unknown contaminated liquid surface with low friction characteristics, thus avoiding instantaneous sideslip failure caused by rigid hard landing impact.

[0109] Figure 2 This is a schematic diagram of a perceptual coordination application system for a quadruped robot according to an embodiment of this application. This system can be implemented through software, hardware, or a combination of both, becoming all or part of the overall system. For example... Figure 2 As shown, the system includes:

[0110] The timing synchronization module 201 is used to acquire a clock synchronization signal and use the clock synchronization signal as a universal time reference to perform clock alignment on the multimodal sensor, and acquire multimodal sensing data with the corresponding sampling frequency based on the aligned time reference;

[0111] The spatial mapping module 202 is used to extract high-frequency pose estimation data output by the inertial measurement unit and spatial vision sensor from the multimodal sensing data, with the trigger time of the clock synchronization signal as the reference time, and to perform integral compensation on the spatial displacement of the olfactory sensor in the low-frequency sampling period using the high-frequency pose estimation data, and to map the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector.

[0112] The feature fusion module 203 is used to input multimodal sensing data containing spatial concentration gradient vectors into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector. The sensing state vector contains gas parameter features generated by the transformation of spatial concentration gradient vectors and foot slip features extracted from the sensing data based on the foot force sensor.

[0113] The anti-slip adjustment module 204 is used to perform motion planning and decision-making for the quadruped robot based on the perceived state vector, and to adjust the anti-slip control sensitivity for foot slip characteristics according to the concentration change trend in the gas parameter characteristics.

[0114] The gait control module 205 is used to determine the potential pollution area based on the concentration change trend, and generate support phase gait adjustment control commands according to the relative spatial distance between the quadruped robot and the potential pollution area.

[0115] Based on the above embodiments, as an optional embodiment, the spatial mapping module 202 is specifically used for: extracting a multi-dimensional pose coordinate sequence based on high-frequency pose estimation data; determining the motion trajectory curve within a low-frequency sampling period based on the multi-dimensional pose coordinate sequence; performing numerical calculations on the motion trajectory curve using a preset numerical integration algorithm to obtain a spatial displacement increment matrix; extracting the numerical change of scalar chemical data within a low-frequency sampling period, and calculating the derivative relationship between the numerical change and the spatial displacement increment matrix to generate a spatial concentration gradient vector.

[0116] Based on the above embodiments, as an optional embodiment, the spatial mapping module 202 is specifically used to: extract the inertial estimation coordinate sequence of the corresponding inertial measurement unit and the visual odometry coordinate sequence of the corresponding spatial vision sensor from the high-frequency pose estimation data; calculate the feature residual values ​​of the inertial estimation coordinate sequence and the visual odometry coordinate sequence through a preset Kalman filter model; perform pose drift correction on the inertial estimation coordinate sequence based on the feature residual values ​​to obtain a smooth pose coordinate sequence; and determine the smooth pose coordinate sequence as a multidimensional pose coordinate sequence.

[0117] Based on the above embodiments, as an optional embodiment, the feature fusion module 203 is specifically used for: inputting multimodal sensing data containing spatial concentration gradient vectors into a preset long short-term memory network model to perform time series feature alignment to obtain an aligned sensing data matrix; extracting the chemical feature sub-matrix corresponding to the olfactory sensor and the mechanical feature sub-matrix corresponding to the foot force sensor from the aligned sensing data matrix; calculating the cross-correlation weights of the chemical feature sub-matrix and the mechanical feature sub-matrix through a preset cross-modal attention mechanism; performing weighted feature extraction on the aligned sensing data matrix based on the cross-correlation weights to obtain cross-modal fusion features, and determining the cross-modal fusion features as gas parameter features; and concatenating the gas parameter features with the foot slip features to generate a sensing state vector.

[0118] Based on the above embodiments, as an optional embodiment, the feature fusion module 203 is specifically used to: extract multi-axis contact force data from the sensing data of the foot force sensor; perform noise reduction processing on the multi-axis contact force data through a preset sliding window filtering algorithm to obtain smooth contact force data; extract the horizontal tangential force component and the vertical normal force component from the smooth contact force data, and calculate the ratio of the horizontal tangential force component to the vertical normal force component to obtain the real-time friction coefficient estimate; calculate the difference between the preset critical slip threshold and the real-time friction coefficient estimate to obtain the foot slip characteristics.

[0119] Based on the above embodiments, as an optional embodiment, the anti-slip adjustment module 204 is specifically used for: extracting the magnitude of the spatial concentration gradient vector in the gas parameter characteristics, calculating the time derivative of the magnitude within a continuous preset time window, and determining the time derivative as the concentration change trend; comparing the concentration change trend with a preset concentration alarm threshold range to obtain a judgment result, and determining the environmental pollution level based on the judgment result; extracting the corresponding preset sensitivity compensation coefficient based on the environmental pollution level; and adding the preset sensitivity compensation coefficient to the preset anti-slip sensitivity base value to obtain the adjusted anti-slip control sensitivity.

[0120] Based on the above embodiments, as an optional embodiment, the anti-slip adjustment module 204 is specifically used to: extract the positive and negative sign features and absolute values ​​corresponding to the concentration change trend; when the positive and negative sign features are positive and the absolute value is greater than or equal to a preset first derivative threshold, the environmental pollution level is mapped to a high-risk leakage pollution level; when the positive and negative sign features are positive and the absolute value is less than the first derivative threshold and greater than or equal to a preset second derivative threshold, the environmental pollution level is mapped to a potential diffusion pollution level; when the positive and negative sign features are negative or the absolute value is less than the second derivative threshold, the environmental pollution level is mapped to a safe decay pollution level, wherein the first derivative threshold is greater than the second derivative threshold, and the first derivative threshold and the second derivative threshold constitute the determination boundary of the preset concentration alarm threshold range.

[0121] It should be noted that the system provided in the above embodiments is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the system and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0122] This embodiment also discloses an electronic device, referring to... Figure 3 The electronic device may include: at least one processor 301, at least one communication bus 302, user interface 303, network interface 304, and at least one memory 305.

[0123] The communication bus 302 is used to enable communication between these components.

[0124] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0125] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0126] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the electronic device (or edge computing unit) using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 305, and by calling data stored in the memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip without being integrated into the processor 301.

[0127] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. Figure 3 As shown, the memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and an application program for a perceptual coordination method for a quadruped robot.

[0128] exist Figure 3 In the electronic device shown, the user interface 303 is mainly used to provide an input interface for the user and to obtain the user input data; while the processor 301 can be used to call the application program stored in the memory 305 for a quadruped robot perception coordination application method. When executed by one or more processors 301, the electronic device executes one or more methods as described in the above embodiments.

[0129] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0130] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0131] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings or direct couplings or communication connections may be through some service interfaces; indirect couplings or communication connections between apparatuses or units may be electrical or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory 305 includes various media capable of storing program code, such as a USB flash drive, external hard drive, magnetic disk, or optical disk.

[0135] The above description is merely an exemplary embodiment disclosed in this application and should not be construed as limiting the scope of this application. Any equivalent changes and modifications made in accordance with the teachings of this application shall still fall within the scope of this application. Those skilled in the art will readily conceive of other embodiments disclosed in this application upon considering the disclosure of this specification. This application is intended to cover any variations, uses, or adaptations disclosed in this application that follow the general principles disclosed in this application and include common knowledge or customary technical means in the art not described in this application. The specification and embodiments are considered exemplary only, and the scope of this application is defined by the claims.

Claims

1. A perception and coordination application method for a quadruped robot, applied to a quadruped robot equipped with an edge computing unit and multimodal sensors, wherein the multimodal sensors include a spatial vision sensor, an inertial measurement unit, a foot force sensor, and an olfactory sensor, characterized in that, The method includes: A clock synchronization signal is acquired, and the clock synchronization signal is used as a universal time reference to perform clock alignment on the multimodal sensor. Multimodal sensing data with the corresponding sampling frequency is acquired based on the aligned time reference. Using the trigger time of the clock synchronization signal as the reference time, high-frequency pose estimation data output by the inertial measurement unit and the spatial vision sensor are extracted from the multimodal sensing data. The high-frequency pose estimation data is then used to perform integral compensation on the spatial displacement of the olfactory sensor within the low-frequency sampling period, and the scalar chemical data output by the olfactory sensor is mapped into a spatial concentration gradient vector. The multimodal sensing data containing the spatial concentration gradient vector is input to the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector, wherein the sensing state vector includes gas parameter features generated by the transformation of the spatial concentration gradient vector and foot slip features extracted based on the sensing data of the foot force sensor. Based on the perceived state vector, the quadruped robot performs motion planning and decision-making, and adjusts the anti-slip control sensitivity for the foot slip characteristics according to the concentration change trend in the gas parameter characteristics. Based on the concentration change trend, potential pollution areas are determined, and support phase gait adjustment control commands are generated according to the relative spatial distance between the quadruped robot and the potential pollution areas.

2. The method according to claim 1, characterized in that, The step of performing integral compensation on the spatial displacement of the olfactory sensor during the low-frequency sampling period using the high-frequency pose estimation data, and mapping the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector, specifically includes: Extract a multidimensional pose coordinate sequence based on the high-frequency pose estimation data; The motion trajectory curve within the low-frequency sampling period is determined based on the multidimensional pose coordinate sequence. The spatial displacement increment matrix is ​​obtained by numerically calculating the motion trajectory curve using a preset numerical integration algorithm. Extract the numerical change of the scalar chemical data within the low-frequency sampling period, and calculate the derivative relationship between the numerical change and the spatial displacement increment matrix to generate the spatial concentration gradient vector.

3. The method according to claim 2, characterized in that, The extraction of multidimensional pose coordinate sequences based on the high-frequency pose estimation data specifically includes: Extract the inertial coordinate sequence corresponding to the inertial measurement unit and the visual odometry coordinate sequence corresponding to the spatial vision sensor from the high-frequency pose estimation data; The characteristic residual values ​​of the inertial extrapolation coordinate sequence and the visual odometry coordinate sequence are calculated using a preset Kalman filter model. Based on the characteristic residual values, pose drift correction is performed on the inertial extrapolation coordinate sequence to obtain a smooth pose coordinate sequence; The smooth pose coordinate sequence is determined as the multidimensional pose coordinate sequence.

4. The method according to claim 1, characterized in that, The step of inputting the multimodal sensing data containing the spatial concentration gradient vector into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector specifically includes: The multimodal sensing data containing the spatial concentration gradient vector is input into a preset long short-term memory network model to perform time series feature alignment and obtain an aligned sensing data matrix. Extract the chemical feature sub-matrix corresponding to the olfactory sensor and the mechanical feature sub-matrix corresponding to the foot force sensor from the alignment perception data matrix; The cross-correlation weights between the chemical feature submatrix and the mechanical feature submatrix are calculated using a pre-defined cross-modal attention mechanism. Based on the cross-correlation weights, weighted feature extraction is performed on the aligned sensing data matrix to obtain cross-modal fusion features, and the cross-modal fusion features are determined as the gas parameter features; The gas parameter features and the foot slip features are concatenated to generate the perception state vector.

5. The method according to claim 1, characterized in that, The extraction steps for the foot slippage features in the perceived state vector specifically include: Extract multi-axis contact force data from the sensor data of the foot force sensor; The multi-axis contact force data is denoised using a preset sliding window filtering algorithm to obtain smooth contact force data. Extract the horizontal tangential force component and the vertical normal force component from the smooth contact force data, and calculate the ratio of the horizontal tangential force component to the vertical normal force component to obtain the real-time friction coefficient estimate. The difference between the preset critical slip threshold and the estimated real-time friction coefficient is calculated to obtain the foot slip characteristics.

6. The method according to claim 1, characterized in that, The step of adjusting the anti-slip control sensitivity based on the concentration change trend in the gas parameter characteristics specifically includes: Extract the magnitude of the spatial concentration gradient vector from the gas parameter features, calculate the time derivative of the magnitude within a continuous preset time window, and determine the time derivative as the concentration change trend; The determination result is obtained by comparing the concentration change trend with the preset concentration alarm threshold range, and the environmental pollution level is determined based on the determination result. Extract the corresponding preset sensitivity compensation coefficient based on the environmental pollution level; The preset sensitivity compensation coefficient is added to the preset anti-slip sensitivity base value to obtain the adjusted anti-slip control sensitivity.

7. The method according to claim 6, characterized in that, The process of comparing the concentration change trend with a preset concentration alarm threshold range to obtain a judgment result, and determining the environmental pollution level based on the judgment result, specifically includes: Extract the positive and negative sign features and absolute values ​​corresponding to the concentration change trend; When the positive and negative sign features are positive and the absolute value is greater than or equal to a preset first derivative threshold, the environmental pollution level is mapped to a high-risk leakage pollution level. When the positive and negative sign features are positive and the absolute value is less than the first derivative threshold and greater than or equal to the preset second derivative threshold, the environmental pollution level is mapped to a potential diffusion pollution level. When the positive or negative sign feature is negative or the absolute value is less than the second derivative threshold, the environmental pollution level is mapped to a safe attenuation pollution level, wherein the first derivative threshold is greater than the second derivative threshold, and the first derivative threshold and the second derivative threshold constitute the determination boundary of the preset concentration alarm threshold range.

8. A perception and coordination application system for a quadruped robot, applied to a quadruped robot equipped with an edge computing unit and multimodal sensors, wherein the multimodal sensors include a spatial vision sensor, an inertial measurement unit, a foot force sensor, and an olfactory sensor, characterized in that, The system includes: The timing synchronization module is used to acquire a clock synchronization signal and use the clock synchronization signal as a general time reference to perform clock alignment on the multimodal sensor, and acquire multimodal sensing data with the corresponding sampling frequency based on the aligned time reference; The spatial mapping module is used to extract high-frequency pose estimation data output by the inertial measurement unit and the spatial vision sensor from the multimodal sensing data, with the trigger time of the clock synchronization signal as the reference time, and to perform integral compensation on the spatial displacement of the olfactory sensor in the low-frequency sampling period using the high-frequency pose estimation data, and to map the scalar chemical data output by the olfactory sensor into a spatial concentration gradient vector. The feature fusion module is used to input the multimodal sensing data containing the spatial concentration gradient vector into the edge computing unit to perform cross-modal feature fusion and generate a sensing state vector, wherein the sensing state vector includes gas parameter features generated by the transformation of the spatial concentration gradient vector and foot slip features extracted based on the sensing data of the foot force sensor. An anti-slip adjustment module is used to perform motion planning and decision-making for the quadruped robot based on the perceived state vector, and to adjust the anti-slip control sensitivity for the foot slip characteristics according to the concentration change trend in the gas parameter characteristics. The gait control module is used to determine the potential pollution area based on the concentration change trend, and generate support phase gait adjustment control commands according to the relative spatial distance between the quadruped robot and the potential pollution area.

9. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions. The user interface and the network interface are both used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images