A method and system for adaptive gait control of robotic dogs in complex terrain

Abstract

This invention relates to the field of adaptive control technology, specifically to a gait adaptive control method and system for robotic dogs facing complex terrain. The specific implementation process includes: quantifying external disturbances generated by unexpected physical interactions through an endogenous disturbance observer; generating compensation control signals using a configurable gait control loop, and adjusting the actual gait by fusing it with the basic gait using a feedforward regulator; continuously collecting and generating historical disturbance sequences, and analyzing and extracting persistent disturbance factors; and using the persistent disturbance factors to perform online adaptive correction of the gait dynamics model and replanning the basic motion. This invention constructs a dual adaptive control architecture, combining feedforward compensation for instantaneous disturbances with online adaptive correction of persistent deviations. This effectively improves the problem of insufficient robustness of robotic dogs facing complex terrain, enhancing the walking ability of robotic dogs in complex terrain.

Description

Technical Field

[0001] This invention relates to the field of adaptive control technology, specifically to a gait adaptive control method and system for a robot dog oriented towards complex terrain. Background Technology

[0002] In the field of robot dog control and regulation, existing technologies typically utilize Model Predictive Control (MPC) to achieve adaptive gait control for robot dogs in complex terrain. MPC uses a pre-established dynamic model describing the robot dog and its interaction with the environment, and performs iterative online optimization calculations via an onboard computer. This predicts the robot dog's motion state over a finite future time domain and solves for a series of optimal control inputs (such as joint torques or foot forces) that achieve the desired goal (e.g., stable walking).

[0003] However, existing technologies have inherent limitations in practical applications. The dynamic model established by MPC is essentially an idealized simplification of the physical world and cannot accurately describe complex nonlinear time-varying factors such as abrupt changes in ground friction coefficients, minute actuator delays, and sensor noise. When the robot dog encounters sudden external disturbances not established by MPC (such as its feet suddenly sinking into soft ground or slipping on wet surfaces), MPC's online optimizer will still perform calculations based on this inaccurate internal model. This results in insufficient ability to cope with sudden disturbances, thus affecting the robustness of the control system in complex terrain.

[0004] To address this, a gait adaptive control method and system for robotic dogs in complex terrain is proposed. Summary of the Invention

[0005] The purpose of this invention is to provide a gait adaptive control method and system for robot dogs in complex terrain, so as to realize the gait adaptive control of robot dogs in complex terrain.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An adaptive gait control method for robotic dogs in complex terrain includes:

[0008] A gait dynamics model is constructed to plan the basic gait of the robot dog under a preset complex terrain. The actual gait of the robot dog is obtained. The basic gait and the actual gait are decoupled by an endogenous disturbance observer to obtain the gait residual generated by unexpected physical interactions. The gait residual is then quantified into an exogenous disturbance vector.

[0009] The exogenous disturbance vector is input into the gait control loop to obtain a compensation control signal; the compensation control signal and the basic gait are fused by a bandwidth feedforward regulator to generate a gait control command to adjust the actual gait.

[0010] Exogenous interference vectors generated during multiple motion cycles are continuously collected to form a historical interference sequence; gait interference analysis is performed on the historical interference sequence to extract persistent interference factors; the persistent interference factors are input into an adaptive correction loop to adaptively correct the gait dynamics model and replan the basic gait.

[0011] Preferably, the specific implementation process of constructing a gait dynamics model and planning the basic movement gait of the robot dog under preset complex terrain includes:

[0012] A gait dynamics model including trunk inertial parameters and joint degrees of freedom is established; combined with the geometric feature data of the pre-set complex terrain, the center of mass trajectory and foot landing point are used as optimization targets; by setting periodic gait functions and stability constraints, the gait dynamics model is solved to generate a basic gait including the driving torques and motion sequences of each joint.

[0013] Preferably, the specific implementation process of obtaining the actual gait of the robot dog and obtaining the gait residual through the endogenous disturbance observer includes: acquiring trunk posture data and joint motion data in real time through an inertial measurement unit and a joint encoder to obtain the actual gait; inputting the actual gait and the basic gait into the endogenous disturbance observer, which estimates the deviation of the actual gait and the basic gait based on the gait dynamics model; separating the deviation caused by unexpected physical interactions from the deviation estimation results to generate the gait residual.

[0014] Preferably, the specific implementation process of quantizing the motion gait residual into an exogenous interference vector includes: transforming the motion gait residual into a dynamic data stream containing external forces and internal parameter deviations through Jacobian matrix transformation; performing high-frequency measurement noise filtering on the dynamic data stream and mapping it to the body coordinate system to obtain a six-dimensional vector containing three force components and three torque components; and generating an exogenous interference vector based on the current dynamic state of the robot dog from the six-dimensional vector.

[0015] Preferably, the specific implementation process of inputting the exogenous interference vector into the gait control loop to obtain the compensation control signal includes: the gait control loop has a preset dynamic gain matrix, which is configured according to the robot dog's expected compliance with external interference; the gait control loop performs matrix multiplication on the exogenous interference vector and the dynamic gain matrix to generate a compensation force and a compensation torque that are proportional to the magnitude of the exogenous interference vector and opposite in direction, and combines the compensation force and the compensation torque into a compensation control signal.

[0016] Preferably, the specific implementation process of fusing the compensation control signal and the basic gait through a bandwidth feedforward regulator to generate gait control commands to adjust the actual gait includes: the bandwidth feedforward regulator, according to the configured bandwidth parameters, feeds forward the compensation force and compensation torque contained in the compensation control signal onto the basic gait through a linear superposition operation, and performs frequency response shaping to generate gait control commands; the gait control commands control the joint drive of the robot dog in real time, thereby adjusting the actual gait.

[0017] Preferably, the specific implementation process of continuously acquiring exogenous interference vectors generated within multiple motion cycles to form a historical interference sequence includes:

[0018] Set up a data acquisition window spanning multiple motion cycles, continuously acquire exogenous interference vectors generated in the current motion cycle at the sampling frequency of the endogenous interference observer; attach a motion cycle index to each acquired exogenous interference vector, mark its time position during the motion process, and store it in a circular buffer memory in sequence; when the exogenous interference vectors in the circular buffer memory reach a preset threshold, output all stored exogenous interference vectors as a whole to generate a historical interference sequence.

[0019] Preferably, the specific implementation process of performing gait interference analysis on the historical interference sequence to extract the persistent interference factor includes: applying a low-pass filter to each exogenous interference vector in the historical interference sequence, extracting a low-frequency trend term whose amplitude and direction exhibit periodicity within multiple motion cycles, and outputting it as a persistent interference factor; the low-frequency trend term characterizes the gait deviation caused by changes in external terrain.

[0020] Preferably, the specific implementation process of inputting the persistent interference factor into the adaptive correction loop to adaptively correct the gait dynamics model and replan the basic movement gait includes: the adaptive correction loop receiving the persistent interference factor, mapping it to a correction vector of the model parameters, and using the correction vector to update the gait dynamics model; the updated gait dynamics model, combined with the geometric feature data of the preset complex terrain, replans the adaptively corrected basic movement gait.

[0021] An adaptive gait control system for robotic dogs in complex terrain includes:

[0022] The interference quantization module constructs a gait dynamics model and plans the basic gait of the robot dog under a preset complex terrain. It obtains the actual gait of the robot dog, decouples the basic gait from the actual gait through an endogenous interference observer, obtains the gait residual generated by unexpected physical interactions, and quantifies the gait residual into an exogenous interference vector.

[0023] The gait control module inputs the exogenous disturbance vector into the gait control loop to obtain a compensation control signal; the compensation control signal is then fused with the basic gait through a bandwidth feedforward regulator to generate gait control commands to adjust the actual gait.

[0024] The adaptive correction module continuously collects exogenous interference vectors generated within multiple motion cycles to form a historical interference sequence; performs gait interference analysis on the historical interference sequence to extract persistent interference factors; inputs the persistent interference factors into the adaptive correction loop to adaptively correct the gait dynamics model and replan the basic gait.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. This invention combines compensatory control with model calibration, enabling the robot dog to respond quickly to external disturbances such as sudden terrain changes, thus improving upon the computational delay inherent in model predictive control. This results in higher stability and adaptability of the robot dog in complex and ever-changing environments, effectively enhancing the robustness of the control system.

[0027] 2. The endogenous disturbance observer proposed in this invention quantifies unmodeled internal and external disturbances into dynamic residuals, effectively reducing the complexity of identifying disturbance sources. This enables the robot dog to quickly compensate for uncertainties, including terrain material, slope changes, and its own parameter deviations, thus broadening its adaptability to complex and unknown terrains.

[0028] 3. The persistent interference factor extraction mechanism based on historical interference sequences used in this invention learns online and identifies long-standing non-random interference factors, automatically adjusting internal model parameters to eliminate biases. This enables the robot dog to self-optimize during movement, adapting to gradual effects caused by load changes or wear of its own components. Attached Figure Description

[0029] Figure 1 This is a flowchart of a machine dog gait adaptive control method for complex terrain proposed in this invention;

[0030] Figure 2 This is a structural diagram of a machine dog gait adaptive control system for complex terrain proposed in this invention;

[0031] Figure 3 This is a schematic diagram of the dual adaptive control architecture for the robot dog's gait proposed in this invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It must be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to constitute any limitation on the scope of protection of this invention. Therefore, all equivalent changes or modifications conceived by those skilled in the art based on the content disclosed in this invention without inventive effort should fall within the scope of protection claimed by this invention.

[0033] Reference Figures 1 to 3 This invention proposes a gait adaptive control method and system for robotic dogs in complex terrain, the technical solution of which is as follows:

[0034] Example 1:

[0035] Reference Figure 1 This embodiment proposes a gait adaptive control method for robotic dogs in complex terrain, including:

[0036] A gait dynamics model is constructed to plan the basic gait of the robot dog under a preset complex terrain. The actual gait of the robot dog is obtained. The basic gait and the actual gait are decoupled by an endogenous disturbance observer to obtain the gait residual generated by unexpected physical interactions. The gait residual is then quantified into an exogenous disturbance vector.

[0037] The exogenous disturbance vector is input into the gait control loop to obtain a compensation control signal; the compensation control signal and the basic gait are fused by a bandwidth feedforward regulator to generate a gait control command to adjust the actual gait.

[0038] Exogenous interference vectors generated during multiple motion cycles are continuously collected to form a historical interference sequence; gait interference analysis is performed on the historical interference sequence to extract persistent interference factors; the persistent interference factors are input into an adaptive correction loop to adaptively correct the gait dynamics model and replan the basic gait.

[0039] Furthermore, the specific implementation process of constructing a gait dynamics model and planning the basic movement gait of the robot dog under preset complex terrain includes:

[0040] A gait dynamics model including trunk inertial parameters and joint degrees of freedom is established; combined with the geometric feature data of the pre-set complex terrain, the center of mass trajectory and foot landing point are used as optimization targets; by setting periodic gait functions and stability constraints, the gait dynamics model is solved to generate a basic gait including the driving torques and motion sequences of each joint.

[0041] Specifically, a gait dynamics model is established that includes information on the robot dog's mass, center of mass position, moment of inertia, and degrees of freedom of each joint. For example, a robot dog's leg typically includes three degrees of freedom at the hip joint and one degree of freedom at the knee joint. These parameters together constitute a complete set of variables describing the robot dog's motion state.

[0042] After the gait dynamics model is built, the basic movement gait needs to be planned in conjunction with preset complex terrain information. The preset terrain here utilizes the geometric features of complex terrain, such as the average slope of the ground, the approximate height and spacing of obstacles, etc. This data is obtained through the robot dog's onboard vision and LiDAR sensors, and can also be used as preset parameter inputs. The core of the planning lies in using the robot dog's center of mass trajectory and the foot placement points of each leg as joint optimization objectives. For example, when traversing a slope with a 15-degree incline, to maintain stability, the optimization objective will plan a center of mass trajectory that is as smooth as possible and close to the slope surface, while simultaneously calculating the sequence of foot placement points that provides the maximum support polygon and avoids interference with obstacles.

[0043] To plan the basic gait, this embodiment introduces periodic gait functions and stability constraints. Periodic gait functions, such as cycloidal functions, are used to generate rhythmic leg swings and support phases that conform to the natural laws of biological walking. For example, a standard "diagonal trot" gait is defined, where the two diagonally opposite legs swing synchronously. To ensure the robot dog does not tip over during movement, stability constraints are simultaneously applied. For example, the zero-moment point must always remain within the support polygon formed by the foot of the current supporting leg. In the basic gait planning for the 15-degree slope described above, the position of the zero-moment point at each instant is continuously verified to ensure it does not exceed the triangular area enclosed by the foot in contact with the ground. By combining periodic gait functions and stability constraints, a complete basic gait is calculated. The basic gait not only includes the detailed movement sequence of the robot dog over multiple future movement cycles, such as the target angles, angular velocities, and angular accelerations of each joint, but also the driving torques of each joint required to achieve these movements.

[0044] This embodiment effectively improves the robot dog's adaptability in complex environments by establishing a gait dynamics model that includes trunk inertia and joint degrees of freedom, and by combining complex terrain feature data to plan the basic movement gait. By introducing stability constraints such as periodic gait functions and zero moment points, the planned basic movement gait is ensured to be physically feasible.

[0045] Furthermore, the specific implementation process of obtaining the actual gait of the robot dog and obtaining the gait residual through the endogenous disturbance observer includes: acquiring trunk posture data and joint motion data in real time through an inertial measurement unit and a joint encoder to obtain the actual gait; inputting the actual gait and the basic gait into the endogenous disturbance observer, which estimates the deviation of the actual gait and the basic gait based on the gait dynamics model; separating the deviation caused by unexpected physical interactions from the deviation estimation results to generate the gait residual.

[0046] Specifically, when the robot dog moves according to its pre-planned basic gait, its actual gait is collected at high frequency by an onboard inertial measurement unit (IMU) and high-precision joint encoders distributed across various drive joints. For example, the IMU installed near the robot dog's center of mass continuously collects the three-axis attitude angles (roll, pitch, and yaw) and angular velocity data of the torso at a sampling rate of 500 Hz, forming a real-time description of the robot dog's overall posture; the encoders installed at the hip and knee joints of each leg collect the rotation angles and angular velocities of each joint in real time at a frequency of 1000 Hz. The data is then fused and processed by the onboard computer to form a dynamic, complete vector, i.e., the robot dog's actual gait.

[0047] An endogenous disturbance observer based on generalized momentum is used to input the real-time acquired actual motion state into the established gait dynamics model, back-calculating the theoretically existing joint driving torque under this actual motion, and comparing the solution with the expected force and torque corresponding to the basic gait. The total deviation between the two includes both the inaccuracy of the model itself and the interference caused by external environmental interactions. Subsequently, the endogenous disturbance observer separates the part caused by unexpected physical interactions from the total deviation. For example, a 30 kg robot dog is walking on a preset flat concrete surface. The basic gait plan is for its right forefoot to land at a speed of 0.2 m / s in the next moment, and it is expected to obtain an instantaneous ground support force of about 75 N. However, the actual ground is covered with a soft sponge with a thickness of 5 cm that was not preset. When the robot dog's right forefoot touches the sponge, the actual situation deviates from the plan: the inertial measurement unit detects that the torso has a 2-degree forward tilt that is not planned, and at the same time, the right foreleg knee joint encoder data shows that the joint is compressed by 0.03 meters more than expected. After receiving this actual motion data, the endogenous disturbance observer analyzed it based on the gait dynamics model and found that the additional trunk forward lean and joint compression could not be explained by the model; that is, the model's intrinsic dynamics could not produce such results. Therefore, the endogenous disturbance observer attributed this unexplainable bias to external disturbances caused by unexpected physical interactions (i.e., stepping on soft sponges) and separated it into gait residuals.

[0048] This embodiment generates a motion gait residual representing external physical interaction through the separation mechanism of the endogenous disturbance observer, enabling the robot dog to make rapid and appropriate responses. This enhances the robustness of the control system and broadens the robot dog's adaptability to complex and unknown terrain.

[0049] Furthermore, the specific implementation process of quantizing the motion gait residual into an exogenous disturbance vector includes: transforming the motion gait residual into a dynamic data stream containing external forces and internal parameter deviations through Jacobian matrix transformation; performing high-frequency measurement noise filtering on the dynamic data stream and mapping it to the body coordinate system to obtain a six-dimensional vector containing three force components and three torque components; and generating an exogenous disturbance vector based on the current dynamic state of the robot dog from the six-dimensional vector.

[0050] Specifically, this embodiment transforms the gait residuals using a Jacobi matrix. The Jacobi matrix is ​​derived from the robot dog's gait kinematics model, establishing a linear mapping between joint space angular velocity and the Cartesian space linear and angular velocities of the end effector. This inversely maps the forces and torques acting on the feet into torques generated at each joint. The gait residuals, containing information such as joint compression and trunk posture deviation, are transformed into a dynamic data stream containing information on external actual forces and internal model parameter deviations using the Jacobi matrix.

[0051] The power data stream inevitably contains high-frequency measurement noise generated by the joint encoder and inertial measurement unit. To eliminate the interference of this noise on subsequent control, this embodiment uses a second-order Butterworth low-pass filter with a cutoff frequency of 50 Hz to filter the power data stream. After filtering, the power data stream is integrated into a unified six-dimensional vector in the body coordinate system. This six-dimensional vector strictly follows the wrench definition in robot dynamics, with the first three components representing the forces acting on the body's center of mass along its front, left, and top orthogonal axes, respectively, and the last three components representing the torques along these three axes.

[0052] Based on the current dynamic state of the robot dog, the six-dimensional vector is identified as the exogenous disturbance vector. For example, the motion gait residual (2-degree forward tilt, 0.03-meter joint compression) output by the endogenous disturbance observer, after Jacobian matrix transformation and noise filtering, is represented as a six-dimensional vector acting in the fuselage coordinate system, with specific values ​​of (Fx=0 N, Fy=0 N, Fz=-51 N, Tx=0 N·m, Ty=5 N·m, Tz=0 N·m). This six-dimensional vector clearly indicates that the robot dog is currently experiencing a 51-Newton vertical downward pressure (generated by the deformation of the sponge) and a 5-Newton pitching moment that causes it to tilt forward. This six-dimensional vector, after the complete processing, is the exogenous disturbance vector.

[0053] This embodiment establishes a Jacobian matrix and uses high-frequency noise filtering to transform the complex motion gait residuals into a physically meaningful six-dimensional vector. This enables the controller to perform precise compensation control for different components of force and torque, effectively improving the accuracy and stability of disturbance estimation.

[0054] Furthermore, the specific implementation process of inputting the exogenous interference vector into the gait control loop to obtain the compensation control signal includes: the gait control loop has a preset dynamic gain matrix, which is configured according to the robot dog's expected compliance with external interference; the gait control loop performs matrix multiplication on the exogenous interference vector and the dynamic gain matrix to generate a compensation force and a compensation torque that are proportional to the magnitude of the exogenous interference vector and opposite in direction, and combines the compensation force and the compensation torque into a compensation control signal.

[0055] Specifically, the exogenous disturbance vector is input into the gait control loop for gain calculation. The core of the gait control loop is a 6x6 dynamic gain matrix. This matrix is ​​not a fixed set of constants, but rather configured based on the current task requirements and environmental characteristics, focusing on the robot dog's expected compliance with external disturbances. Expected compliance is a key parameter describing the robot dog's "rigidity" or "flexibility" when subjected to external forces. For example, when the robot dog performs a task requiring walking on slippery ground, its expected compliance is configured to a high level to prevent slipping due to rigid collisions. Under this configuration, the main gain value on the diagonal of the dynamic gain matrix is ​​set to a relatively low value; for example, the gain coefficient corresponding to vertical forces is set to 50. Conversely, when the robot dog needs to perform high-precision load handling tasks on flat, hard surfaces, where any slight attitude deviation could lead to task failure, its expected compliance is configured to a very low level (i.e., high stiffness). In this case, the same gain coefficient in the dynamic gain matrix is ​​set to 500.

[0056] When an exogenous disturbance vector carrying external disturbance information enters the gait control loop, the processor within the loop immediately performs a matrix multiplication operation, multiplying the six-dimensional exogenous disturbance vector by the currently configured dynamic gain matrix. The disturbance component is weighted to calculate the required compensation force. For example, if the exogenous disturbance vector is (horizontal force Fx = 0 N, pitch force Fy = 0 N, vertical force Fz = -51 N, horizontal moment Tx = 0 N·m, pitch moment Ty = 5 N·m, vertical moment Tz = 0 N·m), and the desired compliance configuration is medium stiffness, the gain corresponding to the vertical force (Fz) in the dynamic gain matrix is ​​80, and the gain corresponding to the pitch moment (Ty) is 60. The matrix multiplication operation performed by the gait control loop will generate a compensation force and a compensation moment that are proportional to the magnitude of the disturbance vector and have the opposite direction. For a vertical downward disturbance force of -51 N, a vertical upward compensation force is generated; for a pitch moment of 5 N·m, a reverse pitch compensation moment is generated. These two newly generated compensation forces and compensation moments, together with the zero values ​​of other components, are combined into a new six-dimensional vector, which is the compensation control signal.

[0057] This embodiment introduces a dynamic gain matrix that can be configured according to desired compliance, endowing the robot dog with behavioral adaptability. The robot dog can switch between "rigid" and "flexible" modes depending on the task, broadening its application scenarios in variable environments. The compensation control signal generated using matrix multiplication improves the robot dog's gait balance ability in complex dynamic environments.

[0058] Furthermore, the specific implementation process of fusing the compensation control signal and the basic gait using a bandwidth feedforward regulator to generate gait control commands to adjust the actual gait includes: the bandwidth feedforward regulator, based on the configured bandwidth parameters, feeds forward the compensation force and compensation torque contained in the compensation control signal onto the basic gait through a linear superposition operation, and performs frequency response shaping to generate gait control commands; the gait control commands control the joint drive of the robot dog in real time, thereby adjusting the actual gait.

[0059] Specifically, the compensation control signal output from the gait control loop and the robot dog's basic gait are simultaneously fed into the bandwidth feedforward regulator. The bandwidth parameter of the bandwidth feedforward regulator defines the frequency range of the compensation control signal. For example, in a scenario where the robot dog needs to quickly traverse a gravel slope, the bandwidth parameter can be configured to 20 Hz to cope with high-frequency ground impacts, allowing the robot dog to react agilely. In a task requiring the robot dog to smoothly transport precision instruments, the bandwidth parameter is configured to 5 Hz to suppress any possible high-frequency jitter, aiming for motion stability.

[0060] Upon receiving the compensation control signal, the bandwidth feedforward regulator performs a linear superposition operation based on the frequency range of the set bandwidth parameters. The compensation force and torque contained in the compensation control signal are superimposed on the basic gait, which includes a series of expected joint driving torques, in a feedforward manner. This superposition is then distributed to the control motors of each joint in the robot dog's four legs, adjusting the force and torque to counteract external interference. For example, one compensation control signal is equivalent to a 30 N vertical upward support force and a -7.5 N·m reverse pitch torque. Subsequently, these compensation torques are added joint-by-joint to the original torques of the corresponding joints in the basic gait.

[0061] After linear superposition, the regulator performs frequency response shaping. For example, with the bandwidth configured at 15 Hz, the frequency response shaping process uses a digital low-pass filter to attenuate all components with frequencies higher than 15 Hz after superposition. The resulting signal retains the ability to respond quickly to interference within 15 Hz while filtering out unstable variations, forming gait control commands. These gait control commands are a set of real-time data streams corresponding to the target torque that the robot dog's joint motors need to output in the next control cycle (e.g., 1 millisecond). Finally, this shaped and fused gait control command set is sent to the robot dog's underlying joint driver. The driver controls the required torque of the motors in real-time with high precision, thereby directly adjusting the robot dog's actual gait.

[0062] This embodiment integrates compensation control and basic planning through a bandwidth feedforward regulator, ensuring the stability of the control system, improving the shortcomings of feedforward control that are prone to oscillation due to inaccurate models or external noise, and effectively enhancing the robustness of adaptive control.

[0063] Furthermore, the specific implementation process of continuously acquiring exogenous interference vectors generated within multiple motion cycles to form a historical interference sequence includes:

[0064] Set up a data acquisition window spanning multiple motion cycles, continuously acquire exogenous interference vectors generated in the current motion cycle at the sampling frequency of the endogenous interference observer; attach a motion cycle index to each acquired exogenous interference vector, mark its time position during the motion process, and store it in a circular buffer memory in sequence; when the exogenous interference vectors in the circular buffer memory reach a preset threshold, output all stored exogenous interference vectors as a whole to generate a historical interference sequence.

[0065] Specifically, to extract regular information from continuous interference signals, a data acquisition window spanning multiple consecutive motion cycles is set. For example, a robot dog walking on sand has a complete motion cycle (i.e., each of its four legs completes one step) of 0.5 seconds. To effectively identify subsidence interference caused by the sand, the data acquisition window is set to cover 8 consecutive motion cycles, i.e., a total duration of 4 seconds. Within this window duration, the exogenous interference vector generated in each control cycle is continuously acquired at the 200 Hz sampling frequency of the endogenous interference observer itself. That is, every 5 milliseconds, an exogenous interference vector representing the external interference at that instant is recorded.

[0066] To preserve the temporal context information of each data point throughout the entire motion process, a motion cycle index is attached to each newly acquired exogenous interference vector. This index acts as a data label, explicitly marking which motion cycle the exogenous interference vector was generated within. For example, in a 4-second acquisition window, the first 100 interference vectors acquired in the first 0.5 seconds are labeled with index "1," the next 100 vectors are labeled with index "2," and so on up to index "8." All indexed exogenous interference vectors are sequentially stored in a circular buffer in the onboard memory according to their generation time, with a storage threshold of 800. When the circular buffer is completely full, the 800 exogenous interference vectors stored in the memory are output as a complete data set. This data set, containing all exogenous interference information within the most recent 4 seconds, constitutes the historical interference sequence.

[0067] This embodiment achieves efficient and continuous recording of recent interference history by setting a cross-cycle acquisition window and using a cyclic buffer. This ensures the accuracy of interference nature identification, thereby effectively improving the robot dog's long-term robustness in complex environments.

[0068] Furthermore, the specific implementation process of performing gait interference analysis on the historical interference sequence and extracting the persistent interference factor includes: applying a low-pass filter to each exogenous interference vector in the historical interference sequence, extracting the low-frequency trend term whose amplitude and direction exhibit periodicity within multiple motion cycles, and outputting it as the persistent interference factor; the low-frequency trend term characterizes the gait deviation caused by changes in external terrain.

[0069] Specifically, the low-frequency trend term extraction process involves independently applying a fourth-order Butterworth filter to each exogenous interference vector in the historical interference sequence, using a cutoff frequency much lower than the robot dog's gait frequency. This filters out interference signals generated by instantaneous high-frequency events such as a single slip or stepping on a small pebble, retaining only the interference components that persist and change slowly over a timescale of several seconds.

[0070] After low-pass filtering, a smooth interference trend data stream is obtained. Using the motion cycle index previously attached to each exogenous interference vector, the sequence is segmented into subsequences corresponding to each independent motion cycle, and these subsequences are aligned in phase. Subsequently, Fourier transform is used to identify whether there is any recurring interference with the same frequency as the motion cycle.

[0071] For example, a robot dog begins to walk continuously uphill on a slope with a gradient of 10 degrees. After low-pass filtering the historical interference sequence (e.g., data containing 8 gait cycles), it is found that in each movement cycle, the horizontal force (Fx) component representing the forward and backward directions consistently exhibits a reaction force of -51 N; simultaneously, the torque (Ty) component representing the pitch direction also exhibits a sinusoidal fluctuation with a period of 0.5 seconds synchronized with the movement, with an amplitude of 4 N·m. The -51 N DC component and the periodic 4 N·m torque fluctuation are extracted and combined into a vector function that varies with the phase of the movement. This vector function is the low-frequency trend term for this scenario and is formally output as the persistent interference factor.

[0072] This embodiment uses low-pass filtering and periodic analysis to identify motion deviations caused by environmental changes from complex raw external interference data. The extracted persistent interference factors provide a crucial basis for subsequent adaptive correction, effectively improving the robot dog's long-term adaptability in unknown and changing environments.

[0073] Furthermore, the specific implementation process of inputting the persistent interference factor into the adaptive correction loop to adaptively correct the gait dynamics model and replan the basic movement gait includes: the adaptive correction loop receives the persistent interference factor, maps it to a correction vector of the model parameters, and uses the correction vector to update the gait dynamics model; the updated gait dynamics model, combined with the geometric feature data of the preset complex terrain, replans the adaptively corrected basic movement gait.

[0074] Specifically, the adaptive correction loop characterizes the persistent disturbance factor of external disturbances, infers it in reverse, and maps it to a correction vector to correct the parameters of the internal gait dynamics model. The mapping process is based on inference rules learned online to determine which model parameter mismatch is most likely to cause the persistent disturbance.

[0075] For example, the persistent disturbance factor received by the adaptive correction loop is a function of a constant backward force of -51 N and a periodic pitching moment of 4 N·m. The inference rules of the adaptive correction loop analyze the characteristics of this external disturbance and conclude that a constant external force opposite to the direction of travel originates from the component of gravity on the inclined plane. Based on this inference, the loop uses the robot's own mass parameters (e.g., 30 kg) to calculate the equivalent slope angle causing this -51 N force, which is approximately 10 degrees. Therefore, the loop generates a correction vector specifically for the environmental model parameters, physically representing an uphill slope of 10 degrees in the current environment.

[0076] In the constructed gait dynamics model, the parameters describing the external environment assume the ground is level. After applying correction vectors, the calculations regarding gravity in the model will be revised to account for the decomposition effect of gravity at a slope of 10 degrees.

[0077] The gait dynamics model, after online learning and adaptive correction, is combined with pre-set complex terrain geometry data (even if the data does not explicitly indicate the slope) to re-plan the robot dog's basic gait. The new basic gait will include the additional forward driving torque required to resist gravity from the beginning, and will automatically plan a body posture with a slightly forward shift of the center of gravity to adapt to uphill.

[0078] This embodiment achieves online learning and self-improvement of the robot dog's gait by mapping persistent interference factors to the correction of model parameters. This effectively improves motion efficiency, reduces the extra energy consumed to combat unexpected external interference, and enhances the robot dog's long-term adaptability to unknown environments.

[0079] This embodiment proposes a gait adaptation method for robot dogs facing complex terrain. By constructing a dual-closed-loop adaptive control architecture, it improves the limitation of robot dog gait control methods in adapting to complex terrain.

[0080] The proposed endogenous disturbance observer can uniformly quantify disturbances generated by unexpected physical interactions into exogenous disturbance vectors, effectively simplifying the complexity of disturbance source identification. By using a bandwidth feedforward regulator to feedforward fuse the compensation control signal with the basic motion gait, rapid and active compensation for disturbances is achieved, effectively improving motion capabilities in complex environments.

[0081] The established adaptive correction loop is no longer limited to passive compensation for instantaneous external disturbances. Through continuous analysis of historical disturbance sequences, it intelligently identifies and extracts persistent disturbance factors caused by environmental changes, and adaptively corrects the gait dynamics model, enabling the robot dog to have online learning and self-optimization capabilities.

[0082] Example 2:

[0083] This embodiment provides a gait adaptive control system for robotic dogs in complex terrain, applicable to gait control of robotic dogs. (Refer to...) Figure 2 The system includes an interference quantization module, a gait control module, and an adaptive correction module.

[0084] The interference quantization module constructs a gait dynamics model and plans the basic gait of the robot dog under a preset complex terrain. It obtains the actual gait of the robot dog, decouples the basic gait from the actual gait through an endogenous interference observer, obtains the gait residual generated by unexpected physical interactions, and quantifies the gait residual into an exogenous interference vector.

[0085] The gait control module inputs the exogenous disturbance vector into the gait control loop to obtain a compensation control signal; the compensation control signal is then fused with the basic gait through a bandwidth feedforward regulator to generate gait control commands to adjust the actual gait.

[0086] The adaptive correction module continuously collects exogenous interference vectors generated within multiple motion cycles to form a historical interference sequence; performs gait interference analysis on the historical interference sequence to extract persistent interference factors; inputs the persistent interference factors into the adaptive correction loop to adaptively correct the gait dynamics model and replan the basic gait.

[0087] Furthermore, the interference quantification module constructs a gait dynamics model and plans the robot dog's basic gait in a preset complex terrain based on this model. The gait dynamics model includes not only inertial parameters such as the robot dog's torso mass, center of mass position, and moment of inertia, but also describes the degrees of freedom of each joint in each leg. Combined with the complex terrain geometric feature data obtained from airborne sensors (e.g., lidar), such as the presence of a 15-degree slope ahead, and using the stability of the center of mass trajectory and foot landing point as a joint optimization objective, the gait dynamics model is solved by setting a periodic gait function and zero-moment point constraints, thereby generating a complete basic gait. The basic gait is a series of refined instructions, containing the target angles, angular velocities, and expected driving torques required to achieve these movements for each joint in the next few seconds. After the robot dog begins to move based on its basic gait, an inertial measurement unit (IMU) installed in its torso continuously collects the torso's posture and angular velocity at a frequency of 500 Hz. Simultaneously, encoders at each joint report angular information at a frequency of 1000 Hz. This data fusion constitutes the robot dog's actual gait. An endogenous disturbance observer takes both the actual gait and the planned basic gait as input and compares them against an established gait dynamics model. It identifies deviations in the actual gait that cannot be explained by the internal gait dynamics model. For example, when the robot dog's right forefoot accidentally steps on a 5 cm thick soft sponge that was not pre-detected, the IMU detects an unplanned 2-degree forward tilt of the torso, and the knee encoder shows an additional 0.03 m compression. The observer determines that this deviation does not originate from the model's own dynamic characteristics but from an unexpected external physical interaction. Based on this, it separates and outputs this kinematic deviation information, including the 2-degree forward tilt and 0.03 m compression, as a gait residual signal. Through Jacobian matrix transformation, the kinematic residuals are converted into a dynamic data stream, which initially contains information about external forces and torques. A low-pass filter with a cutoff frequency of 50 Hz is applied to the data stream to remove high-frequency measurement noise. After filtering, the data is integrated into a six-dimensional vector represented in the fuselage coordinate system, where the first three terms represent force components along the three fuselage axes, and the last three terms represent torque components. After complete processing, the module's final output exogenous disturbance vector clearly quantifies the forces and torques of external disturbances.

[0088] Furthermore, the gait control module converts the input exogenous disturbance vector into a compensation control signal. This process is completed within a gait control loop configured with a six-by-six dynamic gain matrix. When the robot dog needs to walk on slippery ice, to avoid slipping due to rigid contact, the desired compliance of the gait control loop is configured to a high level. In this case, the main diagonal elements in the gain matrix are set to low values, corresponding to a gain coefficient of 50 for the vertical force. Conversely, when performing high-precision platform stabilization tasks, to resist any minor disturbances, the desired compliance is configured to an extremely low level (high stiffness), and the gain coefficient is set to 500. When the exogenous disturbance vector is input into the loop, the loop processor immediately performs a matrix multiplication operation, multiplying the vector by the currently configured dynamic gain matrix. At this time, the system is set to medium stiffness, and based on the set corresponding gain, a compensation force and torque of proportional magnitude and opposite direction are generated. These two values, along with other zero-value components, combine to form a new six-dimensional vector, i.e., the compensation control signal. A bandwidth feedforward regulator fuses the compensation control signal with the basic gait to generate gait control commands. The bandwidth parameter configured in the bandwidth feedforward regulator defines the frequency range of the compensation control signal. In movements requiring rapid traversal of scree slopes, the bandwidth parameter is set to 20 Hz to cope with high-frequency impacts and achieve agile response; while in smooth transport tasks, it is set to 5 Hz to ensure stable movement. The regulator superimposes the compensation control signal onto the basic gait using a linear superposition operation. After superposition, the regulator performs frequency response shaping on the fused signal according to the set bandwidth parameter. For example, if the bandwidth is set to 15 Hz, a digital low-pass filter attenuates all frequency components higher than 15 Hz. After all data processing operations are completed, gait control commands are generated. This command is a data stream corresponding to the target torque output required by all joints of the robot dog in the next control cycle, which is directly sent to the underlying joint actuators to adjust the robot dog's actual gait.

[0089] Furthermore, the adaptive correction module continuously acquires exogenous interference vectors generated over multiple consecutive motion cycles and organizes them into a historical interference sequence. This process is achieved by setting a data acquisition window spanning multiple motion cycles (e.g., 8 gait cycles, totaling 4 seconds) and continuously recording exogenous interference vectors at a high sampling frequency (e.g., 200 Hz) of the endogenous interference observer. To preserve the temporal context of each data point, the module attaches a motion cycle index to each acquired interference vector (e.g., indicating which cycle it belongs to from 1 to 8) and stores it sequentially in a circular buffer with a capacity of 800 samples. When the buffer is full, the data set containing all interference information within the most recent 4 seconds is output as the overall historical interference sequence. Interference analysis is performed on historical interference sequences to extract persistent interference factors. This process applies a digital low-pass filter with an extremely low cutoff frequency (e.g., 0.5 Hz) to each six-dimensional exogenous interference vector in the sequence to filter out interference generated by instantaneous high-frequency events such as single-step slips, retaining only interference trends that persist or change slowly on a timescale of several seconds. Subsequently, the module uses the motion cycle index to perform phase-aligned periodic analysis on the filtered data. For example, when a robot dog begins to walk continuously on an uphill slope with a gradient of 10 degrees, the module identifies a constant backward reaction force of approximately -51 N (originating from the component of gravity along the slope) and a periodic pitch torque fluctuation synchronized with the gait (originating from periodic center-of-gravity adjustments to maintain balance) in each motion cycle. The module extracts this stable DC force component and the periodic torque fluctuation together to form a vector function that varies with the gait phase; this function is the persistent interference factor for this scenario.

[0090] The adaptive correction loop receives persistent disturbance factors and, based on inference rules (e.g., a constant backward force most likely originates from the component of gravity on a slope), maps them back to a correction vector for the model parameters. The loop calculates the equivalent slope angle of approximately 10 degrees based on a force of -51 Newtons and the robot's own mass of 30 kg, generating a correction vector representing "a 10-degree uphill slope in the current environment." This correction vector is then applied to update the parameters describing the external environment in the gait dynamics model, evolving it from a "flatland model" to a more accurate "10-degree slope model." Finally, this gait dynamics model, corrected through online learning, is immediately used to re-plan the basic movement gait.

[0091] Example 3:

[0092] This embodiment fully deploys the aforementioned gait adaptive control method and system for robotic dogs in complex terrain into an X-shaped quadruped robotic dog performing tasks in complex terrain. (Refer to...) Figure 3This enables adaptive control of the robot dog's gait.

[0093] The robot dog weighs 30 kg and is equipped with an inertial measurement unit, encoders for each joint, and a lidar sensor. The robot dog first needs to pick up a package from a flat concrete surface. The package's estimated weight is 5 kg, but its actual weight is 8 kg, leaving an unknown load of 3 kg. The robot dog then needs to carry the package up a 10-degree slope covered with slippery moss.

[0094] Based on the robot dog's known weight of 30 kg and an estimated load of 5 kg, the system establishes a gait dynamics model with a total mass of 35 kg. This model includes the trunk's inertia tensor, the geometric parameters of each link, and the degrees of freedom of each joint. Based on this model, and combined with the flat terrain information initially detected by lidar, the system plans a diagonal trotting gait as the basic movement gait. This gait sequence includes the target angles, angular velocities, and expected joint torques required to drive the movement for each joint in the next few seconds.

[0095] When the robot dog performs a task, its onboard inertial measurement unit collects the posture and angular velocity of its torso in real time at a frequency of 500 Hz, while the encoders of each joint report the actual rotation angle and angular velocity at a frequency of 1000 Hz. These data are fused to form the robot dog's actual gait. The endogenous disturbance observer uses a generalized momentum-based observer architecture, substituting the real-time acquired motion state into the established gait dynamics model to calculate the theoretically existing joint driving torque under this actual motion. This theoretical torque is then compared with the actual driving torque output by the motor controller. The difference between the two is equivalent to the sum of all internal and external disturbances not described by the model. When the robot dog is loaded with an 8 kg package, since the model still calculates based on a 5 kg load, the observer immediately detects an additional equivalent disturbance torque of approximately 29.4 N, equivalent to the gravitational force generated by approximately 3 kg of mass, acting on each supporting leg. The disturbance torque, after inverse transformation of the Jacobian matrix, is transformed into a six-dimensional exogenous disturbance vector acting in the fuselage center-of-mass coordinate system, clearly quantifying this instantaneous overload disturbance.

[0096] An exogenous disturbance vector containing overload information is input to the gait control loop. For the current "load walking" task, a task-mode-based configuration scheme is adopted, and the system calls a "medium stiffness" gain configuration from a preset parameter library. The gain configuration is a 6x6 diagonal matrix, where the gain coefficient corresponding to the vertical force is set to 80, and the gain coefficient corresponding to the pitch moment is set to 60. The gait control loop performs matrix multiplication on the exogenous disturbance vector and the matrix to generate a compensation control signal (containing an upward compensation force of approximately 29.4 Newtons) with opposite direction and proportional magnitude. The compensation control signal is fed forward to the joint moment command of the basic gait in a linear superposition manner through a bandwidth feedforward regulator with a bandwidth set to 20 Hz.

[0097] During the robot dog's movement, exogenous interference vectors were continuously collected and persistent interference factors were extracted. For a diagonal trotting state with a cycle of 0.5 seconds, a data acquisition window of 4 seconds was set, meaning that exogenous interference vectors generated over 8 movement cycles could be continuously collected, forming a historical interference sequence. When the robot dog walked from flat ground onto a 10-degree wet and slippery slope, the endogenous interference observer not only detected the downward force caused by overload but also the persistent resistance in the forward direction. Subsequently, the system applied a fourth-order Butterworth low-pass filter with a cutoff frequency of 0.5 Hz to each exogenous interference vector in the historical interference sequence, filtering out the instantaneous high-frequency interference caused by the robot dog's feet occasionally slipping on the wet and slippery moss, and retaining the persistent and periodically changing low-frequency trend term. After filtering, persistent interference factors were extracted: a downward force of approximately 29.4 Newtons; and a backward force of approximately 65 Newtons (i.e., the gravitational component of a total weight of 38 kg on a 10-degree slope).

[0098] The adaptive correction loop analyzes the received persistent disturbance factors according to the built-in inference rule set. For the downward force of 29.4 N, it can be analyzed as a persistent disturbance in the same direction as gravity and independent of gait, most likely originating from unmodeled load mass. Dividing this force value by gravitational acceleration reveals a 3 kg mass deviation, and the total mass parameter in the gait dynamics model is immediately corrected from 35 kg to 38 kg. For the backward force of 65 N, it can be analyzed as a persistent resistance in the forward direction, its magnitude matching the gravitational component of the current corrected total mass (38 kg) at a certain angle, most likely originating from terrain slope. Based on this force value and the corrected total weight, the equivalent slope of the current terrain is calculated to be approximately 10 degrees. Thus, the environmental model parameters are corrected from "flat ground" to "10-degree slope". After the dynamic model parameters (total mass) and environmental model parameters (slope) have been adaptively corrected online, the system re-plans the robot dog's basic gait based on this new model that better fits the actual terrain.

[0099] This embodiment improves the robot dog's lack of robustness when facing complex terrain by establishing a dual adaptive control architecture. The gait control loop enables the robot dog to respond quickly to sudden disturbances, ensuring stability during movement. Meanwhile, the adaptive correction loop based on historical disturbance analysis gives the robot dog online learning capabilities, effectively enhancing its long-term walking ability in complex terrain.

[0100] It should be clarified that the embodiments described above are merely exemplary and are intended to aid in understanding the present invention, not to limit it. Those skilled in the art can make various changes and modifications after grasping the core ideas of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for gait adaptation control of a complex terrain-oriented robot dog, characterized in that, The application relates to a method for adaptive disturbance compensation of a robot dog. The method comprises the following steps: constructing a gait dynamics model, planning a basic motion gait of the robot dog on a preset complex terrain; obtaining an actual motion gait of the robot dog, decoupling the basic motion gait and the actual motion gait through an endogenous disturbance observer, obtaining a motion gait residual generated by unexpected physical interaction, and quantifying the motion gait residual into an exogenous disturbance vector; the specific quantification process of the exogenous disturbance vector is as follows: the motion gait residual is transformed into a dynamic data stream containing external force and internal parameter deviation through a Jacobian matrix; the dynamic data stream is subjected to high-frequency measurement noise filtering processing and is mapped to a body coordinate system to obtain a six-dimensional vector containing three force components and three moment components; the six-dimensional vector generates an exogenous disturbance vector according to the current dynamics state of the robot dog; the exogenous disturbance vector is input into a gait control loop to obtain a compensation control signal; the compensation control signal and the basic motion gait are fused through a bandwidth feedforward regulator to generate a gait control instruction to adjust the actual motion gait; the exogenous disturbance vector generated in multiple motion cycles is continuously collected to form a historical disturbance sequence; the historical disturbance sequence is subjected to gait disturbance analysis to extract a persistent disturbance factor; the specific extraction process of the persistent disturbance factor is as follows: low-pass filtering is applied to each exogenous disturbance vector in the historical disturbance sequence to extract a low-frequency trend item that presents periodicity in amplitude and direction in multiple motion cycles and outputs the low-frequency trend item as the persistent disturbance factor; the low-frequency trend item is extracted by independently adopting a fourth-order Butterworth filter with a cutoff frequency lower than the gait frequency of the robot dog; The interference signals generated by single-step slipping and stepping on small stones are filtered out, and only the interference components that exist and change in the time scale are reserved; the persistent disturbance factor is input into an adaptive correction loop to adaptively correct the gait dynamics model and re-plan the basic motion gait. The specific implementation process of constructing a gait dynamics model and planning a basic motion gait of the robot dog on a preset complex terrain comprises the following steps: A gait dynamics model containing trunk inertia parameters and joint degrees of freedom is established; the geometric feature data of the preset complex terrain are combined to take the center of mass motion trajectory and the foot landing point as optimization targets; the gait dynamics model is solved by setting a periodic gait function and a stability constraint to generate a basic motion gait containing joint driving moments and motion sequences. The specific implementation process of obtaining an actual motion gait of the robot dog through an endogenous disturbance observer to obtain a motion gait residual comprises the following steps: trunk posture data and joint motion data are collected in real time through an inertial measurement unit and a joint encoder to obtain an actual motion gait; the actual motion gait and the basic motion gait are input into an endogenous disturbance observer; the endogenous disturbance observer estimates the deviation of the actual motion gait and the basic motion gait according to the gait dynamics model; the deviation caused by unexpected physical interaction is separated from the deviation estimation result to generate a motion gait residual.

2. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, ​ ​ 3. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, ​ 4. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, The specific implementation process of inputting the exogenous interference vector into the gait control loop to obtain a compensation control signal comprises that: the gait control loop is built-in with a preset dynamic gain matrix, the dynamic gain matrix is configured according to the expected compliance of the robot dog to external interference; the gait control loop performs matrix multiplication operation on the exogenous interference vector and the dynamic gain matrix, generates compensation force and compensation torque which are proportional to the size of the exogenous interference vector and opposite in direction, and combines the compensation force and the compensation torque into a compensation control signal.

5. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, The specific implementation process of fusing the compensation control signal and the basic motion gait through the bandwidth feedforward regulator to generate a gait control instruction to adjust the actual motion gait comprises that: the bandwidth feedforward regulator performs linear superposition operation on the compensation force and the compensation torque contained in the compensation control signal according to the configured bandwidth parameter, and performs frequency response shaping to generate a gait control instruction by feedforward superposition on the basic motion gait; the gait control instruction controls the joint driving of the robot dog in real time, thereby adjusting the actual motion gait.

6. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, The specific implementation process of continuously collecting the exogenous interference vectors generated in multiple motion cycles to form a historical interference sequence comprises: A data collection window spanning multiple motion cycles is set, and the exogenous interference vectors generated in the current motion cycle are continuously collected at the sampling frequency of the endogenous interference observer; each collected exogenous interference vector is attached with a motion cycle index to mark the time position in the motion process, and is stored in a circular buffer memory in sequence; when the exogenous interference vectors in the circular buffer memory reach a preset threshold, all stored exogenous interference vectors are output as a whole to generate a historical interference sequence.

7. The gait adaptive control method for a complex terrain-oriented dog robot according to claim 1, wherein, The specific implementation process of inputting the persistent interference factor into the adaptive correction loop to adaptively correct the gait dynamics model and re-plan the basic motion gait comprises: the adaptive correction loop receives the persistent interference factor, maps it into a correction vector of the model parameters, and updates the gait dynamics model using the correction vector; the updated gait dynamics model re-plans the basic motion gait after adaptive correction in combination with the geometric feature data of the preset complex terrain.

8. A complex terrain-oriented dog robot gait adaptive control system, characterized by, It comprises: An interference quantization module constructs a gait dynamics model and plans a basic motion gait of the robot dog on a preset complex terrain; An actual motion gait of the robot dog is obtained, the basic motion gait and the actual motion gait are decoupled by an endogenous interference observer, a motion gait residual generated by unexpected physical interaction is obtained, and the motion gait residual is quantized into an exogenous interference vector; A gait control module inputs the exogenous interference vector into a gait control loop to obtain a compensation control signal; a bandwidth feedforward regulator is used to fuse the compensation control signal and the basic motion gait to generate a gait control instruction to adjust the actual motion gait; The adaptive correction module continuously collects the exogenous disturbance vectors generated in multiple movement cycles to form a historical disturbance sequence; performs gait disturbance analysis on the historical disturbance sequence to extract a persistent disturbance factor; inputs the persistent disturbance factor into an adaptive correction loop to perform adaptive correction on the gait dynamics model and re-plan a basic movement gait.

Images