Robot control method, device and equipment and storage medium

Abstract

The invention provides a robot control method and device, equipment and a storage medium. The method comprises the steps that based on the current actual motion state of the robot and foothold information in a control instruction, the expected motion state of the robot is obtained through prediction of an inverted pendulum model; on the basis of the actual speed in the actual motion state and the expected speed in the expected motion state, the speed variation of the robot is obtained through prediction of a speed response model; based on the actual motion state and the expected motion state, predicting a residual speed through a neural network model; determining the execution speed of the robot based on the actual speed, the speed variation and the residual speed; and integral processing is carried out on the execution speed to obtain a motion trail of the robot, and the robot is controlled to move according to the motion trail. According to the invention, the accuracy and stability of robot control can be improved.

Description

Technical Field

[0001] This application relates to the field of robotics technology, and in particular to a robot control method, apparatus, device, and storage medium. Background Technology

[0002] In the field of robot motion control, accurately generating and tracking motion trajectories is crucial for achieving stable walking. Related technologies typically rely on simplified physical models based on kinematics / dynamics to plan the desired motion state. However, this method, based on idealized physical models, often ignores the hysteresis and limitations introduced by the dynamic response of the actual physical system during velocity changes. Furthermore, it is difficult to accurately model nonlinear factors such as ground friction and joint flexibility, leading to significantly increased prediction errors in complex terrain and other conditions. Consequently, robot motion control suffers from large errors and poor stability. Summary of the Invention

[0003] This application provides a robot control method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can improve the accuracy and stability of robot control.

[0004] The technical solution of this application embodiment is implemented as follows: This application provides a robot control method, including: Based on the robot's current actual motion state and the landing point information in the control commands, the robot's desired motion state is predicted using an inverted pendulum model. Based on the actual velocity in the actual motion state and the expected velocity in the expected motion state, the velocity change of the robot is predicted by the velocity response model. Based on the actual motion state and the desired motion state, the residual velocity is predicted by a neural network model. The robot's execution speed is determined based on the actual speed, the speed change, and the residual speed. The execution speed is integrated to obtain the robot's motion trajectory, and the robot is controlled to move according to the motion trajectory.

[0005] This application also provides a robot control device, including: The first prediction module is used to predict the desired motion state of the robot based on the robot's current actual motion state and the landing point information in the control commands, using an inverted pendulum model. The second prediction module is used to predict the speed change of the robot based on the actual speed in the actual motion state and the expected speed in the expected motion state through a speed response model. The third prediction module is used to predict the residual velocity based on the actual motion state and the desired motion state using a neural network model. The determination module is used to determine the execution speed of the robot based on the actual speed, the speed change, and the residual speed. The control module is used to integrate the execution speed to obtain the robot's motion trajectory, and control the robot to move according to the motion trajectory.

[0006] In the above scheme, the first prediction module is further configured to obtain the time length of the control cycle corresponding to the control command; the first prediction module is further configured to determine the model parameters of the inverted pendulum model; obtain the first model expression of the inverted pendulum model, the first model expression indicating the relationship between the actual motion state, the landing point information, the model parameters, the time length and the desired motion state; based on the actual motion state, the landing point information, the model parameters and the time length, perform prediction processing through the first model expression to obtain the desired motion state.

[0007] In the above scheme, the second prediction module is further used to obtain speed change parameters; obtain a second model expression of the speed response model, the second model expression indicating the relationship between the actual speed, the expected speed, the time length, the speed change parameters and the speed change amount; and perform prediction processing based on the actual speed, the expected speed, the time length and the speed change parameters through the second model expression to obtain the speed change amount.

[0008] In the above scheme, the control module is further configured to perform trapezoidal integral processing based on the actual speed, the execution speed, and the time length, using the trapezoidal integral formula to obtain the displacement increment of the control cycle; add the displacement increment to the actual position in the actual motion state to obtain the target position, and generate the motion trajectory from the actual position to the target position.

[0009] In the above scheme, the third prediction module is further used to obtain the terrain features of the terrain where the robot is currently located, and extract the first state features of the actual motion state and the second state features of the expected motion state; construct target features including the first state features and the terrain features; input the target features and the second state features into the multilayer perceptron included in the neural network model, and perform prediction processing on the target features and the second state features through the multilayer perceptron to obtain the residual velocity.

[0010] In the above scheme, the determining module is further used to add the actual speed, the speed change and the residual speed to obtain the summation result; and use the summation result as the execution speed of the robot.

[0011] In the above scheme, the actual motion state is the motion state of the robot at time t, and the execution speed is the predicted speed of the robot at time t+1. The first prediction module is further configured to, after controlling the robot to move according to the motion trajectory, obtain the robot's true speed at time t+1; determine the speed difference between the true speed and the execution speed; update the first model parameters of the inverted pendulum model based on the speed difference to obtain the target inverted pendulum model, and update the second model parameters of the speed response model based on the speed difference to obtain the target speed response model.

[0012] This application also provides an electronic device, including: Memory is used to store executable instructions for a computer; The processor, when executing computer-executable instructions stored in the memory, implements the robot control method provided in the embodiments of this application.

[0013] This application also provides a computer-readable storage medium storing computer-executable instructions or computer programs, which, when executed by a processor, implement the robot control method provided in this application.

[0014] This application also provides a computer program product, including computer-executable instructions or a computer program, which, when executed by a processor, implements the robot control method provided in this application.

[0015] The embodiments of this application have the following beneficial effects: predicting the desired motion state based on the landing point information using the inverted pendulum model ensures that the motion planning conforms to physical laws; calculating the velocity change based on the actual velocity and the desired velocity using the velocity response model reflects dynamic response characteristics; predicting the residual velocity using the neural network model compensates for deviations caused by nonlinear factors that may exist in the physical model (i.e., the inverted pendulum model and the velocity response model); and finally determining the execution speed by combining the actual velocity, velocity change, and residual velocity, integrating physical theory planning, dynamic response, and residual correction, significantly improving the prediction accuracy of the execution speed. As a result, the motion trajectory obtained based on the integral of this execution speed is more consistent with the real situation, thereby improving the accuracy and stability of controlling the robot to move according to the motion trajectory. Attached Figure Description

[0016] Figure 1This is a schematic diagram of the architecture of the robot control system provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application; Figure 3 This is a first flowchart illustrating the robot control method provided in an embodiment of this application; Figure 4 This is a second flowchart illustrating the robot control method provided in the embodiments of this application; Figure 5 This is a third flowchart illustrating the robot control method provided in the embodiments of this application; Figure 6 This is a schematic diagram of the robot control architecture provided in an embodiment of this application; Figure 7 This is a schematic diagram of the fourth process of the robot control method provided in the embodiments of this application; Figure 8 This is a schematic diagram of the fifth process of the robot control method provided in the embodiments of this application.

[0017] It should be noted that the terms "first" and "second" mentioned above are only used to distinguish between different options and do not represent the degree of superiority or inferiority of the options or their priority in the implementation process. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0020] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0021] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of a larger module or unit that includes the functionality of the module or unit.

[0022] Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application.

[0023] In the implementation of this application, the collection and processing of relevant data should strictly comply with the requirements of relevant laws and regulations, obtain the informed consent or separate consent of the personal information subject, and carry out subsequent data use and processing within the scope of laws and regulations and the authorization of the personal information subject.

[0024] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.

[0025] 1) The inverted pendulum model refers to a simplified dynamic model based on physical mechanisms, used to describe the balance and motion of a robot under the assumption of an ideal rigid body. This model abstracts the robot's mass distribution as a geometric relationship between the center of mass and the support point, and performs mathematical deductions based on the principles of gravity and inertia. In this application, its main function is not to directly output the final control commands, but rather to serve as the basis for prior physical knowledge, calculating the theoretically desired motion state that conforms to physical constraints based on the robot's current actual motion state and landing point information.

[0026] 2) Residual velocity refers to the nonlinear velocity compensation component output by the neural network model trained on a large amount of data. It characterizes the systematic deviations of the robot in complex environments caused by nonlinear factors such as ground friction, contact slippage, or model parameter uncertainties, which cannot be captured by the physical model. In this application, this physical quantity is used to correct the results calculated based on the physical model, thereby improving the fidelity of the execution velocity prediction by filling the error between theoretical prediction and real-world interaction.

[0027] 3) Execution speed refers to the synthetic velocity vector used to predict subsequent positions, determined by integrating the robot's current and evolving state, physical dynamic response characteristics, and nonlinear error correction. It is the algebraic sum of the actual velocity, velocity change, and residual velocity, representing the optimal estimate of the robot's true motion trend at the next moment or the end of the current cycle. This execution speed directly determines the accuracy of the motion trajectory obtained through subsequent integration processing and is a core intermediate variable connecting the velocity prediction layer and the position control layer.

[0028] 4) The velocity response model is a mathematical model that describes the dynamic process of a robot transitioning from its current actual speed to its desired speed. This model characterizes the physical inertia and response hysteresis of the robot's actuators, and is used to quantify its actual ability to follow speed commands within a finite control cycle. In this application, its main function is to calculate the velocity change that conforms to the laws of physical response based on the difference between the actual speed and the desired speed, thereby avoiding prediction distortion caused by sudden changes in commands.

[0029] This application provides a robot control method, apparatus, electronic device, computer-readable storage medium, and computer program product, which can improve the accuracy and stability of robot control. The embodiments of this application will now be described in detail based on the above description of the terms and concepts used in the embodiments.

[0030] The robot control system provided in the embodiments of this application is described below. See also... Figure 1 , Figure 1 This is a schematic diagram of the architecture of the robot control system provided in an embodiment of this application. To support an exemplary application, the robot control system 100 includes: a server 200, a network 300, and a robot 400. The robot 400 is connected to the server 200 via the network 300, which can be a wide area network (WAN), a local area network (LAN), or a combination of both, using wireless or wired links for data transmission.

[0031] Here, server 200 sends control commands to robot 400; robot 400 receives the control commands; in response to the control commands, it acquires the current actual motion state of robot 400 and extracts the landing point information from the control commands; based on the actual motion state and landing point information, it predicts the desired motion state of robot 400 using an inverted pendulum model; based on the actual speed in the actual motion state and the desired speed in the desired motion state, it predicts the speed change of robot 400 using a speed response model; based on the actual motion state and the desired motion state, it predicts the residual speed using a neural network model; based on the actual speed, speed change, and residual speed, it determines the execution speed of robot 400; it integrates the execution speed to obtain the motion trajectory of robot 400, and moves according to the motion trajectory.

[0032] The robot control method provided in this application is implemented by an electronic device. For example, it can be implemented by a robot alone, by a server alone, or by a robot and a server working together. The electronic device implementing the robot control method provided in this application can be various types of robots or servers. The server (e.g., server 200) can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The robot (e.g., robot 400) can be various types of robots, such as industrial robots, service robots, medical robots, etc. The robot and the server can be connected directly or indirectly through wired or wireless communication, and this application does not impose any limitations on this.

[0033] The following describes an electronic device for implementing a robot control method provided in an embodiment of this application. See also... Figure 2 , Figure 2 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 500 provided in this embodiment can be a robot or a server. Figure 2 As shown, electronic device 500 includes at least one processor 510, memory 550, at least one network interface 520, and user interface 530. The various components in electronic device 500 are coupled together via a bus system 540. It is understood that the bus system 540 is used to implement communication between these components. In addition to a data bus, the bus system 540 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in… Figure 2 The general labeled all buses as Bus System 540.

[0034] The processor 510 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0035] User interface 530 includes one or more output devices 531 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 530 also includes one or more input devices 532, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.

[0036] Memory 550 may be removable, non-removable, or a combination thereof. Memory 550 may include one or more storage devices physically located away from processor 510. Memory 550 may include volatile memory or non-volatile memory, or both. Non-volatile memory may be read-only memory (ROM), and volatile memory may be random access memory (RAM). The memory 550 described in this application embodiment is intended to include any suitable type of memory.

[0037] In some embodiments, memory 550 is capable of storing data to support various operations. Examples of such data include programs, modules, and data structures, or subsets or supersets thereof, as illustrated below. Operating system 551 includes system programs for handling various basic system services and performing hardware-related tasks, such as a framework layer, core library layer, driver layer, etc., for implementing various basic services and handling hardware-based tasks; network communication module 552 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 520, exemplary network interfaces 520 including Bluetooth, Wireless Fidelity (Wi-Fi), and Universal Serial Bus (USB); presentation module 553 is used to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 531 associated with user interface 530 (e.g., a display screen, a speaker, etc.); input processing module 554 is used to detect and translate one or more user inputs or interactions from one or more input devices 532.

[0038] In some embodiments, the robot control device provided in this application can be implemented in software. Figure 2A robot control device 555 stored in memory 550 is shown. It can be software in the form of programs and plug-ins, including the following software modules: a first prediction module 5551, a second prediction module 5552, a third prediction module 5553, a determination module 5554, and a control module 5555. These modules are logically related and can therefore be arbitrarily combined or further divided according to the functions they implement. The functions of each module will be described below.

[0039] The robot control method provided in the embodiments of this application is described below. As mentioned above, the robot control method provided in the embodiments of this application is implemented by an electronic device, such as a server or a robot alone, or a server and a robot working together. Therefore, the executing entity of each step will not be described again below. See Figure 3 , Figure 3 This is a first flowchart illustrating the robot control method provided in this application embodiment. The robot control method provided in this application embodiment includes: Step 101: Based on the robot's current actual motion state and the landing point information in the control commands, predict the robot's desired motion state using the inverted pendulum model.

[0040] For step 101, the inverted pendulum model refers to a physical model that simplifies the overall dynamic characteristics of a bipedal robot into a three-dimensional linear inverted pendulum. This model describes the evolution of the robot's center of gravity relative to its landing point under gravity. The actual motion state refers to the physical state quantity obtained by the robot at the start of the current control cycle through sensor data acquisition and state estimation. The landing point information in the control command refers to the coordinates of the expected foot landing position issued by the gait planning layer. The expected motion state refers to the theoretically desirable center of gravity state that the robot should achieve in the next moment, calculated based on the physical dynamics equations, and this state conforms to rigid body equilibrium constraints. Specifically, firstly, the robot's current center of gravity position, velocity, and orientation angle are obtained through onboard sensors and a state estimator. These parameters constitute the current actual motion state. Simultaneously, control commands within the current control cycle are acquired, and the desired landing point information (i.e., landing point position and angle) is extracted from the control commands. Next, the state transition matrix and control input matrix of the three-dimensional inverted pendulum model are determined. These two matrices are determined by the control cycle and the model's adaptive adjustment coefficients. Then, a state-space equation describing the evolution of the center of gravity state over time is constructed. The current actual motion state is used as the initial state vector, and the landing point information is used as the control input vector. These are substituted into the state-space equation for forward kinematics calculation. Finally, the calculated center of gravity position, velocity, and orientation angle for the next moment are output as the robot's desired motion state.

[0041] In some embodiments, the following steps may also be performed: obtaining the time length of the control cycle corresponding to the control command; based on this, see [reference needed]. Figure 4 Step 101, "Based on the robot's current actual motion state and the landing point information in the control commands, predict the robot's desired motion state using the inverted pendulum model," can be achieved by executing the following steps 1011-1013: Step 1011, determine the model parameters of the inverted pendulum model; Step 1012, obtain the first model expression of the inverted pendulum model, which indicates the relationship between the actual motion state, landing point information, model parameters, time length, and desired motion state; Step 1013, based on the actual motion state, landing point information, model parameters, and time length, perform prediction processing using the first model expression to obtain the desired motion state.

[0042] The control cycle length refers to the time interval between two adjacent state updates or command issuances in the robot control system, usually denoted as the time step. Model parameters are the physical constants and adjustment coefficients that determine the dynamic characteristics of the inverted pendulum, including but not limited to the center of mass height, gravitational acceleration, and adaptive weights used to calibrate model errors. The first model expression is the mathematical equation describing the evolution of the inverted pendulum's state in the discrete time domain, usually presented in the form of a state-space equation.

[0043] For step 1011, the height of the center of mass is obtained based on the robot's physical configuration. and gravitational acceleration Calculate the natural frequency of the inverted pendulum; simultaneously, read or calculate the adaptive adjustment coefficient in real time by the parameter tuning network. and These two coefficients are used to fine-tune the dynamic characteristics of the standard inverted pendulum model to match the actual hardware response of the current robot.

[0044] For step 1012, the discrete state-space equation (i.e., the first model expression) of the three-dimensional linear inverted pendulum is constructed. This equation defines the linear mapping relationship between the desired motion state at the next moment and the current actual motion state and control input. The state transition matrix and control input matrix in the equation are constructed by the model parameters and the time length of the control period, and the matrix elements contain hyperbolic function terms or their approximate expansion terms involving the time length.

[0045] Specifically, the first model expression is shown in formula (1) below: Formula (1) in, It is the first The state of motion at any given moment (i.e., the actual state of motion, including the actual position) , ), actual speed ( , ) and azimuth angle ); It is the first The motion state at any given moment (i.e., the desired motion state, which includes the same information as the actual motion state, namely the desired position, desired velocity, and azimuth angle). ( ) is the control input (i.e., landing point information, including the landing point) and landing angle ); and This is an adaptive adjustment coefficient.

[0046] in, Here is the state transition matrix. To control the input matrix.

[0047] in, This can be expressed by the following formula (2): Formula (2) in, This can be expressed by the following formula (3): Formula (3) in, To control the duration of the cycle.

[0048] in, This can be expressed by the following formula (4): Formula (4) in, It is the acceleration due to gravity. This is the height of the robot's center of mass (or center of gravity).

[0049] For step 1013, the center of gravity position, center of gravity velocity (corresponding to the actual velocity), and azimuth angle collected at the current moment are assembled into a state vector (i.e., a vector of the actual motion state). The landing point coordinates and landing angle in the control command are assembled into a control input vector (i.e., a vector of landing point information). Subsequently, the state vector and control input vector are substituted into the first model expression, and matrix multiplication and addition are performed using the determined state transition matrix and control input matrix to calculate the center of gravity state values ​​(i.e., center of gravity position, center of gravity velocity (corresponding to the desired velocity), and azimuth angle) corresponding to the next time step, which is the desired motion state.

[0050] Applying the above embodiments, an inverted pendulum model based on physical mechanisms is used as the prediction core. A deterministic relationship between motion state, time duration, and model parameters is established using explicit state-space equations. On the one hand, by introducing specific control cycles and physical parameters (such as center of mass height and adaptive coefficients) to construct a state transition matrix, it is ensured that the generated desired motion state strictly conforms to rigid body dynamics constraints, thereby guaranteeing the robot's theoretical equilibrium stability. On the other hand, this parameterized modeling method can adapt to robots with different heights and mass distributions by adjusting the model parameters, without retraining the network. This significantly improves the algorithm's versatility and transfer efficiency across different robot models, while also reducing computational complexity and meeting the requirements of high-frequency real-time control.

[0051] Step 102: Based on the actual velocity in the actual motion state and the expected velocity in the expected motion state, the velocity change of the robot is predicted by the velocity response model.

[0052] For step 102, the velocity response model refers to a dynamic model used to describe the dynamic lag characteristics of the robot chassis and joint actuators after receiving a velocity command, due to physical factors such as mechanical inertia, motor response time, and system damping. Mathematically, this velocity response model is usually represented as a first-order inertial element or a first-order lag system (such as a first-order oscillation model), used to characterize the asymptotic process of the actual velocity following the commanded velocity. Actual velocity refers to the physical velocity feedback value of the robot at the current moment. Expected velocity refers to the ideal velocity theoretically required to maintain balance, calculated from the preceding inverted pendulum model, and serves as the tracking target or command input (i.e., commanded velocity) for the velocity response model in this step. Velocity change refers to the actual velocity increment that the robot can produce within a single control cycle, constrained by the time constant. Specifically, firstly, the desired speed output by the inverted pendulum model is assigned as the command speed of the current control cycle; simultaneously, the robot's actual speed and the time constant (i.e., speed change parameter) characterizing the dynamic response of the speed response model are obtained. This time constant is set according to the robot's hardware characteristics or adaptively adjusted online by the neural network; next, the difference between the command speed and the actual speed is calculated to obtain the speed tracking error; subsequently, the speed tracking error is solved using the discretized first-order inertial system state equation. This solution process specifically uses an exponential decay function containing the ratio of the control cycle to the time constant, or a linear approximation formula based on Taylor expansion, to calculate the speed increase that can be responded to within the current control cycle; finally, this speed increase is output as the robot's speed change, thereby transforming the ideal kinematic constraints into a dynamic execution increment that conforms to the actual electromechanical characteristics.

[0053] In some embodiments, step 102, "predicting the robot's velocity change based on the actual velocity in the actual motion state and the expected velocity in the expected motion state using a velocity response model," can be achieved by performing the following steps: obtaining velocity change parameters; obtaining a second model expression for the velocity response model, the second model expression indicating the relationship between the actual velocity, expected velocity, time length, velocity change parameters, and velocity change; and performing prediction processing based on the actual velocity, expected velocity, time length, and velocity change parameters using the second model expression to obtain the velocity change.

[0054] Among them, the velocity variation parameter refers to the time constant that characterizes the lag in the robot's dynamic system's response to input commands. This speed change parameter reflects the comprehensive physical properties of the motor response characteristics, mechanical transmission inertia, and system damping (that is, it describes the speed adjustment process; the two are inversely proportional). The second model expression refers to the analytical equation describing the dynamic response law of a first-order inertial system in the discrete time domain, used to accurately calculate the increment of the velocity physical quantity from the current value to the target value within a given control period.

[0055] First, the time constant value stored in the preset memory or output in real time by the online parameter estimation module is read and determined as the velocity change parameter. Second, a preset second model expression is constructed or called. This second model expression is constructed based on the zero-order preserved discretization of the first-order linear differential equation, and its mathematical structure describes the velocity error elimination process through an exponential decay function. Specifically, the second model expression defines the velocity change as equal to the difference between the expected velocity and the actual velocity, multiplied by an exponential factor about the ratio of time length to velocity change parameter. The second model expression is expressed by the following formula (5): Formula (5) in, For the change in velocity, For actual speed, For the desired speed, To control the duration of the cycle (i.e., the above) ), For velocity variation parameters.

[0056] Finally, the actual speed is taken as the current state value, and the desired speed is taken as the target command value. Combined with the current control cycle length and the determined speed change parameters, these values ​​are substituted into the second model expression for numerical calculation. During the calculation, the difference between the desired speed and the actual speed is first calculated, and then the exponential decay coefficient is calculated. Multiplying the two together gives the increase in speed that the robot can actually execute within the current control cycle, which is the speed change.

[0057] By applying the above embodiments and introducing a second model expression that includes a time constant, the physical inertia and hysteresis characteristics of the robot actuator are explicitly incorporated into the prediction process. Compared to rigid predictions or simple linear approximations relying solely on geometric kinematics, this scheme employs an analytical solution based on an exponential function, which can accurately describe the dynamic approximation process of velocity in the time dimension, effectively eliminating the truncation error caused by discretization calculations. This approach makes the predicted velocity change more consistent with the actual response capabilities of the motor and mechanical structure, avoiding prediction overshoot or oscillations caused by sudden command changes, significantly improving the fit between the motion trajectory derived from the velocity integral and the actual trajectory, and enhancing control robustness in high-speed motion scenarios.

[0058] Step 103: Based on the actual motion state and the desired motion state, the residual velocity is predicted by a neural network model.

[0059] For step 103, the neural network model refers to a data-driven nonlinear estimation architecture, typically employing a lightweight multilayer perceptron structure. It is used to compensate for nonlinear dynamic characteristics that are difficult to characterize by the physical foundation models (i.e., the aforementioned inverted pendulum model and velocity response model). This neural network model possesses offline training and online fine-tuning capabilities. Residual velocity refers to the velocity deviation component that exceeds the prediction range of the inverted pendulum model and velocity response model, caused by complex physical factors such as ground friction, foot slippage, terrain undulations, or nonlinearity of the mechanical structure. Specifically, firstly, a multidimensional state feature vector is extracted and constructed. This feature vector encompasses the current center of gravity position and current actual velocity in the actual motion state, the expected center of gravity position and expected velocity in the expected motion state calculated by the inverted pendulum model, and the terrain feature parameters of the robot's current environment. Then, the constructed feature vector is input into the multilayer perceptron network for forward propagation calculation. Nonlinear transformations are performed in the hidden layers using the network's internal weights and bias parameters to map the original deviation prediction value. Finally, convergence accuracy is introduced. As an output constraint, the amplitude range of the original deviation prediction value is truncated or limited (i.e., the predicted value is...). This ensures that the output value is within a preset stable range, thereby suppressing prediction divergence caused by input data noise or network overfitting; finally, the output value after constraint processing is used as the residual velocity, which reflects the subtle difference between the physical model prediction value and the actual physical law under the current environment.

[0060] In some embodiments, step 103, "predicting residual velocity based on actual motion state and desired motion state using a neural network model," can be achieved by performing the following steps: acquiring the terrain features of the terrain where the robot is currently located, and extracting the first state features of the actual motion state and the second state features of the desired motion state; constructing target features including the first state features and terrain features; inputting the target features and the second state features into the multilayer perceptron included in the neural network model, and performing prediction processing on the target features and the second state features through the multilayer perceptron to obtain the residual velocity.

[0061] Among them, terrain feature `terrain_type` refers to a set of parameters characterizing the physical properties and geometric shape of the robot's current foot contact surface, including but not limited to ground slope, surface roughness, friction coefficient, or local elevation information. First state feature refers to the actual motion state (including actual speed) collected from the sensors at the current moment. and actual location The key state components selected from the inverted pendulum model include the actual center of gravity position vector and the actual velocity vector. The second state feature refers to the reference components selected from the theoretical state (i.e., the desired motion state) calculated from the inverted pendulum model for the next moment, such as the desired center of gravity position vector and the desired velocity vector. The target feature refers to the combined feature vector constructed to comprehensively describe the constraints of the current physical environment. A multilayer perceptron is the core computational unit of a neural network model, a fully connected network consisting of an input layer, several hidden layers, and an output layer.

[0062] Specifically, firstly, the robot acquires environmental information using visual sensors, LiDAR, or propagation algorithms, quantifying terrain features. Simultaneously, the actual motion state undergoes dimensionality filtering and normalization to extract first-state features. The same preprocessing is applied to the desired motion state output by the inverted pendulum model to extract second-state features. Secondly, using vector concatenation or feature fusion, the first-state features are combined with the terrain features to construct target features encompassing the robot's current state and external environmental constraints. These target features characterize the physical background that generates nonlinear errors. Finally, the constructed target features and the second-state features, serving as physical priors, are input into a multilayer perceptron (a simple, fast, and real-time-capable component) within the neural network model. The multilayer perceptron uses pre-trained weight matrices and bias parameters to perform forward propagation calculations on the input feature vectors. After mapping through the nonlinear activation functions of the hidden layers, it fits the system dynamics deviations not covered by the physical model, ultimately outputting the predicted deviation value as the residual velocity. .

[0063] Applying the above embodiments, a data-driven strategy guided by physical priors is employed, using both the ideal physical evolution law (second state features) and the actual state including environmental constraints (target features) as network inputs. This design leverages the powerful nonlinear fitting capability of multilayer perceptrons to accurately compensate for prediction biases caused by neglecting ground friction, slippage, or flexible contact in complex terrain using the inverted pendulum model. Compared to purely data-driven methods (i.e., pure neural network models), this approach not only converges faster but also possesses clear physical meaning, significantly improving the trajectory prediction accuracy and motion stability of the robot in unstructured environments.

[0064] Step 104: Determine the robot's execution speed based on the actual speed, speed change, and residual speed.

[0065] For step 104, the execution speed refers to the synthetic velocity vector obtained after dual processing of physical model prediction and neural network residual correction, which characterizes the robot's actual motion state at the next moment or the end of the current control cycle. This physical quantity integrates the robot's rigid body dynamics, the frequency response characteristics of the actuator, and environmental nonlinear disturbance factors, serving as a high-precision velocity benchmark for subsequent trajectory integration and long-term position extrapolation. Specifically, firstly, the actual velocity at the current sampling moment is used as the starting benchmark value for velocity evolution; secondly, the velocity change calculated by the velocity response model is obtained, which represents the robot's physical following increment under time constant constraints; simultaneously, the residual velocity output by the neural network model is obtained, which represents the nonlinear compensation component used to correct the prediction deviation of the physical model; subsequently, the velocity change and residual velocity are added to the actual velocity using a vector linear superposition method, and the algebraic sum of the three is calculated; finally, the synthetic vector obtained is determined as the robot's execution speed, thus completing a single-step extrapolation from the current actual state to the future predicted state. This execution speed will be used as the iterative input for the next prediction cycle or directly used for displacement integral calculation.

[0066] In some embodiments, step 104, "determining the robot's execution speed based on the actual speed, the speed change, and the residual speed," can be achieved by performing the following steps: adding the actual speed, the speed change, and the residual speed together to obtain the sum; and using the sum as the robot's execution speed.

[0067] The summation result refers to the composite vector obtained by performing linear algebraic operations on velocity components from different sources within the multidimensional velocity vector space. This summation result represents the final estimate after multi-source information fusion. The execution speed refers to the optimal speed estimate of the robot at the next moment, inferred by the system after integrating the current motion state, theoretical dynamic response, and data-driven nonlinear compensation. This speed will serve as the basic input for subsequent position integration or the next round of recursive prediction. Specifically, firstly, the actual velocity vector measured by the robot at the current moment is used as the starting reference for state evolution; secondly, the velocity change calculated by the velocity response model is obtained, which reflects the dynamic following capability of the robot's mechanical system to the desired command within the control cycle; simultaneously, the residual velocity output by the neural network model is obtained, which reflects the correction value for nonlinear velocity deviations caused by complex environmental factors such as ground friction and contact slippage; subsequently, a vector superposition operation is performed, adding the velocity change and residual velocity component by component to the actual velocity, that is, performing an algebraic summation of the actual velocity, velocity change, and residual velocity; finally, the calculated superposition result is directly determined as the robot's execution speed, i.e., the execution speed. =Actual speed + Velocity change +Residual speed .

[0068] Applying the above embodiments, an explicit additive fusion strategy is employed to decouple the complex robot motion prediction problem into two independent parts: linear dynamic response and nonlinear residual compensation. By directly superimposing the "main trend" (velocity change) calculated by the physical model with the "micro-correction" (residual velocity) predicted by the neural network, the stability and interpretability of the physical model in macroscopic motion trend prediction are preserved, preventing result divergence. Simultaneously, the fitting advantage of the neural network in microscopic nonlinear characteristics is fully utilized, effectively eliminating systematic errors caused by model simplification. This approach achieves complementary advantages between physical mechanisms (i.e., the basic physical model) and data-driven approaches (i.e., the neural network model) at extremely low computational cost, significantly improving the fidelity of velocity prediction and laying an accurate velocity benchmark for subsequent generation of high-precision long-term motion trajectories.

[0069] Step 105: Integrate the execution speed to obtain the robot's motion trajectory, and control the robot to move according to the motion trajectory.

[0070] For step 105, integration processing refers to the process of converting discrete time-series velocity signals into continuous spatial displacement data using a numerical integration algorithm. Its accuracy directly determines the reliability of long-term prediction in the spatial dimension. The motion trajectory refers to a set of spatiotemporal position coordinates of the robot's center of gravity within a future prediction time window. This trajectory reflects the robot's evolution path under the constraints of physical laws and the correction of neural networks. Specifically, firstly, the preceding physical model prediction and neural network residual correction steps are repeatedly executed through iterative loops to generate a series of discrete execution speed values ​​covering a preset future time window. Secondly, to suppress the cumulative position error that grows over time in long-term prediction, a high-precision trapezoidal integral algorithm is used instead of the conventional Euler integral method to calculate the arithmetic mean of the execution speeds at two adjacent discrete moments, and this average is multiplied by the time step to obtain the single-step displacement increment. Next, starting from the robot's actual center of gravity position at the current moment, the calculated single-step displacement increment is accumulated step by step to construct a motion trajectory containing the coordinates of the center of gravity positions at multiple future moments. Finally, the generated motion trajectory is sent as a reference benchmark to the robot's underlying motion controller, such as a model prediction controller or a whole-body controller. The underlying motion controller calculates joint torques or motor commands based on the motion trajectory to drive the robot to perform real-time tracking motion along the planned path.

[0071] In some embodiments, the step "integrating the execution speed to obtain the robot's motion trajectory" can be achieved by performing the following steps: based on the actual speed, execution speed, and time length, performing trapezoidal integration using the trapezoidal integral formula to obtain the displacement increment of the control cycle; adding the displacement increment to the actual position in the actual motion state to obtain the target position, and generating the motion trajectory from the actual position to the target position.

[0072] The trapezoidal integral formula is an approximate calculation method for numerical integration. It estimates the average velocity within an integration interval by calculating the arithmetic mean of the velocity values ​​at the first and last two moments, and then calculates the displacement, achieving second-order algebraic accuracy. The displacement increment refers to the change in spatial vector of the robot's center of gravity relative to its position at the previous moment within the time domain span of a single control cycle. The target position refers to the theoretical spatial coordinates that the robot should reach at the end of the current control cycle, calculated through integration.

[0073] Specifically, first, the robot's current actual velocity is obtained as the starting velocity of the integration interval, and the execution velocity calculated in the previous steps is obtained as the ending velocity of the integration interval. The duration of the control cycle is then determined. Second, based on the trapezoidal integral principle, the vector sum of the actual velocity and the execution velocity is calculated. This vector sum is divided by 2 to obtain the average velocity vector, which is then multiplied by the time duration to accurately calculate the displacement increment within the current control cycle. Next, the current actual position coordinates of the robot are read, and the calculated displacement increment is added to these coordinates to calculate the target position for the next moment. Finally, the calculated target position is stored in the prediction sequence, and a motion trajectory representing the robot's spatial movement path for the control cycle is generated, starting from the current actual position and ending at the target position. This trajectory is then used to guide the robot's posture adjustment and foot landing point planning. Repeating this motion trajectory generation method yields motion trajectories for multiple consecutive control cycles.

[0074] Applying the above embodiments, the trapezoidal integral formula is used instead of the conventional first-order Euler integral method for displacement estimation. By calculating the average velocity using the velocity information (actual velocity and executed velocity) at both ends of the control cycle, this method effectively improves the accuracy of numerical integration and significantly reduces the truncation error caused by discretization calculation. This second-order precision integration method can better suppress the cumulative position drift during long-term prediction, ensuring a high degree of consistency between the generated motion trajectory and the robot's actual physical motion tendency, thereby improving the smoothness of robot navigation control and the accuracy of terminal positioning.

[0075] In some embodiments, the actual motion state is the robot's motion state at time t, and the execution speed is the robot's predicted speed at time t+1; based on this, see [reference needed]. Figure 5 After executing "controlling the robot to move according to the motion trajectory", the following steps can also be executed: Step 201, obtain the robot's real speed at time t+1; Step 202, determine the speed difference between the real speed and the executed speed; Step 203, based on the speed difference, update the first model parameters of the inverted pendulum model to obtain the target inverted pendulum model, and update the second model parameters of the speed response model based on the speed difference to obtain the target speed response model.

[0076] Here, the true velocity refers to the physical velocity feedback value obtained in real time through the robot's onboard sensor combination (such as an inertial measurement unit and an encoder) at time t+1 on the physical time axis. The velocity difference is an error index used to quantify the model's predictive performance, characterizing the deviation between the execution speed output by the predictive system and the actual physical system response. The first model parameter refers to the core coefficients in the inverted pendulum model that determine the dynamic characteristics of the rigid body's equilibrium (as mentioned above). and This includes, but is not limited to, adaptive weights or physical constant correction terms in the state transition matrix. The second model parameter refers to the time constant in the velocity response model that characterizes the system's dynamic following properties. Or the hysteresis coefficient. The target inverted pendulum model and the target velocity response model refer to the updated models that have higher fitting accuracy and are more adapted to the current working conditions after online parameter calibration.

[0077] Specifically, firstly, the control clock is monitored. When time progresses to time t+1, a sensor data acquisition task is triggered to obtain the robot's current actual physical speed, which is defined as the true speed. Secondly, the execution speed calculated at time t is retrieved from memory. The true speed and the execution speed are then subtracted by a vector to calculate the deviation vector between them, obtaining the speed difference. Finally, this speed difference is used as a feedback signal input to the parameter adaptive adjustment module. Using gradient descent or Kalman filtering algorithms, the adaptive weights of the inverted pendulum model are adjusted according to the magnitude and direction of the difference (as described above). and Fine-tuning was performed to update the parameters of the first model to generate the target inverted pendulum model, while simultaneously correcting the time constant in the velocity response model. The second model parameters are updated to generate the target velocity response model, so that the updated model (including the target inverted pendulum model and the target velocity response model) can be used for prediction in the next control cycle.

[0078] Using the above embodiments, an online adaptive calibration closed loop based on posterior error was constructed. By comparing the predicted value (execution speed) at the previous moment with the observed value (actual speed) at the current moment in real time, it can accurately detect physical model mismatches caused by mechanical wear, load changes, or environmental changes. Based on the speed difference, the first and second model parameters are dynamically updated, achieving real-time approximation of the robot's actual physical characteristics by the inverted pendulum model and the speed response model. This "prediction-verification-correction" mechanism eliminates the cumulative errors generated by static model parameters during long-term operation, significantly enhancing the system's adaptability in unstructured environments and its long-term prediction robustness, ensuring that control accuracy continuously optimizes over time.

[0079] See Figure 6 This application provides a hybrid state model 60 and a high-precision integrator 63. The hybrid state model 60 includes a physical foundation model 61 (including an inverted pendulum model 611 and a velocity response model 612 (i.e., a first-order oscillation model)) and a neural network model 62. The inverted pendulum model 611 is used to calculate the desired motion state, the velocity response model 612 is used to calculate the velocity change, the neural network model 62 is used to calculate the residual velocity, and the high-precision integrator 63 is used for gradient integration.

[0080] See Figure 7 The embodiments of this application include: step 701, collecting sensor data.

[0081] Here, raw signals output from various sensors installed on the robot body and joints are acquired. This data includes, but is not limited to, acceleration and angular velocity information collected by the inertial measurement unit, joint angle information fed back by the encoder, and terrain feature data captured by environmental sensing devices. Step 701 aims to establish a digital foundation reflecting the robot's current physical working condition and external environment, providing a real-time input source for subsequent data processing and state calculation.

[0082] Step 702: The state estimation module determines the actual motion state based on sensor data.

[0083] Here, the collected sensor data undergoes filtering, noise reduction, and information fusion processing. By applying kinematics algorithms or state observers, the discrete raw signals are transformed into a system state vector (i.e., the actual motion state) describing the robot's physical properties at the current moment. The actual motion state determined in step 702 includes the robot's center of gravity position, real-time velocity, and attitude angle, thereby constructing an initial state reference for subsequent dynamic model prediction and trajectory planning.

[0084] Step 703: The hybrid state model outputs the robot's execution speed.

[0085] Here, a hybrid state model is used to comprehensively estimate the robot's future speed trend. This model combines physical mechanism prediction based on an inverted pendulum model and a speed response model, with nonlinear residual compensation based on a neural network model. By vector-superimposing the speed change calculated based on physical constraints, the residual speed output by the neural network, and the current actual speed, an execution speed that integrates dynamic response characteristics and error correction is calculated and output.

[0086] Step 704: The high-precision integrator outputs the robot's motion trajectory.

[0087] Here, the trapezoidal integral algorithm is used to integrate the output execution speed. Based on the preset control cycle length, the displacement increment is calculated using the trapezoidal integral formula, and this increment is accumulated to the current actual position. Step 704 generates a continuous path containing a series of spatiotemporal coordinate points by accurately simulating the position change, which is the robot's motion trajectory, ensuring the smoothness and physical executability of the trajectory.

[0088] Step 705: Send the motion trajectory to the MPC controller.

[0089] Here, the generated motion trajectory is transmitted to the MPC controller, which converts the motion trajectory into standardized low-level control instructions.

[0090] Step 706: The actuator moves based on the motion trajectory of the MPC controller.

[0091] Here, the system receives low-level control commands calculated by the MPC controller, driving the robot's joint motors or hydraulic components to move. The actuators precisely adjust the output torque or angle of each joint according to the commands, controlling the robot's foot landing point and center of gravity displacement, so that it strictly follows the issued motion trajectory for spatial movement, thereby enabling the robot to walk stably or operate dynamically in complex unstructured environments.

[0092] This achieves improved accuracy (position error less than 5 cm, for example, a 62% improvement in pure physical models); improved real-time performance, reducing prediction time per time step; reduced redundant motion trajectories, lowering robot navigation energy consumption; improved adaptability to different environments and different motion control models; and enhanced robustness to varying machine system consistency during mass production.

[0093] See Figure 8 The embodiments of this application also include: step 801, data acquisition.

[0094] Here, historical operational data of the robot in various terrain scenarios is collected as the sample source. The actual motion state and corresponding terrain features at a certain moment are selected from the sensor records as input data. At the same time, based on the actual physical velocity measured at the next moment, the velocity components that cannot be explained by the physical basis model are calculated, labeled as the true residual velocity, and thus a sample dataset containing input data and labels is constructed.

[0095] Step 802, data preprocessing.

[0096] Here, the sample data in the sample dataset undergoes cleaning and standardization operations. For example, outlier noise data caused by sensor malfunctions or communication packet loss is removed, and features of different physical dimensions are normalized. This eliminates the impact of dimensional differences on numerical calculations, standardizes data distribution, and thus improves the convergence speed and numerical stability during subsequent model training, ensuring the effectiveness of feature extraction.

[0097] Step 803, Physical basis model prediction.

[0098] Here, the preprocessed actual motion state is input into a pre-defined physical model. The desired motion state is planned using an inverted pendulum model, and the velocity change based on dynamic hysteresis is calculated using a velocity response model.

[0099] Step 804, residual velocity calculation.

[0100] Here, a target feature vector containing the actual motion state and terrain features is constructed and input into the neural network model to be trained. Forward propagation is performed using a multilayer perceptron within the model to map the current prediction residual velocity. In this way, the model's current estimated value of the nonlinear system bias can be obtained during training iterations, which can then be compared with labeled data to calculate the loss.

[0101] Step 805, neural network model training.

[0102] Here, the numerical difference between the predicted residual velocity and the actual residual velocity (used as a label) is calculated to construct a loss function. Gradient descent and backpropagation algorithms are used to iteratively adjust the weight matrix and bias parameters in the neural network model based on this difference. By minimizing the prediction error, the model gradually fits the nonlinear influence of environmental factors and complex dynamics on the robot's velocity.

[0103] Step 806, Model Validation.

[0104] Here, the trained neural network model is tested on an independent validation dataset. The model's residual velocity prediction accuracy and generalization performance on unseen data are evaluated, and overfitting is checked. If all metrics meet the preset performance thresholds, the model training is considered complete; otherwise, the network structure or hyperparameters are adjusted, and the training process is re-executed to ensure the model's reliability under real-world conditions.

[0105] Step 807: Deploy the application.

[0106] Here, the parameters of the validated neural network model are solidified and ported to the robot's onboard controller. During real-time operation, this neural network model works in conjunction with the inverted pendulum model and the velocity response model, outputting residual velocity in real time to correct the execution speed. This achieves the transformation from offline training to online inference, enabling the robot to have dynamic compensation capabilities for complex environmental disturbances.

[0107] The model training strategy includes offline pre-training and online fine-tuning. Online fine-tuning refers to adjusting various parameters in the physical foundation model online, including the time constant of the first-order oscillation model and the weight coefficients of the three-dimensional inverted pendulum model. The training process also employs a course-learning mechanism. The sample data sequentially includes data on the robot's low-speed linear motion, medium-speed curved motion, full-speed turning motion, and motion on complex terrain. This simulates a student learning a course by first allowing the model to learn simple data and then difficult data, thus improving the model training effect.

[0108] By applying the above embodiments of this application, the desired motion state is predicted based on the landing point information using an inverted pendulum model, ensuring that the motion planning conforms to physical laws; the velocity response model is used to calculate the velocity change based on the actual velocity and the desired velocity, reflecting dynamic response characteristics; the residual velocity is predicted using a neural network model to compensate for deviations caused by nonlinear factors that may exist in the physical model (i.e., the inverted pendulum model and the velocity response model); finally, the execution speed is determined by combining the actual velocity, velocity change, and residual velocity, integrating physical theory planning, dynamic response, and residual correction, which significantly improves the prediction accuracy of the execution speed. As a result, the motion trajectory obtained based on the integral of this execution speed is more consistent with the real situation, thereby improving the accuracy and stability of controlling the robot to move according to the motion trajectory.

[0109] The following description continues to illustrate the exemplary structure of the robot control device 555 provided in the embodiments of this application as a software module. In some embodiments, such as Figure 2 As shown, the software modules stored in the robot control device 555 in the memory 550 include: a first prediction module 5551, used to predict the desired motion state of the robot based on the robot's current actual motion state and the landing point information in the control command using an inverted pendulum model; a second prediction module 5552, used to predict the speed change of the robot based on the actual speed in the actual motion state and the desired speed in the desired motion state using a speed response model; a third prediction module 5553, used to predict the residual speed based on the actual motion state and the desired motion state using a neural network model; a determination module 5554, used to determine the execution speed of the robot based on the actual speed, the speed change, and the residual speed; and a control module 5555, used to perform integral processing on the execution speed to obtain the robot's motion trajectory and control the robot to move according to the motion trajectory.

[0110] In some embodiments, the first prediction module 5551 is further configured to obtain the time length of the control cycle corresponding to the control command; the first prediction module 5551 is further configured to determine the model parameters of the inverted pendulum model; obtain a first model expression of the inverted pendulum model, the first model expression indicating the relationship between the actual motion state, the landing point information, the model parameters, the time length and the desired motion state; and perform prediction processing based on the actual motion state, the landing point information, the model parameters and the time length through the first model expression to obtain the desired motion state.

[0111] In some embodiments, the second prediction module 5552 is further configured to: obtain speed change parameters; obtain a second model expression of the speed response model, the second model expression indicating the relationship between the actual speed, the expected speed, the time length, the speed change parameters, and the speed change amount; and perform prediction processing based on the actual speed, the expected speed, the time length, and the speed change parameters using the second model expression to obtain the speed change amount.

[0112] In some embodiments, the control module 5555 is further configured to perform trapezoidal integral processing based on the actual speed, the execution speed, and the time length using the trapezoidal integral formula to obtain the displacement increment of the control cycle; add the displacement increment to the actual position in the actual motion state to obtain the target position, and generate the motion trajectory from the actual position to the target position.

[0113] In some embodiments, the third prediction module 5553 is further configured to acquire the terrain features of the terrain where the robot is currently located, and extract the first state features of the actual motion state and the second state features of the expected motion state; construct target features including the first state features and the terrain features; input the target features and the second state features into the multilayer perceptron included in the neural network model, and perform prediction processing on the target features and the second state features through the multilayer perceptron to obtain the residual velocity.

[0114] In some embodiments, the determining module 5554 is further configured to add the actual speed, the speed change, and the residual speed to obtain an addition result; and use the addition result as the execution speed of the robot.

[0115] In some embodiments, the actual motion state is the motion state of the robot at time t, and the execution speed is the predicted speed of the robot at time t+1; the first prediction module 5551 is further configured to, after controlling the robot to move according to the motion trajectory, obtain the robot's true speed at time t+1; determine the speed difference between the true speed and the execution speed; update the first model parameters of the inverted pendulum model based on the speed difference to obtain the target inverted pendulum model, and update the second model parameters of the speed response model based on the speed difference to obtain the target speed response model.

[0116] It should be noted that the description of the device embodiments in this application is similar to the description of the method embodiments described above, and has similar beneficial effects as the method embodiments, so it will not be repeated here. Any technical details not covered in the robot control device provided in the embodiments of this application can be understood based on the description of the technical details in the above method embodiments.

[0117] This application also provides a computer program product, which includes computer-executable instructions or a computer program stored in a computer-readable storage medium. The processor of an electronic device reads the computer-executable instructions or computer program from the computer-readable storage medium and executes the computer-executable instructions or computer program, causing the electronic device to perform the robot control method provided in this application.

[0118] This application also provides a computer-readable storage medium storing computer-executable instructions or computer programs. When the computer-executable instructions or computer programs are executed by a processor, the processor will execute the robot control method provided in this application.

[0119] In some embodiments, the computer-readable storage medium may be a memory such as RAM, ROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.

[0120] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0121] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a Hyper Text Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0122] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.

[0123] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.

Claims

1. A robot control method characterized by, The method comprises: based on the current actual motion state of the robot and the landing point information in the control instruction, the desired motion state of the robot is obtained by predicting the inverted pendulum model; based on the actual speed in the actual motion state and the expected speed in the desired motion state, the speed change of the robot is obtained by predicting the speed response model; based on the actual motion state and the desired motion state, the residual speed is obtained by predicting the neural network model; based on the actual speed, the speed change and the residual speed, the execution speed of the robot is determined; the execution speed is integrated to obtain the motion trajectory of the robot, and the robot is controlled to move according to the motion trajectory.

2. The method of claim 1, wherein, The method further comprises: obtaining the time length of the control period corresponding to the control instruction; the desired motion state of the robot is obtained by predicting the inverted pendulum model based on the current actual motion state of the robot and the landing point information in the control instruction, comprising: determining the model parameters of the inverted pendulum model; obtaining the first model expression of the inverted pendulum model, the first model expression indicating the relationship between the actual motion state, the landing point information, the model parameters, the time length and the desired motion state; based on the actual motion state, the landing point information, the model parameters and the time length, the first model expression is used for prediction processing to obtain the desired motion state.

3. The method of claim 2, wherein, the speed change of the robot is obtained by predicting the speed response model based on the actual speed in the actual motion state and the expected speed in the desired motion state, comprising: obtaining the speed change parameter; obtaining the second model expression of the speed response model, the second model expression indicating the relationship between the actual speed, the expected speed, the time length, the speed change parameter and the speed change; based on the actual speed, the expected speed, the time length and the speed change parameter, the second model expression is used for prediction processing to obtain the speed change.

4. The method of claim 2, wherein, the execution speed is integrated to obtain the motion trajectory of the robot, comprising: based on the actual speed, the execution speed and the time length, trapezoidal integral processing is carried out by using trapezoidal integral formula to obtain the displacement increment of the control period; the actual position in the actual motion state is increased by the displacement increment to obtain the target position, and the motion trajectory from the actual position to the target position is generated.

5. The method of claim 1, wherein, the residual speed is obtained by predicting the neural network model based on the actual motion state and the desired motion state, comprising: obtaining the terrain feature of the terrain where the robot is currently located, and extracting the first state feature of the actual motion state and the second state feature of the desired motion state; constructing the target feature including the first state feature and the terrain feature; The target feature and the second state feature are input into a multi-layer perceptron included in the neural network model, and the target feature and the second state feature are predicted and processed by the multi-layer perceptron to obtain the residual velocity.

6. The method of claim 1, wherein, The execution speed of the robot is determined based on the actual velocity, the velocity change amount, and the residual velocity, including: The actual velocity, the velocity change amount, and the residual velocity are added to obtain an addition result; The addition result is used as the execution speed of the robot.

7. The method according to any one of claims 1 to 6, wherein The actual motion state is the motion state of the robot at time t, and the execution speed is the predicted speed of the robot at time t+1; After the robot moves according to the motion trajectory, the method further includes: Obtaining the real speed of the robot at time t+1; Determining the speed difference between the real speed and the execution speed; Based on the speed difference, updating the first model parameter of the inverted pendulum model to obtain a target inverted pendulum model, and updating the second model parameter of the speed response model based on the speed difference to obtain a target speed response model.

8. A robot control device characterized by comprising: The device includes: A first prediction module configured to predict an expected motion state of the robot based on an actual motion state of the robot and foot point information in a control instruction by an inverted pendulum model; A second prediction module configured to predict a velocity change amount of the robot based on an actual velocity in the actual motion state and an expected velocity in the expected motion state by a speed response model; A third prediction module configured to predict a residual velocity of the robot based on the actual motion state and the expected motion state by a neural network model; A determination module configured to determine an execution speed of the robot based on the actual velocity, the velocity change amount, and the residual velocity; A control module configured to integrate the execution speed to obtain a motion trajectory of the robot, and control the robot to move according to the motion trajectory.

9. An electronic device, comprising: The electronic device includes: A memory configured to store computer executable instructions; A processor configured to execute the computer executable instructions stored in the memory to implement the robot control method of any one of claims 1 to 7. 10.A computer readable storage medium storing computer-executable instructions or a computer program, and the computer-executable instructions or the computer program comprise the following steps: The computer executable instructions or computer programs are executed by the processor to implement the robot control method of any one of claims 1 to 7. ​

Images