[0034] Example 1
[0035] like figure 1 As shown, a four-legged robot single leg trajectory tracking control method, including:
[0036] Get the joint parameters of the quadruple robot;
[0037] Establish a single leg dynamics model according to joint parameters;
[0038] Based on disturbance and neural network to correct single leg dynamics model, construct a single leg dynamics model in the case of disturbance, and assume a reference trajectory;
[0039] Using a value-based function approximation method based on neural network learning, the optimal parameters of the reference trajectory are obtained, and the optimal control rate is obtained;
[0040] Based on the optimal control rate to realize the single leg trajectory tracking control of the quadruple robot.
[0041] Further, the joint parameters include joint position, speed, and acceleration vectors.
[0042] Further, according to the joint parameters, the one-leg dynamic model is established, and the control torque vector is obtained according to the symmetrical inertial matrix of the joint position, the speed, the acceleration vector, and the CCC and the centrifugal force matrix and the gravity vector are obtained.
[0043] Specifically, 1.1 four-foot robot single leg 3 free simplified kinetic model is expressed as:
[0044]
[0045] in The joint position, speed, and acceleration vector, M (θ) are symmetrical inertial matrices, For the Corosary and Ceremony Matrix, For the gravity vector, u (θ) is gravity, τ = r n×n For the control torque vector.
[0046] 1.2 According to the actual quadruple robot joint parameters, the specific parameters of the kinetic equation of the four-legged robot 3 freedom single legs are as follows
[0047]
[0048] in,
[0049]
[0050]
[0051]
[0052] in,
[0053]
[0054]
[0055]
[0056]
[0057]
[0058] Further, the neural network is an RBF neural network, using a neural network approximate continuous function, and the weight of the neural network is defined by an error between input and desired output. The process is as follows:
[0059] 2.1RBF neural network for approximate continuous functions h (x): r n → r
[0060]
[0061] Where X is the input variable, It is the connection between the hidden layer and the output layer, It is a base function.
[0062] RBF neural networks can put continuous function h (x) in a tight Ω Z It is approximated to any desired accuracy, as shown below:
[0063]
[0064] Where W * It is an ideal weight, and || Δ || ≤ ε is approximation error, where ε> 0 is a constant.
[0065] Further, the neural network based on the single leg dynamics model considering disturbance includes,
[0066] Redefine the joint parameters of the robot according to the external disturbance parameters received by the robot;
[0067] Set the reference trajectory according to the redefined joint parameter, define the generalized tracking error;
[0068] Introduce virtual control and second error variables and get gain matrices;
[0069] Time-guided gain matrix, based on the Lee Point candidate function to obtain a single leg dynamics model considering disturbance.
[0070] The external disturbance parameters are large-scale disturbances and parameters uncertainty in practical applications. When the perturbation performance is mainly interactively interactively interact with complex terrain, the parameter uncertainty is mainly manifested by the uncertainty of the late disturbance or slip. Some kinetic parameters, such as rotation inertia, etc. are estimated; due to the external disturbance parameters, it is necessary to obtain real-time acquisition according to actual impact scenarios. This application uses disturbance and neural network to correct single leg dynamics models, and construct consider disturbance. The single leg dynamics model is assumed, and the reference trajectory is assumed; the accurate external disturbance parameters are acquired using a value function based on neural network learning, thereby obtaining the optimal control rate, thereby tracking the four-legged robot single leg trajectory tracking control.
[0071] Specifically, the three-legged kinetic equation of the four-legged robot considers the extrusion time, and it can be written as the following form.
[0072]
[0073] Where f is an external disturbance from the robot. Redefine parameters, let X 1 = Θ, available:
[0074]
[0075]
[0076] 3.2 In order to design control torque, the system can achieve better target tracking, and set the reference track x r (t) can be expressed as X r (t) = [ 1r (t), θ 2r (t), θ 3r (t)] T. When status information X 1 X 2 Fully known, define a generalized tracking error z 1 (t) = x 1 (t) -X r (t), and
[0077] 3.3 Introducing a virtual control α 1 (t) and the second error variable defined as z 2 (t) = x 2 (t) -α 1 (t). choose
[0078]
[0079] Gain matrix K 1 = K 1 T 0, then
[0080]
[0081] Time to guide
[0082]
[0083] 3.4 Consider a Li Yapuov candidate function:
[0084]
[0085] Combined equation (9) to V 1 Seek time derivative:
[0086]
[0087] Then consider the Li Yapanov candidate function:
[0088]
[0089] The above formula is guided
[0090]
[0091] 3.5 The model-based control can be designed as follows:
[0092]
[0093] Gain matrix K 2 = K 2 T 0, the equation (14) is brought into the Li Yapanov candidate function:
[0094]
[0095] Further, the optimal parameters of the reference trajectory are obtained using the value of the value based on neural network learning, and the optimal control rate includes,
[0096] Set the normal number of external disturbance parameters Define the operator and establish a model to establish a controller according to the model;
[0097] The optimal parameters of the reference trajectory are obtained using the value of the reference trajectory to obtain the optimal parameters of the reference trajectory, and the controller is determined in the optimal control rate.
[0098] That is, using the RBF neural network estimator to approximate the uncertainty of robot parameters, improve the tracking performance of robots, and the control rate is designed:
[0099]
[0100] here Is the weight of the neural network, Is a base function, this neural network Approximate
[0101] Specifically, f is an unknown amount, that is, there is a normal number vector when Time Because when the flow of f (t) has limited energy, it is bound to be bound, ie f (t) ∈L ∞.
[0102] Definition operator "⊙"
[0103]
[0104]
[0105] Here A = [A 1 , a 2 ] T And b = [b 1 , B 2 ] T It is two two-dimensional vector.
[0106] The controller based on the above model is:
[0107]
[0108] Gain matrix K 2 = K 2 T 0, thus
[0109]
[0110] Due to the parameter M (X 1 ), C (x 1 , X 2 ), G (x 1 ), There is uncertainty in f, using the RBF neural network estimator to approximate the uncertainty of robot parameters, and improve the tracking performance of the robot, and the control rate is redesigned as:
[0111]
[0112] here Is the weight of the neural network, Is a base function, this neural network Approximate which is:
[0113]
[0114] here Is the input variable of the neural network, It is approximate error, and the adaptation rate is designed as:
[0115]
[0116] Γ here i Is a constant gain matrix, and φ i 0, i = 1, 2, ... n, is a small value constant.