Motion control method and system for biomimetic robotic fish based on adversarial structured control

Patent metadata

Field	Value
Publication number	US-10962976-B1
Application number	US-202017094820-A
Country	US
Kind code	B1
Filing date	Nov 11, 2020
Priority date	Nov 29, 2019
Publication date	Mar 30, 2021
Grant date	Mar 30, 2021

Abstract

A motion control method and system for a biomimetic robotic fish based on an adversarial structured control, includes: taking the accuracy and speed of motion to the target point as a reward term, and taking a power sum of servomotors as a loss term to construct an optimization objective function; optimizing parameters of a central pattern generator model that generates a global control quantity of a servomotor, after curing its parameters, optimizing the parameters of the servomotor compensation control model; iteratively optimizing the parameters of the model; obtaining the global control signal and compensation control signal of the biomimetic robotic fish through the trained model, and using the linear combination of the two sets of output signals as the control signal of the servomotor of the robotic fish to realize the motion control of the fish.

First claim

What is claimed is:

1. A motion control method for a biomimetic robotic fish based on an adversarial structured control, comprising:
step S 10 , obtaining a swimming path of the biomimetic robotic fish, and dividing the swimming path into a set of subpaths, wherein the set of subpaths is connected sequentially;
step S 20 , based on a start point and an end point of each subpath in the set of subpaths in sequence, obtaining a global control quantity of each servomotor of the biomimetic robotic fish at a time t through a trained servomotor global control model;
step S 30 , based on obtained position and pose information of the biomimetic robotic fish at the time t and the global control quantity of the each servomotor of the biomimetic robotic fish at the time t, obtaining a compensation control quantity of the each servomotor of the biomimetic robotic fish at the time t through a trained servomotor compensation control model;
step S 40 , summing the global control quantity of the each servomotor of the biomimetic robotic fish at the time t and the compensation control quantity of the each servomotor of the biomimetic robotic fish at the time t, to obtain a control quantity of the each servomotor of the biomimetic robotic fish at a time t+1, wherein the control quantity of the each servomotor is a sum of the global control quantity of the each servomotor and the compensation control quantity of the each servomotor, and performing a motion control on the biomimetic robotic fish at the time t+1 through the control quantity of the each servomotor of the biomimetic robotic fish at the time t+1; and
step S 50 , letting t=t+1, and returning to step S 20 until the biomimetic robotic fish reaches an end of the swimming path.

2. The motion control method according to claim 1 , wherein the trained servomotor global control model comprises a set of servomotor global control submodels, and the trained servomotor compensation control model comprises a set of servomotor compensation control submodels, wherein the set of servomotor global control submodels and the set of servomotor compensation control submodels are constructed for different types of subpaths, and the set of servomotor global control submodels are in a one-to-one correspondence with the set of servomotor compensation control submodels.

3. The motion control method according to claim 2 , wherein the servomotor global control submodels are constructed based on a central pattern generator (CPG) model; the servomotor compensation control submodels are constructed based on a deep deterministic policy gradient (DDPG) network; the servomotor global control submodels and the servomotor compensation control submodels are trained by an iterative adversarial method, and the iterative adversarial method comprises:
step B 10 , constructing an optimization objective function of the servomotor global control submodels and the servomotor compensation control submodels;
step B 20 , optimizing parameters of the servomotor global control submodels by an ES algorithm according to a predetermined gradient descent direction of a first gradient function, until a value of the optimization objective function does not increase or an increased value of the optimization objective function is lower than a predetermined first threshold value, to obtain a first servomotor global control submodel;
step B 30 , based on parameters of the first servomotor global control submodel, optimizing parameters of an action strategy network and an action value network in the servomotor compensation control submodels according to a predetermined gradient descent direction of a second gradient function, until the value of the optimization objective function does not increase or the increased value of the optimization objective function is lower than the predetermined first threshold value, to obtain a first servomotor compensation control submodel; and
step B 40 , based on the parameters of the first servomotor compensation control submodel, returning to step B 20 to iteratively optimize the parameters of the servomotor global control submodels and the servomotor compensation control submodels, until the value of the optimization objective function does not increase or the increased value of the optimization objective function is lower than the predetermined first threshold value, to obtain trained servomotor global control submodels and trained servomotor compensation control submodels.

4. The motion control method according to claim 3 , wherein the optimization objective function is expressed as:
max J ψ =cos(θ e )·{right arrow over ( v m )}−β·{right arrow over (τ)}· , s.t .∥{right arrow over ( v m )}∥≤ v 0 ;

wherein, ψ represents an object optimized by the optimization objective function, namely a parameter of the CPG model and a parameter of the DDPG network; θ e represents a yaw angle between the biomimetic robotic fish and a target point, and θ e ∈(−π,π] is a predetermined range of the yaw angle; {right arrow over (v m )} represents a velocity vector of the biomimetic robotic fish under a world reference system; ∥{right arrow over (v m )}∥ represents a modulus of the velocity vector, v 0 is a predetermined velocity upper limit, wherein the predetermined velocity upper limit is configured to ensure an effect of energy consumption optimization; τ represents a torque vector of the each servomotor of the biomimetic robotic fish and represents an angular velocity vector of the each servomotor of the biomimetic robotic fish; β is a positive value and β the indicates a correlation coefficient between reward and loss.

5. The motion control method according to claim 3 , wherein the first gradient function is expressed as:

∇
θ

⁢

E

E
∼

N
⁡

(

0
,
I

)

⁢

F
⁡

(

θ
+
σɛ

)

=

1
σ

⁢

E

E
∼

N
⁡

(

0
,
I

)

⁢

{

F
⁡

(

θ

Assignees

• Inst Automation Cas

Inventors

• Wu Zhengxing
• Yu Junzhi
• Yan Shuaizheng
• Wang Jian
• Tan Min

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G05D1/10

  Simultaneous control of position or course in three dimensions (G05D1/12 takes precedence) · CPC title

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/094

  Adversarial learning · CPC title