Research On Global Self-stabilizer Of Legged Robot Based On Motion Platform And Training And Learning

The balance control of legged robots is a prerequisite for its practical application.In recent years,the goal of research on balance control problems has shifted from completing actions in a simple environment to how to obtain self-stability with strong robustness in unknown and uncertain environments.The existing balance control methods are mainly based on the dynamics model.As the growth of the driving ability,response speed and other abilities,the legged robots have achieved quite successful experimental results.However,it is still difficult to adapt the unknown environment.Existing robot balance controllers are only designed for one or several disturbances,which leads to weak robustness in uncertain environments.This paper uses the limited random motion of a multi-degree-of-freedom motion platform to simulate random disturbances in uncertain environments.For the first time,this paper proposes obtaining the global self-stabilization ability of legged robots through stability training and intelligent learning.The global self-stability,that is,the ability to adapt to any size,direction,and type of environmental disturbance within the limit of the robot's driving ability.The main content of this paper is divided into the following aspects:The global self-stabilizer of legged robots is proposed and designed based on intelligent learning algorithm.The relationship between the movement of the training platform and the disturbance imposed is analyzed.A general model of the legged robot training system is established.Then the state and action space of the global self-stabilizer are defined.Overall structure of the global self-stabilizer is designed.Specific functional modules are designed by task decomposition and their learning and decision algorithms will be proposed based on Q learning and RBF network algorithms.Regarding the problem of obtaining training data for the global self-stabilizer,balance controllers with different laws are designed.The reflection control of ZMP is extended to arbitrary inclined plane in the three-dimensional space.A posture balance control method based on the Capture Point(CP)is proposed,whose control stability is proved theoretically and verified through simulations.The kinetic energy control method of the legged robot is proposed,whose effectiveness is verified by simulations.The derived balance control laws are combined to construct a variety of model-based balance controllers.These controllers are used to generate training data for the global self-stabilizer of the legged robot.In order to solve the convergence problem caused by the high-dimensional continuous variable space of the global self-stabilizer,accelerated convergence methods of machine learning are studied.For dimension reduction of the system space,a feature selection method called RAFS is proposed based on mutual information theory.The state space abstraction method based on Gaussian function representation is proposed.The state space is divided by the maximum membership principle.In order to verify the effectiveness of the RAFS method,the proposed RAFS method is compared with existing methods on different datasets.The training data obtained under single-degree-of-freedom disturbance is learned.The learning effects before and after the state-space abstraction are compared to verify the effect of the proposed method on accelerating the convergence of the global self-stabilizer.In order to enhance the ability of the motion platform to impose disturbance,its mechanism parameters are optimized.Also,a motion platform system for stability training of biped robots is developed.For solving the forward kinematics problem of the parallel mechanism involved in the mechanism parameter optimization,a direct kinematics numerical algorithm is proposed for the spatial linkage mechanism based on the pseudo-arc-length homotopy continuation method.The parallel mechanism part of a6-DOF series-parallel motion platform is a 4-PSS/PS mechanism.For this mechanism,a mechanism parameter optimization model is established to improve the ability of imposing disturbances.This model is solved by the particle swarm algorithm.The optimal mechanism parameters are obtained and the motion platform for stability training experiment is developed.Simulations and experiments of stability training for the global self-stabilizer are conducted,in which a biped robot GoRoBoT-? is taken as the object.Then the trained global self-stabilizer is tested under the same condition.A virtual prototype system composed of a biped robot and a training platform is sstablished in Matlab and Adams software.Under the condition of limited random disturbance,the stability training and testing simulations are carried out by using different model-based balance controllers.In these simulations,three motions are conducted by the bipped robot,which are standing on both feet,standing on one foot and random stepping respectively.The global self-stabilizer obtained by the simulations is transplanted to the GoRoBoT-? robot's system.Then the experimental research on stability training and testing is conducted.Compare the robot's stability when using the model-based controller and using the trained global self-stabilizer.The action selection process of the global self-stabilizer is analyzed to verify the effect of the proposed global self-stabilizer.