Mobile System Modeling Of Wheeled Robots And Model Learning-Based Research On Tracking Control

The increasing demand of mobile robots task given and diversity of exposure to environment, leads to more and more random factors in the implementation of mobile robots. As the research object of this dissertation, the major difficulties in the current development of wheeled mobile robots include: modeling the dynamical system with high-precision, the external environment cognition, the real applications of interactional mechanics theory between wheel and ground, and so on. All these factors usually lead to the uncertainties of robot models, which bring lots of difficulties to robots control accurately. The advanced control methods are necessary to develop for solving the control problem of the complex system. Therefore, the system models of wheeled robot are established, and the advanced model learning methods are proposed, furthermore, the model learning methods are applied to solve the tracking control problems of wheeled mobile robot.System model is the basic tool of robotics research, kinematics and dynamics models of wheeled robot is the basis for solving control problems. According to nonholonomic constraint of wheeled robots on the hard ground, the kinematics and dynamics models of robot with wheels’ longitudinal slip and lateral skid are established. The dynamics model with wheels’ slip, skid and sinkage on the soft ground is established by combining wheel-ground terramechanics equations and kinematics modeling of wheeled robot.The uncertainties of wheeled robots system model influence the accuracy of tracking control in a large extent. In order to improve the accuracy of control system,neural networks with nonlinear character are applied to identify models as online,where forward radial basic function(RBF) and recurrent neural networks are discussed. Since the hidden layer of recurrent neural network produces the delay time due to the again activation of neurons, the new globally stable conditions of recurrent neural network with delay time are proposed by combining Lyapunov stable theory, linear matrix inequation skill, Lie algebra feature and so on.To learn the robot system model with uncertainties by observed data,nonparametric Gaussian process regression(GPR) model is discussed which is able to identify the potential system model with high-precision by reducing the effect of noise. Cholesky decomposition-based GP model update can ensure the learning speed of model learning as increasing data. According to the similarity between the supervised neural network and Gaussian process in Bayesian regression, the model of Bayesian regression network is established. The novel cluster methods are proposed by local learning theory, and then local Bayesian regression network(LBRN) is established. Model learning for the complex systems, LBRN has thecharacters of rapid learning, great identification accuracy and excellent robustness.To achieve the trajectory tracking control of wheeled robot on the hard ground,model learning methods are used to design the effective control laws. For the situation of wheeled robot with uncertain model parameters, feedback error learning algorithm and RBF neural network-based controller is designed. Neural network is applied to identify the uncertain nonlinear dynamics model as online, and then, an accurate trajectory tracking control is achieved by simulation. To reduce the effect of wheels’ longitudinal slip for trajectory tracking, the formula of computing slip parameters are built. The controller is designed by combining slip parameters compensation and neural network, and then, a trajectory tracking control with wheels’ slip is achieved by simulation. As the situation of wheels occurring the coupling of wheels’ longitudinal and lateral slip, slip parameters are difficult to estimate. LBRN is applied to identify the slip dynamics model as offline, and then,the slip model-based trajectory tracking control is achieved by simulation.To achieve the path following control of wheeled robot with wheels’ slip, skid,sinkage etc. on the soft ground, the effective control law is designed by combining terramechanics and model learning. To identify the unknown parameters of soft ground, according to the analysis of wheel-ground mechanics and single wheel’s test platform, the terrain parameters are identified by dynamic back propagation algorithm-based recurrent neural network. Since the multi-parameters coupling models of wheel-ground mechanics is difficult to apply in driving control of wheels.Gaussian process regression and the observed data by single wheel’s test platform are used to identify the desired model of wheel-ground mechanics, control of wheel’s slip ratio, sinkage and drawbar pull are achieved by wheels driving torque.For the complex robot system with uncertainties on soft ground, the input and output mapping relationship of system model is established. LBRN is applied to identify the system model, and design the effective controller, and then, the accurate path following is achieved by simulation.In this dissertation, the system models of wheeled mobile robot are established,and the complex situation with models uncertainties is discussed. Neural networks,Gaussian process regression and LBRN-based model learning methods are studied and applied to design the control system for wheeled mobile robot, and accurate tracking control of robot on different grounds are achieved.