Research On End-to-End Learning And Optimization Of Neural Network Controller In Robotics

The motion planning of an intelligent car can be divided as decision and planning in the upper layer and controlling in the lower layer.In the process of decision and planning,rule-based methods are usually adopted,such as finite state machine method.These methods are simple and easy to verify,but the number of rules is always finite.Therefore,rule-based methods have poor performances in dynamic and complicated conditions.When the model of an environment is known before,the path planning problem can be relatively simple since the car has all the prior knowledge about the environment and the positions of all the obstacles are still,but the car cannot obtain the change around it in a dynamic environment.Recently,end-to-end learning method based on deep learning has gained a lot of attentions,in this method,a car will not complete a task by human rules but instead learn to achieve a goal through interacting with the environment.This is because learning-based methods have better generalization ability which enable the car to get to the goal in a dynamic and more complicated environment.In the lower layer of controlling,existing dynamic control methods usually face the difficulty to get accurate mathematical model descriptions and lack sufficient nonlinear expression capabilities.To mitigate this problem,a wavelet-Elman fuzzy brain emotional leaning network controller is proposed.In the new controller,an Elman neural network and a wavelet neural network is adopted to represent the amygdala channel and the orbitofrontal cortex channel,respectively,which enables the new controller with temporal memory and faster convergence ability.In addition,a robustcontroller is used as a compensator for the neural network controller to guarantee therobust tracking for a dynamic system.To verify the proposed methods,a simulated lane following task is implementedon ROS and V-rep,and for controlling experiment,different trajectories tracking experiments and auto-landing experiment of a quadrotor are designed to verify the generalization ability and robustness of the proposed controller.