HYBRID TRAJECTORY PLANNING USING REINFORCEMENT AND BACKPROPAGATION THROUGH TIME TECHNIQUES

Abstract

A novel approach for trajectory planning of a mobile robot is presented. The mobile robot is assumed to move in a two-dimensional workspace with continuous input from the surrounding environment. The input is a signal that reflects the distance and position of an obstacle momentarily. The first part consists of using a neural network to direct the robot that moves from some initial point to a given target with constant speed. The neural network uses an original approach of hybrid instantaneous reinforcement learning in addition to a long-term backpropagation through time learning. Both techniques complement each other by providing online and offline learning. The second stage is to test the learned neural network with different obstacles than the ones used in learning. The neural network should be capable of discovering a strategy to steer the robot. The robot is assumed to move with constant speed and zero acceleration; consequently, the neural network output is just the direction of motion. Future work may include robot dynamics such that the output of the neural network will not only be a direction but also amount of motion.