On mapless navigation through deep reinforcement learning

Abstract

We present a deep reinforcement learning based local planer for a mobile robot which can navigate towards goal locations using only a sparse 20-dimensional laser scan and relative goal position as inputs and linear and angular velocity as output. We train multiple models end-to-end without any expert demonstrations or handcrafted features using both on-policy and off-policy methods with prioritized experience replay. Traditional local motion planning methods rely on an obstacle cost map that assumes a relatively static environment while our method can continue to operate even under significant environmental changes. Through the use of a stacked recurrent intermediate model architecture, our policies are able to scale more efficiently with environment complexity and can handle dynamic environments significantly better than prior work. We demonstrate that the learned policies can also generalize to novel environments not encountered during training while incurring no additional training cost.