Universal Successor Features Based Deep Reinforcement Learning for Navigation

Abstract

This thesis outlines a research work on modelling a novel approach for robot navigation by using an advanced Deep Reinforcement Learning (DRL) algorithm. Visual navigation is a core problem in the robotics and machine vision. Previous research used map-based, map-building or map-less navigation strategies. The first two approaches were favoured in the past, however, they essentially depend on the accurate mapping of the environment and a careful human-guided training phase, which overall limits generalizability. With recent developments in DRL, map-less navigation experienced major advancements. A current challenge for DRL algorithms is learning new tasks or goals. This ability is called transfer learning. To cope with the challenges in transfer learning and performance, we present a new approach using Universal Successor Features (USF) in this thesis. We propose several models that we applied for the task of target driven visual navigation in a complex photo-realistic environment using a simulator named as AI2THOR. With the evaluation of our proposed models in AI2THOR, we demonstrate that an agent is able to successfully improved the ability to reach goals which the agent was initially not trained for.