In this paper we consider the problem of robot navigation in simple maze-likeenvironments where the robot has to rely on its onboard sensors to perform thenavigation task. In particular, we are interested in solutions to this problemthat do not require localization, mapping or planning. Additionally, we requirethat our solution can quickly adapt to new situations (e.g., changingnavigation goals and environments). To meet these criteria we frame thisproblem as a sequence of related reinforcement learning tasks. We propose asuccessor feature based deep reinforcement learning algorithm that can learn totransfer knowledge from previously mastered navigation tasks to new probleminstances. Our algorithm substantially decreases the required learning timeafter the first task instance has been solved, which makes it easily adaptableto changing environments. We validate our method in both simulated and realrobot experiments with a Robotino and compare it to a set of baseline methodsincluding classical planning-based navigation.
展开▼
