A design methodology for deep reinforcement learning in autonomous systems

Autonomous systems such as mobile robots will play an important role in fields like industrial production, transportation or in hostile environments such as space. One of the most fundamental problem in autonomous mobile robotics is autonomous navigation. It is imperative for a mobile robot to learn to navigate in complex environments such as roads or buildings. The most popular approach to this problem is to utilize different algorithms for mapping the environment, self-localization in the map, planning a trajectory to the given goal and executing this trajectory. However, there are some drawbacks of these approaches. We often make assumptions about the environment such as no dynamic or transparent objects. Moreover, there is considerable overhead, they do not learn from failures and operation scenarios. This prompts us to search for alternative approaches for autonomous navigation, such as deep reinforcement learning. However, the application of deep reinforcement learning to a particular task involves a series of non-trivial design decisions. Previous work have failed to address the need for a design methodology for deep reinforcement learning systems. In this paper, we propose design methodology and discuss relevant design decisions for deep reinforcement learning in autonomous systems. We apply the methodology to the problem of autonomous navigation.