Reinforcement Learning for Autonomous Navigation in Dynamic Environments

Abstract

Reinforcement Learning (RL) has become a strong computational model that empowers robots and autonomous systems to acquire strong navigation behaviors via environment interaction. In dynamic environments where obstacles move, layouts change and the human activity is unpredictable, the traditional rule-based navigation solutions often do not work since they lack adaptability to the ever-changing environment. Deep reinforcement learning (DRL) and other RL methods provide the possibility to discover the best navigation policies that can be generalized to new unobservable situations. The article focuses on the application of RL in autonomous navigation in dynamic environments, including warehouses, outside streets, and multi-agent robots. It gives a systematic overview of the literature available, presents the methodology used, assesses benchmark analysis, and summarizes significant results. Safety, computational demand, real-time decision-making, and sim-to-real transfer are the other issues that are discussed in the paper. Future work recommendations and better system reliability are mentioned.