Sample Efficient Reinforcement Learning for Navigation in Complex Environments

Abstract

Navigation of mobile robots in unstructured, time-varying environments is challenging. It becomes even more complicated in disaster scenarios where logistical difficulties, as well as technical issues such as reactive and time-varying obstacles, exist. These scenarios are too complex for classical obstacle avoidance methods to navigate through successfully. This paper presents a sample efficient reinforcement learning algorithm for navigation in complex environments. The approach augments training data with randomly generated target location data to accelerate learning. A Q-learning approach is implemented, which is capable of quick training with limited episodes. The procedure is tested in four scenarios in Gazebo and one scenario in a real-world experiment. In the two simulation scenarios with no obstacles, the method can learn to navigate towards the target in fewer than 200 episodes. For environments with moving obstacles, training takes slightly longer, but the process is still able to learn an effective policy quickly.