Application of reinforcement learning based on neural network to dynamic obstacle avoidance

This paper focuses on the application of reinforcement learning to obstacle avoidance in dynamic environments. Behavior-based control architecture is more robust and better in real-time performance than conventional model based architecture in the control of mobile robot. An intelligent controller is proposed by integrating reinforcement learning with the behavior-based control architecture and applied to the obstacle avoidance. Neural network is used to approximate the Q-function to store the Q-value. By using the reinforcement learning, the mobile robot can learn to select proper behavior online without knowing the exact model of the system. In experiments, dynamic and static obstacles are placed in the environments separately. Experiment results show that the mobile robot can get to the target point without colliding with any obstacle after a period of learning.