Vision Based Leader-Follower Control of Wheeled Mobile Robots using Reinforcement Learning and Deep Learning

Vision-based control of mobile robots often involves complex calculations to derive a control law. The reinforcement learning algorithm (Q-learning) offers a machine learning method to extrapolate a control law from an environment given discretized actions, without the need of complex calculations. In this paper, a vision-based controller is created using Q-Learning to enable tracking in a leader-follower configuration of two nonholonomic autonomous mobile robots. The follower robot gathers its desired trajectory values by using a deep learning SSD model to identify a distinguishing visual feature on the leader robot and uses a lidar to determine the distance between two robots. These parameters are utilized to select an optimal action of the follower robot through reinforcement learning. The emulated results in a ROS Gazebo environment show this method to be effective in enabling a wheeled mobile robot to follow another, while simultaneously avoiding obstacles.