Asynchronous deep reinforcement learning for the mobile robot navigation with supervised auxiliary tasks

In this paper, we present the method based on asynchronous deep reinforcement learning adapted for the mobile robot navigation with supervised auxiliary tasks. We apply the hybrid Asynchronous Advantage Actor-Critic (A3C) algorithm CPU/GPU based on TensorFlow. The mobile robot is simulated as the navigation tasks on the OpenAI-Gym-Gazebo-based environment with the collaboration with ROS Multimaster. The supervised auxiliary tasks include the depth predictions and the robot position estimation. The simulated mobile robot shows the capability to learn to navigate only the input from raw RGB-image and also perform recognition of the place on the map. We also show that the combination of all possible auxiliary tasks leads to the different learning rate.