IEEE/RSJ International Conference on Intelligent Robots and Systems
1-10-2018 to 5-10-2018
[en] Model Learning for Control ; Reactive and Sensor-Bas
[en] Navigation in unknown indoor environments with fast collision avoidance capabilities is an ongoing research topic. Traditional motion planning algorithms rely on precise maps of the environment, where re-adapting a generated path can be highly demanding in terms of computational cost. In this paper, we present a fast reactive navigation algorithm using Deep Reinforcement Learning applied to multi rotor aerial robots. Taking as input the 2D-laser range measurements and the relative position of the aerial robot with respect to the desired goal, the proposed algorithm is successfully trained in a Gazebo-based simulation scenario by adopting an artificial potential field formulation. A thorough evaluation of the trained agent has been carried out both in simulated and real indoor scenarios, showing the appropriate reactive navigation behavior of the agent in the presence of static and dynamic obstacles.