Mapless ground robot navigation in forests through Reinforcement Learning

Due to the complexity and heterogeneity of the landscape, autonomous navigation for ground robots in outdoor unstructured environments still poses many challenges. The main difficulty lies in the wide variety of potential obstacles, thus hindering the possibility to fully and accurately model these types of scenarios. We propose a neural network architecture which, given only 3D LIDAR information and a goal, outputs a velocity command for the low-level navigation controller. The network is trained with reinforcement learning without any assumption on the kinematic model of the vehicle, but taking into account only the inability to perform sideways motion and the physical size of the vehicle. The experiments show convincing results of the architecture, allowing the robot to successfully perform point goal navigation and avoid obstacles with smooth trajectories.