Risto Kojcev (Erle Robotics), Alejandro Hernández Cordero (Erle Robotics), Asier Bilbao Calvo (Erle Robotics), Irati Zamalloa Ugarte (Erle Robotics), Yue Leire Erro Nuin (Erle Robotics), Víctor Mayoral Vilches (Erle Robotics)
Rather than programming, training allows robots to achieve behaviors that generalize better. However, such training requires big amount of experimentation which isn’t always feasible for physical robots. In this talk, we discuss gym-gazebo , a toolkit for developing and comparing reinforcement learning algorithms using ROS and Gazebo. In contrast to other alternatives, gym-gazebo applies roboticists’ tools, simplifying the development and deployment processes on real robots. We discuss the architecture of gym-gazebo and introduce early results that demonstrate that, for simple tasks, gym-gazebo accelerates the robot training time by more than 33% while maintaining same levels of accuracy and repeatability.