Evaluating deep reinforcement learning algorithms for quadrupedal slope handling

Abstract

In recent years, a number of deep reinforcement learning (DRL) algorithms have emerged that promise to automate the development of locomotion controllers and map sensory observations to low-level actions. However, legged locomotion still is a challenging task for DRL algorithms, especially when slope handling is required. As a result, a framework using commonly used tools (ROS, Gazebo, etc.) and specific slope handling scenarios would enable the evaluation of recent DRL algorithms in order to choose the appropriate algorithm for a given task. In this work, an evaluation framework is proposed that combines DRL with trajectory planning at toe level aiming at reducing training time and facilitate decision-making in slope-handling cases. The proposed evaluation scheme is extensively tested in a Gazebo environment and valuable results are produced using three state-of-the-art DRL algorithms.