Generating Diverse Challenging Terrains for Legged Robots Using Quality-Diversity Algorithm

Recent progress in legged robotics [1]-[4], particularly through the use of reinforcement learning (RL), has led to significant improvements in their performance in navigating complex terrains. However, despite these advances, significant challenges remain in ensuring the robustness of such systems, particularly when navigating unstructured terrains. Traversing unstructured terrains is crucial in applications that require the exploration of hazardous areas, such as rescue operations or underground inspections. Many studies rely on hand-crafted terrains, such as stairs, slopes, and discrete obstacles, or employ uncontrollable noise or Perlin noise [4], [5] to generate them. While these methods allow for training and testing controllers on a variety of terrains, their scope is limited, and they do not ensure the controller's reliability across all possible terrains. Critical corner cases may be missed, and, given the diversity of terrains the robot might encounter, these cases can be difficult to identify, especially as weaknesses can differ widely depending on the controller's design. Moreover, such weaknesses are often difficult to discover manually. In [6], 100 volunteers were asked to identify weaknesses in a quadruped robot's controller by applying pushing forces and overwriting velocity commands.