Learning Locomotion over Challenging Terrain
ANYmal on Science Robotics cover
The ability to traverse deformable surfaces (mud or snow) and dynamic footholds (rubble), as well as handle impediments such as thick vegetation and flowing water, is key to successfully navigating unstructured natural environments. Lee et al. have developed a locomotion controller that uses deep reinforcement learning to teach a quadruped robot how to navigate unseen and unstructured environments without the need for external sensors. The quadruped, ANYmal, was deployed in various outdoor settings to demonstrate that it could robustly traverse a range of challenging terrain relying solely on proprioception. This month's cover is a photograph of ANYmal atop Tenner Chrüz, Tenna, Switzerland (see also the Focus by Ha). [IMAGE CREDIT: SIMON TANNER/NZZ]
