Learning to Navigate Over Clutter in Indoor Environments using Vision

Anonymous

Keywords: Legged Locomotion, Reinforcement Learning, Visual Navigation

TL;DR: We present Visual Navigation and Locomotion over obstacles (ViNL), which enables a quadrupedal robot to navigate unseen apartments while stepping over small obstacles that lie in its path (e.g., shoes, toys, cables).

Abstract: We present Visual Navigation and Locomotion over obstacles (ViNL), which enables a quadrupedal robot to navigate unseen apartments while stepping over small obstacles that lie in its path (e.g., shoes, toys, cables). ViNL consists of: (1) a visual navigation policy that outputs linear and angular velocity commands that guides the robot to a goal in novel indoor environments; and (2) a visual locomotion policy that controls the robot’s joints to avoid stepping on obstacles while following provided velocity commands. These two policies are trained independently, and can be seamlessly be coupled together upon deployment by feeding the velocity commands from the navigation policy to the locomotion policy. While several related prior works have demonstrated learning visual navigation policies or learning robust locomotion control, to the best of our knowledge, this is the first fully learned approach that leverages vision to accomplishes both (1) intelligent navigation in new environments, and (2) intelligent visual locomotion that aims to traverse cluttered environments without disrupting obstacles. We find that ViNL outperforms prior work that was trained to robustly walk over challenging terrain using privileged terrain maps (+32.8\% success and -4.4 collisions per meter).