Memory-based Deep Reinforcement Learning for Obstacle Avoidance in UAV\n with Limited Environment Knowledge

摘要

This paper presents our method for enabling a UAV quadrotor, equipped with a\nmonocular camera, to autonomously avoid collisions with obstacles in\nunstructured and unknown indoor environments. When compared to obstacle\navoidance in ground vehicular robots, UAV navigation brings in additional\nchallenges because the UAV motion is no more constrained to a well-defined\nindoor ground or street environment. Horizontal structures in indoor and\noutdoor environments like decorative items, furnishings, ceiling fans,\nsign-boards, tree branches etc., also become relevant obstacles unlike those\nfor ground vehicular robots. Thus, methods of obstacle avoidance developed for\nground robots are clearly inadequate for UAV navigation. Current control\nmethods using monocular images for UAV obstacle avoidance are heavily dependent\non environment information. These controllers do not fully retain and utilize\nthe extensively available information about the ambient environment for\ndecision making. We propose a deep reinforcement learning based method for UAV\nobstacle avoidance (OA) and autonomous exploration which is capable of doing\nexactly the same. The crucial idea in our method is the concept of partial\nobservability and how UAVs can retain relevant information about the\nenvironment structure to make better future navigation decisions. Our OA\ntechnique uses recurrent neural networks with temporal attention and provides\nbetter results compared to prior works in terms of distance covered during\nnavigation without collisions. In addition, our technique has a high inference\nrate (a key factor in robotic applications) and is energy-efficient as it\nminimizes oscillatory motion of UAV and reduces power wastage.\n