Deep Reinforcement Learning for Mapless Navigation of Unmanned Aerial Vehicles

摘要

This paper presents a deep reinforcement learning-based system for goal-oriented mapless navigation for Unmanned Aerial Vehicles (UAVs). In this context, image-based sensing approaches are the most common. However, they demand high processing power hardware which are heavy and difficult to embed into a small-autonomous UAV. Our approach is based on localization data and simple sparse range data to train the intelligent agent. We based our approach in two state-of-the-art Deep- Rl techniques for terrestrial robot: Deep Deterministic Policy Gradient (DDPG) and Soft Actor Critic (SAC). We compare the performance with a classic geometric-based tracking controller for mapless navigation of UAVs. Based on experimental results, we conclude that Deep- Rl algorithms are effective to perform mapless navigation and obstacle avoidance for UAVs. Our vehicle successfully performed two proposed tasks, reaching the desired goal and outperforming the geometric-based tracking controller on the obstacle avoiding capability.