Exploring Deep Reinforcement Learning for MAV Trajectory Tracking and Collision Avoidance

Abstract

In this extended abstract, propose a novel Deep Reinforcement Learning (DRL) approach that enables Micro Aerial Vehicles (MAVs) to autonomously track reference trajec- tories while effectively avoiding collisions using only depth maps and the robot's current pose. The proposed method imposes no constraints on the drone's mobility, allowing free navigation in three-dimensional space and does not require detailed informa- tion about the environmental characteristics and the shape and distribution of obstacles. Through comprehensive evaluations in photo-realistic simulated environments, we demonstrate the effec- tiveness and generalization capabilities of our strategy compared to state-of-the-art baselines.