Exploring Deep Reinforcement Learning for Reactive Collision Avoidance

Abstract

In this extended abstract, we introduce our previously published work [1]. Semi-autonomous capabilities in Micro Aerial Vehicles (MAVs) are crucial for assisting teleoperation and avoiding collisions. We present an approach where operators provide a simple speed and direction command to the MAV. The drone executes the instruction by exploiting a Deep Reinforcement Learning (DRL) model that processes RGB images and the current robot position to follow the command while avoiding collisions. We demonstrate the effectiveness of our approach in simulated environments and compare it against a state-of-the-art baseline.