Fly-Crash-Recover: A Sensor-based Reactive Framework for Online Collision Recovery of UAVs

Abstract

Unmanned Aerial Vehicles (UAVs) are becoming increasingly popular thanks to the multiplicity of operations in which they can be deployed such as surveillance, search and rescue, mapping, transportation, hobby and recreational activities. Although sensors like LIDARs and cameras are often present on such systems for motion planning to avoid obstacles, collisions can still occur in very dense and unstructured environments, especially if disturbances are present. In this work, we research techniques to recover UAVs after a collision has occurred. We note that the on-board sensors, especially the inertial sensor used to stabilize the UAV, run at a high frequencies obtaining hundreds of data points every second. At run-time, this can be leveraged at the moment of a collision to quickly detect and recover the system. Our approach considers knowledge of UAV system dynamics to predict the expected behavior of the vehicle under safe flight conditions and leverage such expectations together with inertial data to detect collisions rapidly (on the order of milliseconds). We also propose a potential field-based approach to map the collision and create the correct reactive maneuver to avoid the collided object and bring the system back to a stable and safe configuration. Experiments are executed using ROS on two micro-quadrotor UAV platforms having different dynamics and performances, while colliding with poles and walls positioned in different configurations. In our results, we are able to show that the UAVs are successfully able to detect and avoid a collision, while also providing a rigorous analysis of the conditions in which the system can recover from imminent collisions.