Development and validation of a safe reinforcement learning drone controller

Abstract

This paper presents the work in progress that aims to develop, validate, and verify the use of Reinforcement Learning~(RL) and neural networks for safety real-time systems. A \gls{UAV} controller is utilized as a use case. These techniques are particularly promising for autonomous control as they learn from a dynamic environment without human intervention. The proposed solution shows the behavior of a UAV that has been trained to maintain altitude and avoid incoming obstacles such as other UAVs. The AirSim simulator was used to simulate a realistic flight scenario with two UAVs. One of the vehicles contains the RL controller, while the other is constantly attempting to collide with it. Preliminary results show that neural networks trained with the Soft Actor-Critic (SAC) algorithm can avoid collisions, especially in cases unseen during training. Despite the satisfactory results, further research is needed to verify that the RL agent can operate correctly in a safe and real-time environment.