Exploring Deep Reinforcement Learning for Autonomous Powerline Tracking

Panin Pienroj, Sandro Schönborn, Robert Birke

Abstract: Deep reinforcement learning (DRL) has recently received increasing attention due to unprecedented ability achieved in playing games. Since then DRL has been successfully applied across different domains. Here we explore the potential of DRL in autonomous power transmission lines inspection. Regular inspection is essential to ensure an optimal efficiency void of hot spots or corona discharges, but challenging due to the extensive geographic scope of the electric network often crossing areas difficult to reach. Traditional inspection is done via helicopters equipped with infrared and ultraviolet cameras. This solution however is costly and dangerous for the crew making solutions based on unmanned aerial vehicles (UAV) an attractive alternative. In this work we use realtime simulations based on AirSim to test several state-of-the-art DRL algorithms in controlling the flight of a quadrotor. The agent is trained on simple and complicated tracks under different landscapes and ambient conditions. Elements from imitation learning are used to speed up training. Visual disturbance is added to the drone in order to test the robustness of the learned policies as well as their adaptivity to different training conditions.