Improving Energy Efficiency in UAV Attitude Control using Deep Reinforcement Learning
Varun Agarwal, Rajiv Ranjan Tewari
Abstract
Attitude control refers to controlling the rotational motion of an Unmanned Aerial Vehicle (UAV) about its axes. Any movement requires energy consumption and UAV batteries store limited energy. We propose the use of reinforcement learning to optimise energy usage in UAVs. We use Proximal Policy Optimization (PPO) algorithm to train the model, modifying the existing algorithm to incorporate sigmoid activation function. We have introduced Ornstein Uhlenbeck noise to the policy function with the intention of adding the unpredictability found in real world environments. We have designed the reward function of our algorithm such that the quadcopter aims to change its existing angular velocity to achieve the target angular velocity. While this attitude control takes place, energy is spent due to motor/propeller revolutions. Our reward function minimizes the rapid change in motor speeds which causes fast battery depletion, thus saving energy and enhancing UAV flight time.