Development of Intelligent Control Strategy for an Unconventional UAV: A Novel Approach

Use of intelligent systems have marked a major breakthrough in the domain of aerial plat- forms. Under the ambit of Machine Learning (ML), Reinforcement Learning (RL) has started to address the inherent limitations of the conventional controllers and emerges as the most ac- tive, conceptually prudent and best suited machine learning category for autonomous control. In this paper paper, we present two innovating RL based control strategies namely ‘Reward Selective Dynamic Programming (RSDP)’ and ’Optimal Deep Deterministic Policy Gradient (O-DDPG)’ to control the dynamics of an unconventional UAV. Both the framework have been developed specifically to accommodate the continuous state and action domains of the aerial platform. Both the RL algorithms turned out to have satisfactory computational performance, and the agent was effectively trained for entire state and action space. The effectiveness of both the proposed strategies were verified through extensive 6-DOF simulations. Results indicate that both outperforms the classical control architecture primarily by eliminating the explicit requirement of gain scheduling for various equilibria during the trajectory. A perspective analysis was also carried out to perform a comparative analysis between both the proposed RL strategies.