Deep Reinforcement Learning for Trajectory Generation and Optimisation of UAVs

In recent years, the rapid advancements in Machine Learning have led to substantial research in control systems for autonomous aerial vehicles. Particularly, Reinforcement Learning (RL) has attracted a lot of attention for the design and development of such control algorithms. This paper examines the control issues of autonomous flight and how these are addressed using RL approaches. The objective is to investigate how RL algorithms like Deep Deterministic Policy Gradient may be applied particularly for control actions in an unmanned aerial vehicle (UAV). This learning paradigm acts as a mechanism that continuously generates policies for tasks such as attitude and position control, which converges into an optimized trajectory. As an outlook, the application of Continual Reinforcement Learning is proposed. This is a novel RL methodology that holds the potential to advance the control system of a UAV operating in dynamic, unknown environments with the ability to reapply learnt behavior and flexibly adapt to new situations.