Intelligent Algorithms for UAV Automatic Landing on-board a Moving Platform

Abstract

Reinforcement learning based control algorithms require large amount of training. During training the agent explore many possible safe and unsafe states. In the simulation environment exploration of unsafe states can be affordable. But many times, training in simulation world is not a feasible option because of issues of modeling of real-world scenarios. In such scenarios direct real-world training seems to be feasible solution. In such real-world training exploration of unsafe state is very dangerous for the agent as well as for the environment. In this work, safe learning is carried out for multi agent deep deterministic policy gradient (MADDPG) algorithm with the help of control barrier functions (CBF). This MADDPG augmented with CBF is applied for an application of long duration autonomy. Battery driven multiple ground robots are deployed for surveillance of a given region with fix charging stations. The robots are trained for surveillance considering safe learning i.e. robots will not drain out of battery during training and testing for surveillance.