An Approach to Modeling and Studying the Behavior of Firefighting Drones Using Unity ML-Agents

This paper presents an approach to model and study the behavior of a firefighting drone. Unity ML-Agents Toolkit and two advanced Deep Reinforcement Learning algorithms are used. They are respectively Soft Actor-Critic (SAC) and Proximal Policy Optimization (PPO) which are supported. Their effectiveness in training a virtual agent drone to put out a fire is being studied. For the experiment, a fire is realized at different places randomly in the training virtual environment. The results show that the agents can detect and extinguish many of the fires and achieve adaptive behavior. From this, it can be concluded that these algorithms can be used for real purposes and the Unity ML Agent toolkit is suitable for conducting experiments.