Comparison of Reinforcement and Imitation Learning algorithms in autonomous sailboat Digital Twins

This project aims to study the performance of two reinforcement machine learning algorithms, namely the Proximal Policy Optimization and Soft Actor Critic, in the simulation of autonomous sailboats and their response to different wind directions while avoiding obstacles detected by image analysis and following defined target check-points. Also, the effect of the imitation learning algorithms Behavioral Cloning and Generative Adversarial Imitation Learning combined with the first mentioned algorithms is studied. The proposed scenarios consist of areas filled with random static or moving obstacles and with the presence of favorable or crosswinds. The motivation for the project comes from the lack of studies of the mentioned algorithms in autonomous sailboats, issue which the current study tries to address. The Unity platform and ML-Agents machine learning toolkit are used for development and the methodology that guides the project can be similarly applied to other reinforcement learning problems. Through agent training, it is possible to compare the results and observe that the Proximal Policy Optimization obtains better performance within the proposed scenarios, both with and without the support of imitation learning algorithms.