Learning Legged Swimming Gaits from Experience

We are interested in making underwater robots agile and adaptable to challenging ocean conditions. It is impractical for humans to engineer controllers for all desired swimming motions and for all circumstances that may arise during operation, such as damage to flippers or motors. Therefore, we propose to use modern reinforcement learning techniques to learn swimming gaits, either from scratch or from easy-to-obtain prior information such as a realistic simulator. In this way, new motions only require the robot to swim and perform learning computations, saving time for human engineers, and achieving new and exciting motions for the robot.