Advances in Artificial Intelligence for the Underwater Domain

Where most so-called autonomous underwater vehicles are automated, they have limited autonomy in the traditional sense of being free to choose new courses of action. Onboard software systems have a narrow range of sensor processing and mission replanning capabilities due to the expense and the complexity of these operations, as well as a lack of operator trust in the validity of plans that operators cannot inspect before execution. The advent of machine learning and the maturation of new ideas in automated planning will reduce the cost and difficulty of fielding more autonomous systems, attacking the first difficulty. The second difficulty-that of trust-is severe for underwater robotics, where the more traditional approach of intelligent systems acting as operator aids is made difficult by the physics. Increasing trust will require technical advances in assured autonomy that are only just beginning. We survey the history of artificial intelligence (AI) for robotics, highlighting important developments that will influence future systems.