An adapted subsumption architecture based on learning to improve autonomous water navigation techniques
Autonomous Navigation, Autonomous Sailboats, Adapted Subsumption Architecture, Machine Learning
Autonomous sailing navigation faces a different paradigm when searching for better efficiency in route planning. Unlike land navigation, where the best path is a result of the shortest route, with less traffic or obstacles, in the water other factors harden this dynamic. It is still unknown for the aquatic robotic's community the development of an efficient navigation technique, involving high-level control for such vehicles. Then, the biggest issue presented in this work is how to program and choose a control and behavior architecture for completely autonomous water vehicles. Nevertheless, also being effective and easy to use. Thus, we established as an objective to improve autonomous aquatic navigation techniques for robotics vessels through the use of subsumption architecture, combined with deep machine learning techniques. Hence, the great control problem is divided into several tasks that are isolated in behaviors, to be solved independently by machine learning. The big difference between the control algorithms usually adopted and of this proposal is based on the use of a reactive model, resulting in quick decision making - essential to autonomous aquatic navigation- combined with a machine-learning technique, allows a pursuit for the best parameters of behaviors and control architecture. Therefore, the hypothesis raised in this work is that through a behavioral architecture control dynamically adaptable to the environment, enables obtaining better results consistently than when in static architectures.