Planificación óptima de movimiento y aprendizaje por refuerzo en vehículos móviles autónomos

Abstract

This research work is focus on the proposal of an algorithm able to perform an optimal motion planning in two-wheel-drive (front or rear) and front steering autonomous mobile vehicles. The algorithm is based on optimal control techniques in closed loop applied to four-wheel vehicles. These vehicles are non-linear dynamic systems in which the motion planning and its control are difficult tasks due to they are non-holonomic systems. All studies done have been developed taking into account the reinforcement learning and cell state space concepts. Possible changes in the environment or in the physical or mechanical structure of the vehicle do not affect the new algorithm. The optimal controller will be generated when the learning stage has been previously performed and therefore, these changes have been implicitly taken into account. The learning is oriented to reach a specific goal from each origin, according to an optimization criterion (e.g. minimum time). The vehicle learns its kinematics and dynamics from its own experience. There is no need to have any kind of mathematical model of the system. Also, the algorithm is able to extend the local knowledge acquired in a specific zone of the state space to the rest of the space without the vehicle moving physically to those zones. Once the learning stage finishes, the planning is performed in closed loop applying the optimal control actions associated to the state of the vehicle in real time.