Abstract

This paper presents a new method of learning visual docking skills for non-holonomic vehicles by direct interaction with the environment. The method is based on a reinforcement algorithm, which speeds up Q-learning by applying memorybased sweeping and enforcing the “adjoining property”, a filtering mechanism to only allow transitions between states that satisfy a fixed distance. The method overcomes some limitations of reinforcement learning techniques when they are employed in applications with continuous non-linear systems, such as car-like vehicles. In particular, a good approximation to the optimal
behaviour is obtained by a small look-up table. The algorithm is tested within an image-based visual servoing framework on a docking task. The training time was less than 1 hour on the real vehicle. In experiments, we show the satisfactory performance of the algorithm.