A Framework of Autonomous Knowledge Transfer for Robot Navigation Task

This paper proposes a framework of knowledge transfer used in the robot navigation problem, in order to speedup the learning process when the robot meets new or more complex task. Take the graph of source navigation maps as the template library, Principal Component Analysis method is adopted to match the target graph with source template. Then, the robot recognizes the similar source task. The value-function of the selected source task, which correlates the cumulative rewards with the sensory information corresponding to the generalized distance information to each beacon, is transferred to the target task. With this approach, the robot can choose the most related knowledge from the source tasks and transfer the generalized knowledge to the target navigation task automatically without handed code mapping between tasks. Experimental results demonstrate that the presented transfer method can yield a remarkable speedup for learning process in the robot navigation task.