抄録
Today's outdoor autonomous robot is not so much having ability to cope with unknown terrain. Giving all pre-knowledge on the target terrain is not realistic, and adaptive control is needed. However, the learning approach usually takes a long time to get knowledge about the environment as the precision of sensory acquired states increases. This paper proposes a method to get terrain information by classifying sensory data by its features and learning to move by classification. The main point is that the classification is done by the distance of learning results and the system repeats learning and classifying, which gives practically feasible adaptive learning rules.
