抄録
A method for acquiring an internal representation of a vision-based agent by associating visual input with a tactile sensor is presented. Agents that act in the physical world always suffer from the constraints of embodiment. However, embodiment plays a very important role in the formation of visual functions. Cognitive psychological findings assert that vision does not work without tactile motion experiences. Thus, our method features a tactile sensor for acquiring an internal representation. An internal representation corresponding to the physical relationship between the agent's body and environment in the visual input is acquired as a utility function. The utility value is calculated by a dynamic programming method with rewards given in a tactile state and transition probability model obtained beforehand. We conducted two computer simulation experiments, one being obstacle avoidance in which the agent needs to determine whether it can pass through a gap between objects or not. The other is object recognition in which the agent can distinguish differences in object shape by tactile motion.
