Vision-Based Path Learning for Home Robots

Our purpose is to develop a robot which can find and learn paths between important places in a home environment based only on vision. Reinforcement Learning (RL) is suitable for such tasks because it can be low-cost and robust to the sensor and actuator noises. A method which uses a Bag-of-Visual-Words representation for each image as an input to RL has been proposed to solve this problem. We think that an RL method which is more strongly oriented to exploitation would work better with this input space. This paper proposes a new vision-based path learning method for home environments. For this purpose, we develop a strongly exploitation-oriented RL method. We introduce the concept of the value of a state so that unnecessary states can be reduced quickly. We have verified the effectiveness of our method by simple simulation experiments.