Abstract
In recent years, robots have served conspicuously in dangerous environments including disaster areas and outer space. Within such environments, however, robots suddenly may fall into dangerous situations when commands from people to avoid certain risks don't reach them in time. Accordingly acquisition of autonomous risk-avoidance behavior in robots is required. It is thought that the technique for realizing this capability will make use of reinforcement learning. As a new reinforcement-learning framework for avoiding risk, we propose a probability-based reinforcement learning method and apply it to behavior acquisition in robots.
