Humans interact with other people in their daily lives. The aim of this research is to realize a robot coexisting with humans and capable of smooth human-robot interaction. In an actual environment, we face various situations and human behavior is diverse. Since the rule to interact with human is complicated and ambiguous, it is difficult to implement an explicit rule on robot. In this research, we bring a robot in an actual environment and conduct experiments. We are developing a robot that autonomously learns the rule for human-robot interaction by a reinforcement learning algorithm using a deep neural network through its own trial-and-error.