Autonomous Behavior of Biped Robot by Learning Camera Images

COVID-19 infection has been prevalent worldwide from the end of 2019, and people have spent more time at home avoiding contact with others. In such an environment, it is expected that avatars with physicality will act in the real world. This makes it possible to promote social life and economic activities through the avatars of each person, even if humans do not engage in face-to-face activities. In such a world, it is desirable for humans to remotely control avatar robots, and for robots to learn from those logs and behave autonomously. Therefore, in this paper, we investigated a method for robots to learn from human operation logs and continue autonomous behavior without colliding with the environment. The method proposed in this paper is divided into the following three steps. First, a human operates a bipedal walking robot via a GUI, and records the image of the robot-mounted camera and the selection behavior at that time. Next, machine learning is performed from the recorded images and selection behavior. Finally, the robot is autonomously behaved using the learned results. As a result, it was confirmed that the robot continues to behave autonomously for a relatively long time without colliding with the environment.