Communication Interface for Human-Robot Partnership

摘要

This paper deals with learning for a person/robot/computer agent partnership communication interface. Although it is difficult for robots to learn behavior through interaction with people based on human intention in an actual environment, the robot easily obtains environmental information using sensors. Learning in computer simulation is relatively easy because contact patterns are restricted in the virtual environment, but the computer agent cannot collect environmental information on people. Robot and computer agents thus play different roles. Interface design is vital to computer agent, because people is interfaced to the computer’s virtual environment. Human intention should be extracted through communication with the computer agent in the virtual environment. In this study, we consider interaction between a robot and a person through a computer agent, and the task given to the person is to guide the robot to a target point based on human intention. For this, we use a computer agent, assuming it gets energy at a specific point in the virtual environment. We propose a method for extracting human intentions using multiple state-value functions. A state-value function is selected based on a human tapping pattern on the PDA used as an interface to the computer agent, and is updated by a reinforcement learning algorithm based on a reward. Experimental results demonstrate the effectiveness of the proposed method.