Adaptive Behavior Generation of Social Robots Based on User Behavior Recognition

Abstract

For natural human-robot interaction, social robots should understand a user behavior and respond appropriately. In particular, when generating a behavior to interact with the user, it is important to adapt its behavior to the user’s posture and position rather than repeating the predefined motion. To this end, we propose a method for generating the robot behavior in three steps, i.e. user behavior recognition, robot behavior selection, and robot behavior adaptation. First, the user behavior is recognized by using a Kinect v.2 sensor and a long short-term memory-based neural network model. The weights of the model are trained using the AIR-Act2Act, which is a human-human interaction dataset. Then, according to the behavior selection rules designed by referring to the interaction scenarios in the dataset, the robot selects an appropriate behavior for the recognized user behavior. Finally, the key pose of the selected behavior is modified in consideration of the user’s posture and position. To demonstrate the feasibility of the proposed method, experiments were conducted using a Pepper robot in a 3D virtual environment. The experimental results showed that the proposed method has an accuracy of 99% in recognizing the user behavior, and the robot behavior can be modified naturally even if the user’s intention is misunderstood at first.