Siamese robot for multi-human robot interaction with developmental episodic memory-based mechanism of thought

Abstract

To use a service robot in a crowded area, the robot should be able to interact with multiple people at the same time. However, the existing robots with one head have a limitation in interaction with a crowd since they cannot recognize a person behind them while interacting with another person in front of them. Moreover, the existing behavior selection methods have difficulties in learning deeply or in creating new behaviors. To solve these problems, we propose a novel robot platform, i.e. Siamese robot, and its behavior selection method using the developmental episodic memory-based mechanism of thought (DEM-MoT). The proposed robot platform consists of two heads, two arms, and a body to cover all directions. Each head is controlled by an intelligent software agent that selects its own behaviors by DEM-MoT and the authority to control the shared limbs is distributed to one of the individuals according to the strengths of their current internal states and contexts. DEM-MoT is a human-like thought process that selects a proper behavior based on the past experiences stored in developmental episodic memory (DEM). If the expected reward of the selected behavior is too low, an experience is transferred from the other individual or a new behavior is created by randomly arranging a set of primitive behaviors. To develop intelligence in an adaptive and unsupervised way, DEM is implemented with an adaptive resonance theory-based algorithm and its weights are pre-trained using the simulation experiences produced by the degree of consideration-based mechanism of thought. To show the effectiveness of the proposed mechanism of thought for behavior selection, experiments were carried out with a physical Siamese robot in a real environment.