Cultivating Expressivity and Communication in Robotic Objects: An Exploration into Adaptive Human-Robot Interaction

This work introduces a model to personalize ‘robjects’ - everyday objects with embedded robotic capabilities - using deep reinforcement learning (RL). The method creates user mood maps from emotional states and environmental conditions, allowing the robot to convey its status via light and movement. Two RL agents, with distinct reward systems, are deployed: The motor agent aligns the robot’s expression with the human emotion, while the light agent correlates with the robot’s expression adherence to the Pleasure-Arousal-Dominance (PAD) scale. The learning process establishes a base model for real-world use, starting with synthetic interactions in simulation. This model underwent a 10-day experiment aimed at assessing the model’s reception and user experience. The model enabled continuous interaction; all participants reported a pleasant experience, expressing the desire to continue interacting with it. The robot’s adaptivity is highlighted by a mean Jensen-Shannon distance of 0.334 between user emotion and the robot expression distributions, showcasing a responsive, adaptive model.