Two-Level Reinforcement Learning Framework for Self-Sustained Personal Robots

Abstract

As social robots become integral to daily life, effective battery management and personalized user interactions are crucial. We employed Q-learning with the Miro-E robot for balancing self-sustained energy management and personalized user engagement. Based on our approach, we anticipate that the robot will learn when to approach the charging dock and adapt interactions according to individual user preferences. For energy management, the robot underwent iterative training in a simulated environment, where it could opt to either “play” or “go to the charging dock”. The robot also adapts its interaction style to a specific individual, learning which of three actions would be preferred based on feedback it would receive during real-world human-robot interactions. From an initial analysis, we identified a specific point at which the Q values are inverted, indicating the robot’s potential establishment of a battery threshold that triggers its decision to head to the charging dock in the energy management scenario. Moreover, by monitoring the probability of the robot selecting specific behaviours during human-robot interactions over time, we expect to gather evidence that the robot can successfully tailor its interactions to individual users in the realm of personalized engagement.