Personalised Interactive Reinforcement Learning with Multi-Task Pre-training

Personalised robots have immense potential to enhance daily life through tailored interactions, yet achieving efficient personalisation remains challenging. This paper introduces a Multi-task Interactive Reinforcement Learning (MIRL) framework aimed at improving the efficiency of interactive learning with evaluative feedback. We demonstrate that pre-training the robot across diverse tasks significantly reduces the learning steps required during fine-tuning, thereby enhancing sample efficiency. Our approach effectively aligns robot behaviours with user preferences, as evidenced by experimental results. These advancements promise to advance the usability and effectiveness of personalised robotics in diverse applications.