Learning more from limited demonstrations: methods for efficient and informative human-robot interaction

Abstract

Learning from Demonstration (LfD) offers a promising paradigm for enabling socially assistive robots (SARs) to acquire complex skills and social behaviors by observing human demonstrations. However, conventional LfD methods rely heavily on large-scale, high-quality demonstration datasets, which are difficult to obtain in realworld healthcare and education settings due to privacy, cost, and data scarcity. This dissertation addresses these challenges by proposing a series of approaches to enhance learning efficiency, improve adaptability, and maximize information extraction from limited demonstration data in human-robot interaction (HRI) scenarios.First, a hierarchical deep reinforcement learning framework is introduced that incorporates auxiliary classifier generative adversarial networks (ACGAN), dynamic experience replay strategies, and Deep Q-learning Networks (DQN) to improve learning performance without additional data. Second, a task-oriented Meta-inverse reinforcement learning (Meta-IRL) approach is proposed to enhance adaptation to new tasks by leveraging encoders and exploring transformer-based multi-head and layer feature extraction strategies. Finally, a novel framework integrating a Global Attention Mechanism (GAM) with multi-layer feature fusion and latent Dirichlet allocation (LDA) topic modeling is developed to enrich feature representations and optimize prompt generation in few-shot learning. Experimental validation on robot-mediated therapy tasks and other datasets demonstrates that the proposed methods enhance performance under limited data conditions. Overall, this work integrates efficient learning mechanisms with personalized intervention strategies, enabling the model to acquire richer and more informative representations that enhance its performance, while contributing to the broader goal of facilitating effective deployment of intelligent SARs in data-constrained, real-world environments.