Causal MDP for Reminiscence Therapy using LLMs for Pepper Robot (RA-L 2024)

Developed a Causal MDP Framework: Designed a decision-making model leveraging Causal Markov Decision Processes to structure therapy sessions for effective human-robot interaction in RT. Integrated Large Language Models (LLMs): Benchmarked and implemented models like GPT-4 and Llama to enhance conversational capabilities, enabling the Pepper Robot to engage naturally with dementia patients. Optimized Conversational Engagement: Improved therapy session effectiveness by achieving a 14% increase in conversational engagement through tailored dialog strategies driven by LLMs and causal inference. Enhanced Q-Value Convergence: Achieved 3x faster Q-value convergence in the reinforcement learning framework, enabling more efficient policy updates and real-time decision-making. Multimodal Integration: Incorporated vision-based features and contextual understanding using tools like Vision Transformers (ViT) to enrich the robot’s perceptual and interactive capabilities. Real-Time Adaptation: Enabled the Pepper Robot to dynamically adjust therapy interactions based on patient responses, supported by the causal MDP and reinforcement learning models. Human-Centered Design: Focused on designing a system that prioritizes empathy, responsiveness, and patient comfort in reminiscence therapy.