Evaluating Efficiency and Engagement in Scripted and LLM-Enhanced Human-Robot Interactions