A Goal-Oriented Agentic Framework For Collaborative Branching Human-Robot Interaction

As robots and technology grow more advanced, teaching humans how to operate them becomes increasingly complex and difficult to master. This complexity often requires longer explanations, yet extended instructional periods can make it harder for users to maintain focus and avoid cognitive fatigue. To address this, we designed and built an agentic AI system for conveying complex, branching information through goal-oriented interaction. The framework consists of three agents, Strategic, Conversational, and Persistent, with separate responsibilities for planning, communication, and state management. The system is driven by smaller language models capable of running on local hardware, supporting data privacy and reducing reliance on centralized processing. To test the system, we built an online collaborative puzzle game that required participants (N=70) to interact with the agentic AI to solve conditionally branched tasks. Data were collected through pre- and post-interaction questionnaires using 7-point Likert scale items derived from validated constructs, including NARS, trust, and competence. Results demonstrated a 100\% completion rate, with high perceived trust and social presence above the neutral midpoint (d>1.2). A supplementary RAG baseline study showed comparable interaction length, but significantly lower overall subjective interaction quality than the agentic condition (M=4.18 vs. M=5.51), Welch's t(37.88)=4.22, p<.001, Hedges' g=1.04. These findings suggest that the benefit of the agentic framework lies primarily in interaction quality rather than conversational efficiency. Furthermore, the agentic architecture displayed robustness by recovering from non-normative inputs while maintaining strict goal alignment.