Summary
A PRISMA-compliant systematic review of 86 articles examining how LLMs are transforming Human-Robot Interaction (HRI). The paper argues that existing technical reviews overlook human-centered considerations — user modeling, appropriate autonomy, and human-oriented understanding — and provides consolidated guidelines for future LLM-driven HRI research.
對 86 篇文章進行 PRISMA 系統性文獻回顧，探討 LLM 如何轉型人機互動研究。論文指出現有技術性回顧忽略了以人為本的考量（用戶建模、適當自主性、人本理解），並為未來 LLM 驅動的 HRI 研究提供整合性指導方針。
Key Points
- LLMs enable contextual sensing: robots infer situational context from language, not just sensors
- LLMs enable socially grounded interaction: natural dialogue, intent inference, common-sense reasoning
- Research landscape is highly fragmented — inconsistent experimental setups and evaluation metrics across studies
- Critical gap: human-centered aspects (user modeling, levels of autonomy) are underrepresented vs. purely technical papers
- Human-oriented understanding is the weakest link in current LLM-HRI systems
Insights
- The fragmentation finding mirrors the ICLR 2026 VLA survey: rapid field growth creates noise, and lack of standardized evaluation makes cross-paper comparison nearly meaningless
- The emphasis on “levels of autonomy” is significant: LLM capabilities create a new design space where robots can be more or less autonomous depending on context — but researchers aren’t explicitly studying this dimension yet
- “Socially grounded interaction” is what distinguishes HRI from regular robotics: the goal isn’t just task completion but interaction that humans find natural and appropriate
- The CHI venue (not a robotics conference) signals that HRI is increasingly a human factors and design problem, not just an engineering one
Connections
- GR-Dexter: VLA for Bimanual Dexterous Robot Control
- State of VLA Research at ICLR 2026
- Embodied AI
- Vision-Language-Action Models
- Natural Language Processing
- Human-Robot Interaction
Raw Excerpt
LLMs fundamentally reshape how robots perceive context and generate socially appropriate responses. Current research remains exploratory with inconsistent experimental setups, methodologies, and evaluation metrics across studies.