Embodied Intelligence in Multi-Robot Systems: Foundations, Applications, and Future Directions

Embodied intelligence in multi-robot systems (MRS) represents the collective ability of multiple physically situated agents to perceive their surroundings, reason about shared goals, and act cooperatively in dynamic environments. In contrast to single-robot autonomy, MRS embodiment highlights mutual perception, adaptive coordination, and interactive decisionmaking within a physically coupled setting. Achieving such integrated intelligence remains difficult due to non-stationary interactions, limited communication bandwidth, and heterogeneous sensing and actuation capabilities. Over time, research on MRS has evolved from explicitly designed coordination rules and optimization-based formulations to data-driven learning and foundation-model paradigms, which have enhanced adaptability and semantic understanding but also led to fragmentation across separate methodological streams. This survey presents a structured and comprehensive analysis of embodied intelligence in multi-robot systems, organized around three major perspectives: foundations, applications, and future directions. The foundational part introduces a unified taxonomy that characterizes multirobot embodiment in terms of task structure, agent configuration, control architecture, and coordination mechanism, and reviews four principal methodological paradigms: rule-based, optimization-based, learning-based, and LLM-based approaches. The applications part summarizes available benchmarks and evaluation settings and compares representative studies across navigation and exploration, manipulation, perception, and task allocation and planning. The final part discusses open challenges and outlines future research opportunities focused on scalability, adaptability, interpretability, and trustworthiness. Together, this survey provides a consolidated view and forward-looking roadmap for advancing embodied intelligence in MRS.