This paper introduces EmbodiedAgent, a hierarchical framework for heterogeneous multi-robot control. To address the hallucination problem arising from unrealistic tasks, EmbodiedAgent integrates a next-action prediction paradigm and a structured memory system to decompose tasks into executable robot actions and dynamically validate actions based on environmental constraints. Furthermore, we present the MultiPlan+ dataset, which contains over 18,000 annotated planning instances across 100 scenarios, including a subset of unrealistic cases to mitigate the hallucination problem. To evaluate performance, we propose the Robot Planning Assessment Schema (RPAS), which combines automated metrics with LLM-assisted expert evaluation. Experimental results demonstrate that EmbodiedAgent outperforms state-of-the-art models, achieving an RPAS score of 71.85%. Real-world validation on an office service task highlights the ability of EmbodiedAgent to coordinate heterogeneous robots toward long-term goals.
