SynerMind: Towards a Long-Horizon Multi-Agent System for Embodied Collaboration

Multi-robot collaboration systems are essential for complex robotic tasks. While existing systems can coordinate multiple robots in short-horizon scenes, their predefined static task allocation makes them difficult to adapt to dynamic long-horizon execution. Addressing this challenge, we introduce **SynerMind**, a unified system for long-horizon collaboration among heterogeneous robots. It introduces **SynerMind-Brain**, a collaborative agentic model. The brain maintains global task awareness throughout dynamic long-horizon task execution by continuously aligning evolving task progress with heterogeneous robot capabilities. Based on this, the multi-agent framework **SynerMind-Flow** organizes perception, planning, execution, and reflection into a closed-loop collaboration process. Meanwhile, we introduce **Collaborative-State Mining Group Relative Policy Optimization** to mine difficult collaborative states from group reward statistics. It prioritizes unresolved coordination failures during policy optimization, thereby improving long-horizon collaboration. Notably, we construct ***SynerAwareness-20$k$*** and ***SynerAgentic-102$k$***, two specialized real-world datasets for long-horizon heterogeneous collaboration. Extensive experiments demonstrate that **SynerMind** exhibits excellent performance in simulation and achieves a **1.53$\times$** higher success rate than existing methods in dynamic real-world environments.