Auto-Regressive Model Unifies Multi-Agent Task Allocation and Routing
A new AI framework, ARMATA (Auto-Regressive Multi-Agent Task Assignment), jointly optimizes area allocation and visitation routing for multi-agent systems over spatially distributed areas. Proposed in arXiv:2605.04225, ARMATA uses a centralized, fully end-to-end auto-regressive decoder that generates allocation decisions and routing sequences in a single pass, maintaining a global state. This approach avoids local optima common in decentralized heuristics and outperforms existing decoupled methods in experiments.
Key facts
- ARMATA stands for Auto-Regressive Multi-Agent Task Assignment.
- The framework is centralized and end-to-end auto-regressive.
- It jointly generates allocation decisions and routing sequences.
- A multi-stage decoding mechanism unifies high-level allocation and low-level routing.
- The model maintains a centralized global state.
- It implicitly balances workload distribution with routing efficiency.
- It avoids local optima common in decentralized methods.
- Extensive experiments show significant performance improvements.
Entities
Institutions
- arXiv