VECTOR: Decentralized Task Planning with Verification and Adaptive Execution for Multi-Robot Systems

Decentralized multi-robot systems must coordinate robots with heterogeneous capabilities, manage action dependencies, and remain robust to unexpected interruptions, all of which are critical to reliable task completion. However, existing frameworks often overlook robots’ physical and task constraints, execution verification, and coordinated replanning, leaving these challenges largely unaddressed. To address this problem, we propose VECTOR, a decentralized framework that combines constraint-aware peer-to-peer task negotiation, symbolic plan grounding, dependency-based parallel scheduling, and plan verification. Before execution, VECTOR verifies that the combined multi-robot plan will achieve the global task goal. During execution, the system monitors the shared symbolic state and responds to environmental changes through targeted recovery strategies such as motion-level adjustment or decentralized replanning. We evaluate VECTOR on two collaborative tabletop manipulation tasks under four interruption conditions in both simulation and real robot experiments. Our results show that VECTOR achieves an average task success rate of 91%, significantly outperforming existing vision-language planning approaches while maintaining low coordination overhead. These results highlight that combining decentralized coordination with symbolic verification and adaptive execution enables more reliable multi-robot task completion in dynamic environments.