Systematic Multi-Agent Vision-and-Language Navigation: Formulation, Benchmark, and Method

AI Digest

本文提出多智能体视觉-语言导航（MAVLN）的系统化方法，通过形式化任务约束、构建基准测试并设计TRISS系统，揭示了多智能体协作的挑战。

This paper introduces a systematic approach to multi-agent vision-and-language navigation (MAVLN) via task formalization, benchmark creation, and TRISS system design, highlighting coordination challenges.

Key points

• 首次将多智能体VLN形式化为带约束的协调问题，包含子任务依赖与资源约束。 Formalizes multi-agent VLN as constrained coordination with subtask dependencies and resource constraints.
• TRISS系统整合LLM调度器、共享拓扑记忆与冲突消解机制实现协同导航。 TRISS integrates LLM scheduler, shared topological memory, and conflict resolution for collaborative navigation.
• 实验验证TRISS作为基线，揭示调度/规划/执行环节的显著优化空间。 Experiments show TRISS as a baseline with substantial improvement potential in scheduling/planning/execution.

Takeaway: 系统化MAVLN框架为多智能体协作提供新范式，揭示实际部署的关键技术瓶颈。 / The systematic MAVLN framework provides new paradigms for multi-agent collaboration, revealing key technical bottlenecks for real-world deployment.

Why it matters 为多智能体协作提供新框架，对机器人集群应用有直接指导价值。