Systematic Multi-Agent Vision-and-Language Navigation: Formulation, Benchmark, and Method
Paper formalizes multi-agent vision-language navigation and releases the MAVLN benchmark plus TRISS.
The paper formalizes multi-agent vision-and-language navigation as coordination under dependency and resource constraints, including presence locks and holding chains. Its MAVLN benchmark contains 11,724 episodes across 145 scenes, with teams of up to four agents and three instruction regimes. The authors introduce TRISS, combining an LLM subtask scheduler, shared topological memory, and conflict-aware collision-free execution. Experiments position TRISS as a baseline and report substantial remaining gaps in scheduling, planning, and execution.
- MAVLN formalizes multi-agent VLN as constrained coordination.
- The benchmark has 11,724 episodes across 145 scenes.
- Teams of up to four agents face three instruction regimes.
- TRISS pairs an LLM scheduler with shared memory and conflict-aware routing.
- Experiments show large gaps in scheduling, planning, and execution.
From the original article
1 min read at huggingface.co
Vision-and-Language Navigation (VLN) has largely focused on a single agent following a single instruction, yet many real-world applications require teams of robots to tackle tasks beyond the capabilities of any individual agent.
By Yunzhe Xu, Zhe Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.