Systematic Multi-Agent Vision-and-Language Navigation: Formulation, Benchmark, and Method
The paper turns multi-robot language navigation into a constrained coordination benchmark with 11,724 episodes.
The task setup, called MAVLN, covers 145 scenes and teams of up to four agents. Its missions include dependency and resource constraints, including presence locks and holding chains. The authors also introduce TRISS, a baseline system with LLM-based subtask scheduling, shared topological memory, and conflict-aware execution for collision-free routes. Experiments show the baseline still leaves major gaps in scheduling, planning, and execution.
HF Daily Papers' note
The task setup, called MAVLN, covers 145 scenes and teams of up to four agents. Its missions include dependency and resource constraints, including presence locks and holding chains. The authors also introduce TRISS, a baseline system with LLM-based subtask scheduling, shared topological memory, and conflict-aware execution for collision-free routes. Experiments show the baseline still leaves major gaps in scheduling, planning, and execution.
HF Daily Papers' note
score 5