CoNavBench: Collaborative Long-Horizon Vision-Language Navigation Benchmark

Abstract

CoNavBench evaluates how multiple robots cooperate to complete long-horizon vision-language navigation tasks. The benchmark contains 4,048 single-robot and collaborative episodes, with graph annotations and collaboration categories covering object handoffs and rendezvous. Its NavCraft platform generates tasks through a hierarchical process: it creates a primary robot’s mission, assigns supporting robots and subgoals, and validates the resulting coordination plan. Habitat-Sim scene graphs support checks of reachability, travel time, and interference, allowing schedules to be revised before execution. A collaborative policy built on Qwen2.5-VL-3B provides a reference baseline. Experiments show that coordination can shorten completion time and improve task reliability, while highlighting challenges such as congestion and synchronization.