SenseNav: Zero-Shot Vision Language Navigation with WordNet

The ability to locate and navigate to specific objects in unfamiliar environments is essential for robots to perform real-world tasks. To this end, we introduce SenseNav, a zero-shot navigation method that guides robots to target objects in novel settings. SenseNav constructs an occupancy grid from RGB-D observations to identify obstacles and frontiers.It further uses these observations to evaluate frontiers by integrating two metrics: semantic similarity between the target object and objects identified in the observations using WordNet, and the likelihood that the frontier could lead to the target object, estimated through frontier descriptions generated by a vision-language model and reasoning with a large language model. Extensive experiments conducted on the Habitat-Matterport 3D dataset (HM3D) using the Habitat simulator demonstrate that SenseNav achieves state-of-the-art performance in Success Rate (SR), while also achieving a strong performance in Success Weighted by Inverse Path Length (SPL).