SAP-Nav: Zero-shot robot navigation beats trained models by 12.2%

SAP-Nav tackles hierarchical open-vocabulary object navigation (OVON), where robots must follow free-form instructions like 'find the red mug in the kitchen on the counter.' Existing methods like LangMap struggle under partial observations because spatial grounding needs persistent environment-level evidence, while target verification demands clear candidate views. SAP-Nav solves this with a fully online, zero-shot framework. It incrementally builds a Queryable Spatial-Semantic Representation from actively acquired room views, allowing the robot to query spatial semantics from any location. Its Active Viewpoint Verification module checks whether the current observation is sufficient and, if not, repositions the agent to a more informative viewpoint before verifying candidates against category and attribute constraints.

The framework supports both hierarchical and standard category-level OVON without task-specific training or precomputed scene maps. Experiments on LangMap and HM3D-OVON benchmarks show SAP-Nav achieves the overall best performance, including a 12.2% improvement in success rate (SR) over training-based methods on region-level navigation. Real-world robot experiments further demonstrate its practical feasibility, proving the framework works beyond simulation. The code will be released upon acceptance, enabling other researchers to build on the approach. For developers in robotics and embodied AI, SAP-Nav offers a robust path toward more flexible, instruction-following agents in unseen environments.