IntentNav: Learning Spatial-Visual Object Navigation from Human Demonstrations

This research introduces IntentNav, a framework for object navigation in unknown environments by learning from human demonstrations. It employs Frontier-based Human-Intent Labeling to infer search intent and uses a spatial-visual candidate space with BEV memory and egocentric visual cues. A VLM policy is trained with an Intent-Aligned Objective to select candidates, achieving state-of-the-art success rates on MP3D, HM3D-v1, and HM3D-v2 benchmarks and enabling zero-shot transfer to wheeled, quadruped, and humanoid robots without further fine-tuning.