Schr\"odinger's Navigator: Imagining an Ensemble of Futures for Zero-Shot Object Navigation

Schrödinger's Navigator introduces a belief-aware framework for zero-shot object navigation (ZSON) that operates without task-specific fine-tuning or pre-built maps. The model employs a trajectory-conditioned 3D world model to predict multiple hypothetical observations, utilizing an adaptive occluder-aware sampler and a Future-Aware Value Map (FAVM) to enhance decision-making in cluttered environments. Experimental results demonstrate significant improvements in hidden-target discovery and risk-aware waypoint selection compared to existing ZSON methods, underscoring its potential for robust navigation in uncertain real-world scenarios.