Brief

Researchers have developed PSG-Nav, a novel approach to open-vocabulary navigation for embodied agents that addresses perception uncertainty. The system constructs a 3D Probabilistic Scene Graph to represent semantic ambiguities and model errors. It employs Multiverse Decision making to sample likely world states and evaluate navigation landmarks based on their compatibility with these sampled states. An Evidential Experience Calibrator is also introduced for online adaptation to mitigate false positives. AI

A new research paper questions the extent to which large language models (LLMs) contribute to zero-shot gains in instruction-guided navigation systems. The study introduces two training-free variants, FPE and SHF, which rely on engineered geometric approaches and lightweight semantic heuristics respectively. Results indicate that careful geometric engineering can match or surpass LLM-driven performance, suggesting language models are more effective as heuristics than end-to-end planners in this domain. AI