언어 기반 네비게이션에서의 불확실성 재고

초록

Language-driven navigation is a key challenge for autonomous robots, requiring the integration of natural language understanding with robust perception and localization. Existing approaches in visual place recognition and navigation mainly rely on similarity-based feature mat- ching, while variance estimation or explicit uncertainty modeling has rarely been a fundamental component. Although a number of prior studies have investigated uncertainty in tasks such as object detection and target search, these efforts remain limited. In this paper, we revisit uncertainty in language-driven navigation and argue that it should be addressed at the distributional level rather than reduced to point estimates such as similarity search. Leveraging a large language model capable of semantic reasoning, we analyze how non-uniform distributions of semantic features and class activations affect navigation performance on benchmark datasets. The results show that conventional similarity-based measures are insufficient to capture the variability inherent in semantic maps, underscoring the necessity of redefining uncertainty modeling as a foundation for reliable perception, robust decision-making, and deployment of language-driven robotic navigation.

키워드