SkillNav: Modular Skill-Based Framework for Vision-and-Language Navigation

Researchers propose SkillNav, a modular framework for Vision-and-Language Navigation (VLN) that decomposes navigation into interpretable atomic skills such as Vertical Movement, Area and Region Identification, and Stop and Pause. Each skill is handled by a specialized agent within a Transformer-based architecture. To enable targeted skill training without manual annotation, the authors develop a synthetic dataset pipeline that generates diverse, linguistically natural instruction-trajectory pairs. The approach aims to improve generalization to unseen scenarios requiring complex spatial and temporal reasoning. The work is detailed in arXiv preprint 2508.07642.