Prior Knowledge Guided Vision‐Language Navigation of High‐Level Instructions for Household Service Robots

Visual‐language navigation (VLN) plays a crucial role in household service robots, enabling them to understand and execute natural language commands while interacting with dynamic indoor environments. However, current VLN methods often struggle with efficiently navigating to small objects due to low pixel occupancy and frequent occlusion, especially in unseen environments. To this end, a knowledge‐guided navigation policy is proposed that employs a two‐stage strategy combined with Exploration Memory to improve navigation efficiency. Specifically, prior object‐container relationships are utilized to improve small object localization. Specifically, the agent is first guided toward relevant containers using semantic masks, reducing unnecessary exploration, and then switches to the small object's semantic mask once it is found. Experimental results on the ALFRED benchmark demonstrate that incorporating prior knowledge can improve navigation efficiency.