Towards Socially Intelligent Robots: Research on Human-Aware Vision Language Navigation

Abstract

As robots are increasingly deployed in human-centric environments like homes and hospitals, their ability to navigate safely while respecting social norms is essential. Vision-and-Language Navigation (VLN) offers a promising interface by allowing robots to follow natural language instructions. However, current VLN research is largely confined to static, unpopulated simulations, where humans are treated as irrelevant or merely as obstacles. This paper aims to address the critical gap between mechanical instruction-following and socially intelligent behaviour. The author reviews the limitations of existing datasets and simulators, arguing that a fundamental shift is necessary. The author posits that integrating external Large Language Models (LLMs) is a hard requirement for achieving human-centric VLN, as the author provides the commonsense reasoning, intent prediction, and adaptive planning that current architectures lack. The study concludes that by leveraging LLMs as a reasoning backbone, VLN agents can transition from rigid path execution to responsive, socially aware navigation. This integration is crucial for bridging the divide between static language commands and the dynamic reality of shared spaces, ultimately enabling robots to become acceptable and effective collaborators in daily life.