Socially-aware Object Goal Navigation

Recently, there has been a growing interest in “Embodied AI”, the research field that focuses on Artificial Intelligence problems for virtual agents, where the learning paradigm exploits data perceived from the agent’s egocentric point of view and its interactions with the environment. The tasks that a robot/agent must solve are numerous and can be grouped into three main categories: visual exploration, visual navigation, and embodied question answering. This work focuses on a particular visual navigation task called “Object Goal navigation”, which requires the agent to navigate to an instance of a given object category in unseen indoor environments. In particular, we aim to extend the aforementioned task by introducing a social component, namely the presence of humans within the environment, as similarly done in the recent Social Navigation task. Social tasks in general are designed to deal with the complexities of a more realistic environment, involving dynamic challenges such as moving pedestrians and to produce more comprehensive agents capable of coexisting efficiently with humans.