Socially-aware ObjectGoal Navigation using Proximity-based Auxiliary tasks

There has been a growing interest in “Embodied AI” in recent years. In an embodied AI task, an agent no longer learns from a static dataset extracted from the internet (e.g. image classification) but from the interaction with the environment. Three main problems are studied in embodied AI: Visual exploration, Visual navigation, and Embodied Question answering. In this thesis, I have focused my work on a specific visual navigation task: Object-goal navigation. In object-goal navigation, the agent is asked to navigate to a particular object in an unseen environment. This task has been further extended by introducing a social component: the presence of simulated human beings. The task of navigating in an environment with the presence of simulated people and finding an object-goal instance is called social object-goal navigation. Introducing a social component allows the agent to face more complex situations such as avoiding humans and, more in general, coexisting with them