Skill-Dependent Representations for Object Navigation

Object navigation tasks aim to guide an agent to reach a target object with a monocular visual RGB sensor. The agent is required to possess various abilities, such as exploring unknown areas for environmental learning and navigating to the discovered target, to accomplish the tasks. However, most existing methods implicitly establish a holistic representation of these abilities, lacking interpretability and increasing training difficulty. In skill-based approaches, different skills are often fed with the same features, necessitating further learning of more suitable features for each skill by the agent. Therefore, we propose a Skill-Dependent Representations navigation frAMe-work (SDRAM), including a skill selection strategy, and two skill modules: exploration and navigation. The exploration skill utilizes high-level object representations and target embedding, while the navigation strategy employs detailed visual representations that are widely used in the state-of-the-art (SOTA) methods. Our method outperforms the SOTA models in the AI2Thor environment with higher navigation success rate, and we demonstrate that our framework has good generalizability.