Interpretable Brain-Inspired Representations Improve RL Performance on Visual Navigation Tasks

Visual navigation is a complex but increasingly relevant task Related work in robotics and in machine learning (ML). Research in this field touches on a wide range of agent capabilities, including This section explains how previous works on navigation the parsing of tasks (Wang et al. 2021), locating objects toin ML literature have addressed the extraction of location, interact with (Lyu, Shi, and Zhang 2022), mapping out theheading and pose information. It also presents an overview environment (Chaplot et al. 2020) and planning (Gupta et al. of the relevant prior works on SFA for navigation, which stem from the field of computational neuroscience.2017). A basic capability in navigation, however, is that the agent always has to move around and find a path to its target. Localization for Navigation Representation learning in Finding a path to a location, crucially, requires awarenessthe context of RL and navigation is often approached of one's own location and heading. Unsurprisingly, it hasthrough auxiliary tasks (Lange et al. 2023; Jaderberg et al. been found that an agent's ability of self-localization is im-2017; Ye et al. 2021a; Mirowski et al. 2017), often withportant for navigation and especially long-term planning inout explicitly considering position, orientation or pose of an ML (Zhu et al. 2021).