Abstract
We present MINOS, a simulator designed to support the development ofmultisensory models for goal-directed navigation in complex indoorenvironments. The simulator leverages large datasets of complex 3D environmentsand supports flexible configuration of multimodal sensor suites. We use MINOSto benchmark deep-learning-based navigation methods, to analyze the influenceof environmental complexity on navigation performance, and to carry out acontrolled study of multimodality in sensorimotor learning. The experimentsshow that current deep reinforcement learning approaches fail in largerealistic environments. The experiments also indicate that multimodality isbeneficial in learning to navigate cluttered scenes. MINOS is releasedopen-source to the research community at http://minosworld.org . A video thatshows MINOS can be found at https://youtu.be/c0mL9K64q84