Embodied Navigation at the Art Gallery

Embodied agents, trained to explore and navigate indoor photorealistic environments, have achieved impressive results on standard datasets benchmarks. So far, experiments evaluations involved domestic working scenes like offices, flats, houses. In this paper, we build release a new 3D space with unique characteristics: the one of complete art museum. We name environment ArtGallery3D (AG3D). Compared existing scenes, collected is ampler, richer in visual features, provides very sparse occupancy information. This feature challenging for occupancy-based agents which are usually crowded environments plenty Additionally, annotate coordinates main points interest inside museum, such as paintings, statues, other items. Thanks manual process, deliver benchmark PointGoal navigation space. Trajectories dataset far more complex lengthy than ground-truth paths Gibson Matterport3D. carry extensive experimental evaluation using our prove that methods hardly adapt scenario. As such, believe availability model will foster future research help improve solutions.