Embodied Navigation

Embodied navigation studies how an agent uses sensory input and movement to reach objects or places, follow language instructions, explore unfamiliar spaces, and act over long periods. Approaches include learned policies and vision-language models, semantic or topological maps with memory, frontier-based exploration, world-model planning, and audio-visual or floor-plan guidance; mobile robots also use localization and traversability methods based on cameras, LiDAR, or radar. Methods are commonly tested on ObjectNav, PointNav, and vision-and-language tasks in Habitat, HM3D, MP3D, and R2R-CE, with increasing use of real-robot and dynamic-environment evaluations. A central challenge is making performance reliable beyond familiar simulation settings, especially over long routes, under changing or corrupted observations, and when exploration, recovery, safety, and goal verification must work together.