3D Spatial Reasoning

The field studies how AI can recover the structure of 3D scenes, track objects across viewpoints, and reason about distance, direction, movement, and possible actions. Common approaches add depth and camera cues to vision-language models, build explicit maps or scene graphs, or use geometric tools and active multi-view exploration; others train structured reasoning with supervised data or reinforcement learning. These methods support tasks such as visual grounding, navigation, scene editing, and robotic manipulation, and are tested on image, video, point-cloud, and embodied benchmarks. A central challenge is that strong question-answering scores can hide reliance on 2D or linguistic shortcuts: models still struggle with occlusion, changing viewpoints, metric accuracy, and carrying spatial beliefs reliably into action.