Articulated Object Reconstruction

Articulated object reconstruction infers an object's 3D shape, rigid parts, joint structure, and motion from visual or depth observations, enabling interactive digital twins and robotic planning. Methods use motion tracking and part segmentation, explicit geometric and kinematic optimization, or feed-forward and generative models, often with Gaussian splatting, meshes, or geometric primitives. Research increasingly targets casual monocular video, sparse views, and single images, while drawing on synthetic benchmarks and real-world interaction data. A central challenge is recovering complete geometry and physically valid kinematics despite occlusion and limited observations, while generalizing across object categories and producing assets reliable in simulation.