Figure 1: Articulated object recognition by splatting screw axes and Gaussians. Articulated objects with movable parts - such as doors, laptops, and drawers - are common in everyday environments, and manipulating them requires understanding both their 3D geometry and underlying kinematic structure (e.g., joint types and axes). While prior work has addressed this using large-scale datasets of 3D objects with annotated joint axes in supervised settings [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], such methods struggle to generalize to unseen categories - a natural limitation of supervised learning. In this work, we tackle a more challenging yet practical scenario: inferring kinematic structure directly from multi-view RGB images under varying object configurations, without relying on category-specific supervision (see the left of Figure 1). Spurred in part by the success of neural rendering-based 3D reconstruction methods that require no supervised training [12, 13, 14, 15], recent works have adapted these frameworks for articulated object recognition [16, 17, 18, 19, 20], achieving promising results using raw RGB observations. However, a key drawback of these methods lies in their reliance on strong assumptions, such as a known number of articulated components or predefined joint types.