Kinematic Kitbashing for Modeling Functional Articulated Objects

TL;DR

Kinematic Kitbashing tackles automatic synthesis of functional articulated objects by reusing parts from existing models. It couples a kinematics-aware geometric attachment energy $E^{\mathrm{km}}$, computed via vector distance fields, with black-box functionality energies $E^{\mathrm{func}}$ and optimizes them through annealed Langevin dynamics to place parts so that attachments remain valid across the full articulation range while meeting task goals. Key innovations include VDF-based snapping, a training-free, gradient-free optimization that accommodates non-differentiable functionals, and optional attachment priors to guide placement. The approach yields diverse, functional assemblies and outperforms state-of-the-art baselines on geometric, kinematic, and functional metrics, enabling rapid creation of interactive articulated assets across cross-category parts.

Abstract

We introduce Kinematic Kitbashing, an automatic framework that synthesizes functionality-aware articulated objects by reusing parts from existing models. Given a kinematic graph with a small collection of articulated parts, our optimizer jointly solves for the spatial placement of every part so that (i) attachments remain geometrically sound over the entire range of motion and (ii) the assembled object satisfies user-specified functional goals such as collision-free actuation, reachability, or trajectory following. At its core is a kinematics-aware attachment energy that aligns vector distance function features sampled across multiple articulation snapshots. We embed this attachment term within an annealed Riemannian Langevin dynamics sampler that treats functionality objectives as additional energies, enabling robust global exploration while accommodating non-differentiable functionality objectives and constraints. Our framework produces a wide spectrum of assembled articulated shapes, from trash-can wheels grafted onto car bodies to multi-segment lamps, gear-driven paddlers, and reconfigurable furniture, and delivers strong quantitative improvements over state-of-the-art baselines across geometric, kinematic, and functional metrics. By tightly coupling articulation-aware geometry matching with functionality-driven optimization, Kinematic Kitbashing bridges part-based shape modeling and functional assembly design, empowering rapid creation of interactive articulated assets.

Figures (14)

• Figure 1: We introduce Kinematic Kitbashing -- a framework that assembles reusable kinematic parts into functional, articulated objects driven by user-specified tasks. Left: Three lamp modules are kitbashed with a weighted base and support bars; our framework optimizes part placements so all heads reach their target points, producing a multi-head lamp. Right: A tabletop, drawers, doors, and a fold-down screen are rearranged into a reconfigurable workstation; the optimized layout allows the entire assembly to fold into a cube that fits inside a moving truck, while still deploying smoothly when needed.
• Figure 2: Kinematics-aware attachment for a novel part pair. A wheel mounted on a trash can is reattached to a car body. Our vector distance field energy guides it to the fender recess, matching its original socket.
• Figure 3: Effect of kinematic attachment. A drawer slide and a hinged door taken from the table are attached to a cabinet. Without kinematic-aware attachment (top), i.e., the placement is optimized for a single pose and ignores the full articulation, the doors are misoriented. With the kinematic dimension (bottom), our method positions both parts plausibly.
• Figure 4: ICP vs. VDF attachment. ICP ignores the clearance required between parts and snaps the arm directly into the torso surface. When the arm articulates, this oversight becomes critical, producing collisions. VDF, in contrast, keeps the necessary gap (red arrows) for collision-free articulation.
• Figure 5: Kinematic attachment under different priors. Our kinematics-aware attachment energy alone may settle in different local minima depending on the initialization. The data-driven prior guides the wheels into the canonical hub slots, providing placement with less dependence on initialization. A user may pin the wheel to any desired location, even one absent from the training set, such as the Airbus wing, forcing the wheel to follow this out-of-distribution location and enabling customized assemblies.