Functionalization via Structure Completion
and Motion Rectification

Mingrui Zhao, Sai Raj Kishore Perla, Kai Wang, Sauradip Nag, Duc Anh Nguyen, Jiayi Peng, Ruiqi Wang, Angel X. Chang, Manolis Savva, Ali Mahdavi-Amiri, Hao Zhang

Simon Fraser University · ShanghaiTech University

SIGGRAPH Asia 2026

Functionalization teaser
A visually plausible cabinet that cannot function: its door has no hinge, its drawer has no rails, and opening the door collides with the side panel. Our neural graph functionalizer (GraFu) completes the functional graph with the missing connectors and parts, and geometric realization installs actual hinges and rails, rectifying the motion so the model becomes physically operable.

Why functionalization?

Most 3D assets are made to be looked at, not used: doors without hinges, drawers without rails, missing tops and interiors, and human-annotated motions that collide or detach. We introduce functionalization, the task of transforming such models into functional, physically operable counterparts. Our key idea is to treat it as functional graph completion, where labeled part nodes and typed edges (contact, hinge, rail, attached) expose every structural deficiency as a missing node or edge, followed by geometric realization that grounds the predicted graph with real mechanical parts. Because part motion is driven by an installed hinge or rail rather than a virtual axis, erroneous motions are rectified at the same time.

Video

How it works

Pipeline
GraFu encodes each part from its category, point cloud, and bounding box, relates parts with a graph transformer over a fully connected topology, and decodes a functionalized graph with a DETR-style slot decoder: per-slot categories, boxes, and motion axes, and per-pair edge types. A geometric fixing stage then snaps hinge and rail templates between the predicted parts.

Results

Functionalization gallery
Dynamic functionalization installs hinges and rails while rectifying motion ranges, axes, connectivity, and collisions (a-g). Static functionalization completes tops, interiors, and handles with user-selectable options (h-j).
Graph completion results and comparisons
From a partially labeled input graph (left), GraFu completes the functionalized graph with new nodes and functional edges (highlighted), which grounds the articulated result (Ours). Our functionalization matches or exceeds motion-prediction baselines while producing connected, collision-free articulation.

Interactive functionalization demo

Three PartNet-Mobility cabinets, functionalized live in your browser — one per collision-tested hinge class. Pick a cabinet, customize what functionalization adds (top-panel shape, handle style, rail type, interior shelving), then drag the slider or press play to operate it. Rotate with the mouse, zoom with the wheel.

Motion source
Operate

BibTeX

@article{zhao2026functionalization,
  title={Functionalization via Structure Completion and Motion Rectification},
  author={Zhao, Mingrui and Perla, Sai Raj Kishore and Wang, Kai and Nag, Sauradip and Nguyen, Duc Anh and Peng, Jiayi and Wang, Ruiqi and Chang, Angel X and Savva, Manolis and Mahdavi-Amiri, Ali and others},
  journal={arXiv preprint arXiv:2605.18010},
  year={2026}
}