Check out the project page for more results and details! Many thanks to my amazing collaborators: Qixing Huang, Mikaela Uy, and @nmwsharp.bsky.social — and to the incredible NVIDIA Spatial Intelligence Lab for wonderful internship experience. 8/n
The resulting convex components can be seamlessly integrated into simulators such as MuJoCo and NVIDIA Newton for fast collision handling. 7/n
Our hierarchical decomposition algorithm also supports granularity control. It allows users to interactively adjust the threshold to achieve the desired level of detail. 6/n
Our model shows strong open-world generalizability across a diverse range of data in various modalities, including CAD models, 3D scans, and gaussian splats. 5/n
Datasets of convex decompositions do not exist to learn from — that’s where our key idea comes in. We introduce an unsupervised contrastive loss derived from the definition of convexity, which pulls features of convex pairs of points together while pushing features of non-convex pairs apart. 4/n
We train a model that converts any shape into a feature field, where the distance between two points indicates whether they should belong to the same convex component. We then apply clustering on this field to obtain a hierarchical convex decomposition. 3/n
Convex decomposition breaks a non-convex 3D shape into a set of convex components. It is widely used in physics engines as it accelerates geometric computation like collision detection. However, existing approaches either rely on expensive heuristic search or generalize poorly to openworld shape.2/n
Excited to share our new work at CVPR 2026: Learning Convex Decomposition via Feature Fields.
We introduce the first feedforward openworld model that generates high-quality convex decompositions for any 3D shapes in seconds, enabling faster simulation.
Project: research.nvidia.com/labs/sil/pro...
Check out the project page for more results and details!
Many thanks to my amazing collaborators: Qixing Huang, Mikaela Uy, and @nmwsharp.bsky.social — and to the incredible NVIDIA Spatial Intelligence Lab for wonderful internship experience. 7/n
The resulting convex components can be seamlessly integrated into simulators such as MuJoCo and NVIDIA Newton for fast collision handling. 6/n
Our model shows strong open-world generalizability across a diverse range of data in various modalities, including CAD models, 3D scans, and gaussian splats. 5/n
Datasets of convex decompositions do not exist to learn from — that’s where our key idea comes in. We introduce an unsupervised contrastive loss derived from the definition of convexity, which pulls features of convex pairs of points together while pushing features of non-convex pairs apart. 4/n
We train a model that converts any shape into a feature field, where the distance between two points indicates whether they should belong to the same convex component. We then apply clustering on this field to obtain a hierarchical convex decomposition. 3/n
Convex decomposition breaks a non-convex 3D shape into a union of convex components.It is widely used in physics engines as it accelerates geometric computations like collision detection. However, existing approach either rely on expensive heuristic search or generalize poorly to open-world shapes.
