Event
StackGen3D: Generating Stable Block Stacking from Multi-View Silhouettes
1 reports · 1 sources
Xiaopeng Zhang, Luzhe Sun, Yiqing Xu, and Matthew R. Walter propose the StackGen3D framework, which generates stable block assembly structures from a given block inventory and three orthogonal silhouette views. It first uses a physics-informed graph neural network (PI-GNN) to model block support relationships and filter a stable-structure dataset, then trains a conditional flow matching model to learn a multimodal distribution of block configurations that conform to the target silhouettes. In preliminary experiments, Best-of-10 selection raised physical stability from 45.3% to 72.4%, with a silhouette IoU of 0.970. During inference, lightweight geometric filtering and correction reduce inter-block interpenetration without requiring physical simulation; the authors state that PI-GNN generalizes markedly better than a parameter-matched multilayer perceptron from simple stacking to complex structures.
Report timeline
- StackGen3D: Generating Stable 3D Block Stack Structures from Multiview Silhouettes
The StackGen3D framework generates stable block assembly structures from a given block inventory and three orthographic silhouette images. It first uses a physics-informed graph neural network (PI-GNN) to model support relationships between blocks and to curate a dataset of stable structures, then trains a conditional flow-matching model to learn the multimodal distribution of block configurations that match the target silhouettes. In preliminary experiments, Best-of-10 selection raised physical stability from 45.3% to 72.4%, with a silhouette IoU of 0.970.
arXiv cs.RO ·