Degradable physical interactions are studied with adaptive spatial tokenization
This paper was accepted at the AI for scientific workshop on neurips 2025.
Measuring interactions between solids is important in fields such as materials science, mechanical engineering, and robotics. While learning-based methods with Graph Neural networks (GNNs) are effective in solving complex physical systems, they encounter scalability issues when modeling virtual physical interactions. To model communication between objects, global edges must be rendered dynamically, which is cumbersome and impractical for large maps. To overcome these challenges, drawing on views from geometric perspectives, we propose a spatial tokenial tokenization (AST) method for the effective representation of physical spaces. By dividing the simulation space into a grid of cells and mapping meshes to this structured grid, our method naturally joins groups of adjacent mesh nodes. We then add an attention module to count sparse cells in compact, random-length embeddings, serving as tokens for the whole body. Attention modules are employed to predict the next situation in these icons in the background space. This framework discovers the efficiency of Pokenization and the light power of attention methods to achieve precise and stubborn simulation results. Extensive testing shows that our method significantly outperforms state-of-the-art methods in modeling the disabled body. Notably, it remains effective at large scales with meshes exceeding 100,000 locations, where existing methods are restricted by policy limitations. Additionally, we contribute a novel large-scale dataset covering a wide range of physical interactions to support future research in this area.
