RigidFormer: Learning Rigid Dynamics Using Transformers (arxiv.org)

Researchers have introduced RigidFormer, an innovative Transformer-based model designed for learning rigid-body dynamics without relying on traditional mesh connectivity. This advancement is crucial as it tackles the complexities of simulating multi-object interactions in rigid dynamics, addressing limitations of existing methods that often depend on dense vertex-level message passing. RigidFormer employs an object-centric approach, utilizing compact anchors to enhance local vertex features while ensuring computational efficiency and maintaining contact-relevant geometry. The significance of RigidFormer lies in its ability to handle mesh-free inputs, making it applicable to various representations such as point clouds. It demonstrates superior performance compared to traditional mesh-based baselines on standard benchmarks, showcasing faster processing speeds and the capability to generalize across different datasets and object resolutions. Additionally, its design incorporates mechanisms like Anchor-based RoPE to manage unordered object interactions and employs differentiable Kabsch alignment to enforce rigid dynamics effectively. This model not only advances the state of the art in physics-based simulation but also opens avenues for future applications, including conditionally controlling articulated bodies by treating their components as interacting objects.