RigidFormer: Learning Rigid Dynamics using Transformers

Researchers have introduced RigidFormer, a novel object-centric Transformer-based model designed to simulate multi-object rigid-body dynamics using mesh-free inputs like point clouds. Traditional simulation methods often struggle with discontinuous contacts and high computational costs due to reliance on mesh connectivity. RigidFormer addresses these challenges by reasoning at the object level through compact anchors enriched with local vertex features via Anchor-Vertex Pooling. The model employs Anchor-based RoPE to integrate geometry into attention mechanisms while maintaining permutation equivariance and invariance to anchor reindexing. Rigidity is enforced by projecting updates onto the rigid-body manifold using differentiable Kabsch alignment. Benchmarks indicate that RigidFormer outperforms or matches mesh-based baselines in accuracy while offering faster processing speeds. It demonstrates strong generalization capabilities across unseen point resolutions and datasets, scaling effectively to simulations involving over 200 objects. Additionally preliminary extensions show potential for handling command-conditioned articulated bodies by treating body parts as interacting components. This advancement significantly improves the efficiency and applicability of learning-based physics simulations in computer vision and graphics.