The model casts trajectory prediction as a denoising diffusion process over mesh coordinates. Factorized attention across time, space, and objects improves efficiency and supports permutation-invariant multi-object reasoning, while the original mesh topology is imposed at inference to assemble predicted vertices into 4D meshes.
PhysiFormer is useful for robotics simulation, graphics, physical design, and world-model research where future dynamics are uncertain. It is trained on more than 100,000 simulated trajectories, supports unseen geometries and larger object counts, and provides public code, checkpoints, an interactive viewer, and a Hugging Face demo.
