PhysiFormer: Learning to Simulate Mechanics in World Space

A new diffusion transformer model called PhysiFormer can simulate physically plausible 3D object motion directly in world coordinates, outperforming autoregressive baselines on trajectory accuracy and physical consistency, according to research posted to arXiv on 25 June 2026 [1]. The model, developed by Yiming Chen, Yushi Lan, and Andrea Vedaldi, represents objects as 3D meshes and samples future vertex trajectories given initial positions, velocities, and material type — rigid or elastic [1][2]. Unlike earlier neural physics approaches that rely on ad-hoc latent spaces or explicitly enforced rigidity constraints, PhysiFormer casts vertex trajectory prediction as a single denoising diffusion process without such inductive biases [1][4]. The probabilistic formulation allows the model to capture uncertainty in learned dynamics, generating multiple plausible futures from the same initial conditions [1][5]. This capability could prove useful in settings where unobserved variables affect outcomes, the authors note [1]. PhysiFormer uses attention factorised over time, space, and objects, which enables permutation-invariant multi-object reasoning without requiring explicit object encoding [1][2]. The architecture was trained on over 100,000 simulated trajectories and generalises to mixed-material settings, unseen real-world geometries, and larger object counts than those seen during training [1][4]. In evaluations, the model substantially outperformed autoregressive baselines across three metrics: trajectory accuracy, rigidity preservation, and momentum-based physical consistency [1][5]. The authors position coordinate-space diffusion as a step toward view-invariant, geometry-aware world modelling for applications in robotics, graphics, and physical design [1][2]. arXiv, where the paper appeared, is an open-access repository of electronic preprints moderated but not peer-reviewed, hosting more than two million articles as of late 2021 [9]. The PhysiFormer paper was submitted under the Computer Vision and Pattern Recognition category [1]. Code, models, and visualisations are available on the project page at yimingc9.github.io/physiformer [1][5].