Toward Goal-Agnostic Joint-Embedding Predictive Control of Partial Differential Equations

Short summary

The authors propose a goal-agnostic control framework for PDEs using a joint-embedding predictive architecture (JEPA) with a frozen ViT encoder and action-conditioned latent dynamics. A model-predictive path integral controller reuses the frozen world model, and planning against explicit physical observables (like kinetic energy) outperforms raw latent-space distance minimization. On Navier-Stokes benchmarks, kinetic-energy-probe planning reduces velocity-field RMSE by 53% on dissimilar targets and supports stabilization control with 2.7% mean relative error.

• •JEPA-based latent dynamics trained offline without reward, frozen and reused by MPPI controller
• •Planning against physical observables beats latent-L2 distance on Navier-Stokes PDE control
• •Frozen model supports both trajectory matching (R²=0.989) and stabilization tasks

Generated with AI, which can make mistakes.