LaWM: Least Action World Models for Long-Horizon Physical Consistency from Visual Observations

Researchers Qixin Xiao and Maani Ghaffari have introduced Least Action World Models (LaWM), a novel framework for embodied AI that enhances long-horizon physical consistency in visual predictions. Published on arXiv, this study addresses limitations in existing latent world models and video generation systems, which often suffer from compounding errors and energy drift due to unconstrained neural transition functions. LaWM operationalizes the Principle of Least Action within a learned visual latent space, utilizing a latent variational integrator to govern future rollouts via a learned Lagrangian action functional. By encoding observations into generalized coordinates and solving discrete integration conditions, the model ensures that physical structure defines the latent transition rule itself rather than merely regularizing trajectories. Benchmark tests on synthetic dynamics and robot interaction scenarios demonstrate that LaWM significantly improves physical invariance, background consistency, motion smoothness, and geometric prediction metrics compared to current baselines. This advancement offers promising applications for model-based reinforcement learning and robotic planning by providing a structure-preserving bias for accurate long-term visual forecasting.