Flow-JEPA Enhances Robustness in AI World Models

Joint-Embedding Predictive Architectures (JEPAs) have demonstrated considerable promise in learning compact, predictive representations, with LeWorldModel (LeWM) extending this to reconstruction-free latent world modeling from pixel data. However, LeWM's reliance on deterministic, autoregressive predictors can lead to error accumulation and sensitivity to irrelevant visual noise. This research introduces Flow-JEPA (F-JEPA), a novel approach that addresses these limitations. F-JEPA employs a conditional flow matching dynamics model to generate a sequence of future latent states, conditioned on the current observation and actions. This method replaces the point-wise transition regression with a more robust stochastic trajectory-level prediction. By using a Gaussian distribution as the flow source, the model learns to transport perturbed latent trajectories towards cleaner future representations. F-JEPA significantly improves performance, raising mean success rates from 86% to 92% under clean conditions and from 67% to 86% under noisy observations, indicating a substantial increase in robustness for JEPA world models.