Temporal-Distance JEPA (TD-JEPA) introduces a directed temporal cost to Joint-Embedding Predictive Architectures by mining supervision from reward-free trajectories. This allows latent world models to move beyond Euclidean distance for ranking imagined futures, enabling more effective multi-step planning.
HOW THIS AFFECTS YOU
●
researcherThis provides a more robust way to integrate temporal order into latent representation learning for predictive control.