Temporal-Distance JEPA: Plan-Aware Representation Learning for Latent World Model Predictive Control

The paper introduces Temporal-Distance JEPA (TD-JEPA), extending the LeWM encoder-predictor backbone with a directed temporal cost mined from reward-free offline trajectories. Same-trajectory ordering provides positive supervision, cross-trajectory pairs provide heuristic negatives, and rollout consistency aligns representation learning with the planning horizon. Under locked evaluation, the mined cost reaches 100.0% success on Two-Room versus 97.4% for LeWM. Shared Euclidean planning on the temporally trained checkpoint improves OGB-Cube by 14.2 points over LeWM and also improves Push-T. The authors report that TD-JEPA matches or exceeds LeWM and the concurrent RC-aux baseline across all evaluated environments.