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

Researchers introduced TD-JEPA, a JEPA-based world model that learns a directed temporal cost from reward-free trajectories for planning and representation learning.

It uses same-trajectory ordering, cross-trajectory negatives, and rollout consistency to infer progress-aware latent distances.

In locked evaluation, TD-JEPA improved success on navigation and manipulation benchmarks and matched or outperformed competing baselines.

The method helps offline world-model planning by reducing the mismatch between representation learning and control.