FF-JEPA: Long-Horizon Planning in World Models with Latent Planners

A new hierarchical planning method called Forward-Forward-JEPA (FF-JEPA) aims to overcome the long-horizon limitations of existing world models by eliminating the need for explicit goal images, according to a preprint published on arXiv [1]. Joint Embedding Predictive Architectures (JEPAs) have demonstrated useful world modeling capabilities, allowing planning in latent space by optimizing action trajectories with techniques such as the Cross-Entropy Method (CEM) [1]. However, these methods are computationally expensive and break down when planning over extended sequences [1]. They also typically require an explicit image of the goal state, a condition that is not always feasible in real-world tasks [1]. The proposed FF-JEPA architecture addresses both constraints by introducing a hierarchical structure that relies on two forward dynamics models [1]. Alongside a standard action-conditioned forward model, the researchers add an action-free latent planner that predicts the next subgoal directly from the current state [1]. This decomposition turns a complex, long-horizon trajectory into a series of tractable, short-term optimization problems [1]. Preliminary experiments on the PushT task indicate that FF-JEPA avoids the long-horizon collapse observed in flat world models [1]. The authors describe the approach as a promising direction for goal-free planning [1]. The work appears as a preprint on arXiv and has not yet been peer-reviewed [1].