Handling uncertainty: How JEPA predicts multiple valid outcomes without producing "blurry" averages.
In the real world, the same action can lead to different results. If a ball hits a peg, it might fall left or it might fall right. Generative models (like MAE) often produce blurry predictions because they attempt to "average" all possible futures.
JEPA handles this by introducing a Latent Variable $z$. The predictor receives both the context and $z$. By varying $z$, the model can explore different sharp, physically valid future trajectories rather than a single blurry one.
Adjust the slider z to see different imagined futures. JEPA doesn't predict 'a bit of both'; it predicts a single, sharp trajectory conditioned on the latent variable. This allows agents to plan for specific scenarios (e.g., 'What if the car slips?' vs 'What if the car grips?').