Visual navigation typically relies on world models that propose complete action candidates and then predict their consequences to pick or refine a plan. InJePA, a new paper on arXiv, argues that this limits predictions to evaluating actions rather than guiding their stepwise construction. The authors instead propose an intention-first paradigm centered on latent states.

InJePA couples two JEPA-style world models: a goal-conditioned model proposes latent intentions, while an action-conditioned model predicts future latent states. Both operate in a shared latent space, so predicted consequences feed back into the next intention proposal and action inference. This creates a closed loop of intention, action, and consequence, extending the world model's role from evaluation to actively leading candidate construction.

The authors evaluate InJePA through closed-loop image-goal navigation in Habitat MP3D, reporting effective multistep navigation. Experiments across different encoders also support the method's applicability across visual representations. The findings suggest latent intentions can serve as the central planning object, with world models coordinating reasoning about intentions and consequences to directly guide plan construction.