Advanced Planning with JEPA Models

Deep learning models capable of reasoning about their environment require the ability to capture the underlying environmental dynamics. Joint-Embedded Predictive Architectures (JEPA) offer a promising approach to model such dynamics, learning representations and predictors through a self-supervised prediction objective.

A recent study focuses on improving action planning within JEPA models. The research proposes shaping the representation space so that the negative goal-conditioned value function for a reaching cost in a given environment is approximated by a distance (or quasi-distance) between states.

A practical method has been introduced to enforce this constraint during the training phase. The results show a significant improvement in planning performance compared to standard JEPA models, particularly in simple control tasks.