From Passive Observation to Active Interaction: Learning the causal effects of agency while maximizing compute efficiency.
Passive JEPAs predict the future based on observation alone. **Act-JEPA** introduces agency by conditioning predictions on an Action Sequence. The model asks: "Given the current state $S_t$, what is the representation of the world at $S_{t+1}$ if I apply force $A_t$?"
Predicting in Latent Space is fundamentally different from predicting in Pixel Space. While a generative model must render every surface detail, a JEPA model only predicts the abstract mathematical features required for control.
By predicting only the (x, v) state vector, Act-JEPA consumes minimal power. The prediction is mathematically perfect because it ignores irrelevant pixel noise and focuses on the underlying physics.