Latent world models

Compress the world you already have into latents and predict forward in that space. The JEPA line (LeCun's position paper, VICReg, V-JEPA 2) plus DreamerV3 and TD-MPC2. My project picture: V-JEPA 2 as frozen eyes, MuJoCo as the world, and a small action-conditioned predictor as the only thing I train, with planning done in latent space by CEM or MPC. The glove feeds the scene.