JEPA as a Latent World Model for Model-Predictive Control in Simulated Physics and Language-Specified Goals

A world model predicts how the environment evolves, which allows an agent to plan without acting. We study a joint-embedding predictive architecture (JEPA) that predicts future latent states conditioned on actions, and use it inside a model-predictive controller (MPC). We evaluate planning success in simulated control tasks and check whether the latent space preserves the dynamics using known simulator states. Goals are specified in natural language and mapped to target latents through a frozen text encoder. A pilot arm with a linear model saturates at R^2 0.998, so the JEPA results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.