26.0131 [email protected], linkedin.com/in/terry-taylor-biz, Lab notes (Gdrive), Git

First JEPA demo doc.

see Ch 10 Demo 10a describes a demo for Planning in Latent Space with JEPA. The goal of this demo is to show that a JEPA world model can support planning directly in latent space. Images are encoded by I-JEPA, a latent dynamics model predicts how the latent state evolves under actions, and a planner searches over action sequences to reach a goal.

• Train: 10a1_train_toy1d_predictor_TRICK.py learns
• Plan: 10a2_goal_conditioned_random_shooting_TRICK.py uses Random Shooting MPC to choose the best 10-step sequence

The resulting GIF shows the system moving from start → goal using only its latent world model, demonstrating the architecture behind JEPA-style model-based control.