Lecture 13 - LeWorldModel | A JEPA World Model on One GPU

In this lecture, we look at LeWorldModel (Paper: https://arxiv.org/abs/2603.19312), the opposite of DINO-WM: the encoder is trained end-to-end from pixels, each frame becomes a single CLS token, and the whole model is ~15M parameters (≈48× faster planning). The key idea is the SIGReg loss, which prevents representation collapse by pushing random 1-D projections of the embeddings towards a Gaussian — no EMA, no stop-gradient. We finish with project ideas you can try on an SO-101 arm. Project: Train a 15M-parameter LeWorldModel on real SO-101 robot data on one GPU and use it to plan, detect surprises, or solve four tasks with one model https://github.com/RajatDandekar/build-a-world-model-from-scratch/tree/main/lecture-13-leworldmodel All slides, code & Colab notebooks for this series: https://github.com/RajatDandekar/build-a-world-model-from-scratch If you're enjoying the series, a ⭐ on the repo helps more learners find it.