I built Micro-JEPA: A lightweight JEPA (Joint Embedding Predictive Architecture) in Python

I’ve always been fascinated by Yann LeCun’s vision of world models and autonomous agents, so I decided to build a minimal, lightweight implementation of a Joint Embedding Predictive Architecture (JEPA) from scratch, which I call Micro-JEPA.

In this project, the agent learns a representation of the environment, predicts future states in the latent space using a learned world model, and utilizes a cost/energy function to plan its steps toward a target while actively avoiding dynamic or static obstacles.

There is a video of it working in the README

GitHub Repo: https://github.com/Jacopos311/Micro\_JEPA