← Retour au blog
tech 15 July 2026

LeMario: Training a JEPA World Model on Super Mario Bros

Explore how a JEPA model is revolutionizing the learning of Super Mario Bros dynamics, and the lessons learned from this unique experience.

Article inspired by the original source
LeMario: Training a JEPA World Model on Super Mario Bros ↗ www.benjamin-bai.com

Introduction

Super Mario Bros, a timeless classic in video gaming, becomes the testing ground for a Joint-Embedding Predictive Architecture (JEPA) model. As artificial intelligence continues to make significant strides in understanding and predicting dynamic environments, applying these technologies to iconic video games offers unique opportunities to test and showcase their capabilities.

What is the JEPA Model?

The JEPA (Joint-Embedding Predictive Architecture) model is designed to learn world dynamics from pixels and actions. Initially used for reward-free planning in the Push-T environment, the model was adapted for Super Mario Bros by Benjamin Bai, a gaming enthusiast and AI researcher.

The main goal of JEPA is to predict future game steps based on past images and actions. This anticipatory capability allows for more strategic planning and better interaction with the game's environment.

The Building Process

Bai started by encoding game images into latent representations, transforming each Mario image into a 192-dimension private numerical description. This representation enables the model to grasp the essential elements of each screenshot.

Each pair of observations is separated by five emulator frames, with six possible button states per frame. The model uses these button sequences to forecast how the game will evolve, integrating this information into the causal predictor.

The Challenges Faced

While the model excelled at predicting game evolution over short distances, it struggled with more distant goals. For instance, the model could move Mario towards nearby goals but had difficulty navigating to farther target images.

These challenges highlighted the current limitations of JEPA models in handling the complexities of more advanced video games. Prediction errors increase with the distance and complexity of the goals to be reached.

Experiences and Lessons Learned

The experiments underscored the importance of integrating more robust learning mechanisms to enhance the model's ability to navigate complex environments. Future adjustments might include improving action encoding techniques and updating the architecture of transformer blocks.

Conclusion

Applying the JEPA model to Super Mario Bros provided a deeper understanding of video game dynamics and identified areas for improvement in future AI architectures. This experience reminds us that even advanced technologies require continuous adjustments to tackle the challenges of complex interaction with digital environments.

Let's discuss your project in 15 minutes.

JEPA Super Mario Bros intelligence artificielle modèle prédictif jeu vidéo
Deepthix newsletter · 100% AI · every Monday 8am

An AI agent reads tech for you.

Our AI agent scans ~200 sources per week and ships the best articles to your inbox Monday 8am. Free. One click to unsubscribe.

Visit the newsletter page →

Want to automate your operations?

Let's talk about your project in 15 minutes.

Book a call