# Genie: Generative Interactive Environments

**Authors:** Jake Bruce, Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai, Aditi Mavalankar, Richie Steigerwald, Chris Apps, Yusuf Aytar, Sarah Bechtle, Feryal Behbahani, Stephanie Chan, Nicolas Heess, Lucy Gonzalez, Simon Osindero, Sherjil Ozair, Scott Reed, Jingwei Zhang  
**arXiv:** [2402.15391](https://arxiv.org/abs/2402.15391)  
**Date:** 2024-02-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.15391v1/

## Key points

- Genie is a generative interactive environment trained using unlabelled Internet videos
- Capable of generating action-controllable virtual worlds through text, synthetic images, photographs, and sketches
- Consists of 11 billion parameters including a spatiotemporal video tokenizer, autoregressive dynamics model, and scalable latent action model
- Users can interact within the generated environments on a frame-by-frame basis without needing ground-truth action labels or domain-specific requirements
- Learned latent action space enables training agents to mimic behaviors observed in unseen videos
- Research team includes various members such as Michael Dennis, Ashley Edwards, Jack Parker-Holder, etc.
- Genie opens up new possibilities for creating interactive virtual environments and training advanced AI systems with minimal supervision

## Summary

In their paper titled "Genie: Generative Interactive Environments," Jake Bruce and a team of researchers introduce Genie - a groundbreaking generative interactive environment trained in an unsupervised manner using unlabelled Internet videos. This innovative model has the ability to generate a diverse range of action-controllable virtual worlds through text, synthetic images, photographs, and sketches. With an impressive 11 billion parameters, Genie serves as a foundational world model consisting of a spatiotemporal video tokenizer, an autoregressive dynamics model, and a scalable latent action model. One of the key features of Genie is its capability to allow users to interact within the generated environments on a frame-by-frame basis without the need for ground-truth action labels or other domain-specific requirements typically found in existing world models. Furthermore, the learned latent action space produced by Genie enables training agents to mimic behaviors observed in unseen videos - paving the way for the development of versatile generalist agents in the future. The research team behind Genie includes Michael Dennis, Ashley Edwards, Jack Parker-Holder, Yuge Shi, Edward Hughes, Matthew Lai,
Aditi Mavalankar,
Richie Steigerwald,
Chris Apps,
Yusuf Aytar,
Sarah Bechtle,
Feryal Behbahani,
Stephanie Chan,
Nicolas Heess,
Lucy Gonzalez,
Simon Osindero,
Sherjil Ozair,
Scott Reed,
Jingwei Zhang,
Konrad Zolna,
Jeff Clune,
Nando de Freitas,
Satinder Singh,and Tim Rocktäschel. Their work on Genie opens up new possibilities for creating interactive virtual environments and training advanced AI systems capable of adapting to various tasks and scenarios with minimal supervision. For more information on Genie and its applications, interested readers can visit their website at https://sites.google.com/corp/view/genie-2024/.