INSAIT paper “Exploration-Driven Generative Interactive Environments,” led by INSAIT doctoral student Nedko Savov, has been accepted for publication at CVPR 2025. This work is a collaborative effort between INSAIT, ETH Zurich, and TU Munich.
Research Highlights
Training AI to simulate, control, and generate virtual environments typically requires large-scale video datasets and expert demonstrations, which are costly and time-intensive to collect. Our approach addresses this bottleneck by leveraging automatically collected interactions across a diverse range of virtual environments.
Key innovations of this research include:
- AutoExplore Agent: A novel AI agent that explores solely based on the simulator’s uncertainty, eliminating the need for human demonstrations or environment-specific rewards.
- RetroAct Dataset: A large-scale dataset of labeled virtual environments, designed to validate the effectiveness of our approach.
- GenieRedux-G: A simulator model trained across numerous environments, achieving superior video quality and control with the assistance of the AutoExplore Agent.
For more details, please visit the project website.
Significance
This work lays the foundation for more robust and generalizable world models, moving us closer to AI systems that can learn, adapt, and explore in ways similar to human intelligence.
We extend our congratulations to the INSAIT team members: Nedko Savov, Naser Kazemi (INSAIT Summer Intern 2024), Mohammad Mahdi, Dr. Danda Pani Paudel, Dr. Xi Wang, and Prof. Luc Van Gool.
Additional Achievements
In addition to this publication, INSAIT authors have achieved a total of seven papers accepted at CVPR 2025. We commend all contributors on these significant accomplishments.
A complete list of accepted papers can be found here.