Introducing RoadGPT: AI-Powered Scenario Generation for AV Testing

We’re excited to share a new research project: RoadGPT, a tool that can automatically generate 3D road scenarios for autonomous vehicle testing from simple text descriptions.

Highlights:

  • Turns natural language prompts into fully functional 3D roads in BeamNG.tech.

  • Powered by OpenAI’s ChatGPT, combining AI with high-fidelity simulation.

  • Validated with expert user studies — showing strong usability and potential.

  • Could become a standard tool for AV scenario creation, making testing faster and more flexible.

This opens up interesting possibilities for the community: creating custom test scenarios, experimenting with unusual road layouts, or even training AV systems in diverse environments, all from a textual description!

We’d love to hear your thoughts:

  • How would you use something like RoadGPT in your projects?

  • Any ideas for scenarios you’d like to generate automatically?

Explore More:

For a deeper dive into RoadGPT, including access to simulation videos and survey data, visit the Zenodo repository and / or check out the paper here: BeamNG.tech Research

:writing_hand: Authors:
Alessio Gambi, Dejan Nickovic, Benedikt Steininger, Chrysanthi Papamichail, David Stark

Let’s start a conversation! :automobile::dashing_away:

2 Likes

Just wanted to point out that the link to the research paper points to the list of research papers, that however doesnt seem to contain a paper about this project. Maybe im just not looking right?

Yes, you are right. You can find it under this title “Automatically Generating Content for Testing Autonomous Vehicles from User Descriptions”.

1 Like

This would be interesting, previously I saw some BeamNG work that uses NLPs to re-generate scenarios from crash report. It would be nice to see those scenarios being generated with the powerful LLMs we have right now.

2 Likes