Driving Simulation from Real-World Data: How Well Can We Render and Drive?

Welcome to the RealADSim Workshop organized at

Join us on 19 Oct 2025 from 12:00 - 17:45 HST

Introduction: Given the safety concerns and high costs associated with real-world autonomous driving testing, high-fidelity simulation techniques have become crucial for advancing the capabilities of autonomous systems. While classical driving simulators offer closed-loop evaluation, they still exhibit a domain gap compared to the real world. In contrast, offline-collected driving datasets avoid this gap but struggle to provide closed-loop evaluation. Novel View Synthesis (NVS) has recently opened up new possibilities by enabling closed-loop driving simulation directly from real-world data, which has attracted great attention. This creates a promising alternative for evaluating autonomous driving algorithms in dynamic, interactive environments. However, while NVS-based simulation unlocks new opportunities, there are two key questions that are yet to be answered: 1) How well can we render? 2) How well can we drive?

News

30 Jun 2025 — The Workshop website is launched.
11 Mar 2025 — The Workshop is accepted!

Important Dates

30 Jun 2025 — Challenge Release
31 Aug 2025 — Challenge Submission Due
05 Sep 2025 — Release Results & Submit Technical Report
20 Sep 2025 — Technical Report Due

To be eligible for awards, teams are required to submit a technical report of no more than 4 pages. Please note that these reports will not be included in the official ICCV proceedings.

Tentative Schedule

The workshop will take place on 19 Oct 2025 from 09:00 - 12:15 HST.

NOTE: Times are shown in Hawaii Standard Time. Please take this into account if you plan to join the workshop virtually.

Time (PDT)	Event
09:00 - 09:10	Welcome & Introduction
09:10 - 09:40	Keynote-1
09:40 - 10:10	Keynote-2
10:10 - 11:10	Awards / Challenge winner Presentation
11:10 - 11:40	Keynote-3:
11:40 - 12:10	Keynote-4:
12:10 - 12:15	Closing remarks

Invited Speakers

Yue Wang
Assistant Professor
University of Southern California

Jyh-Jing Hwang
Research Scientist
Waymo

Yue Wang is an Assistant Professor at USC CS, leading the Geometry, Vision, and Learning Lab. His current focus includes simulation, perception, and decision making. He obtained the Ph.D. degree from MIT EECS in 2022.

Yuexin Ma is an Assistant Professor in SIST, Shang- haiTech University. She received the PhD degree from the University of Hong Kong in 2019. Her current research focuses on scene understanding, multi-modal learning, autonomous driving, and embodied AI.

Jyh-Jing Hwang is a Research Scientist at Waymo Research, a technical lead for end-to-end autonomous driving. He received his Ph.D. degree in Computer and Information Science from the University of Pennsylvania.

Peter Kontschieder is the Director of Research at Meta. He received his PhD in 2013 from Graz University of Technology. His research interests include photorealistic 3D scene reconstruction, semantic scene understanding, image-based 3D modeling, and generative models for 3D synthesis.

Competitions

Tracks

We are holding two tracks in the workshop competitions:

Track 1: Extrapolated Urban Novel View Synthesis

In this track, we investigate the question: how well can we render? While NVS methods have made significant progress in generating photorealistic urban scenes, their performance still lags in extrapolated viewpoints when only a limited viewpoint is provided during training. However, extrapolated viewpoints are essential for closed-loop simulation. Improving the accuracy and consistency of NVS across diverse viewing angles is critical for ensuring that these simulators provide reliable environments for driving evaluation.
Track 2: Autonomous Driving in a Photorealistic Simulator

In this track, we investigate the question: how well can we drive? Despite challenges in extrapolated viewpoint rendering, existing methods enable photorealistic simulators with reasonable performance when trained on dense views. These NVS-based simulators allow autonomous driving models to be tested in a fully closed-loop manner, bridging the gap between real-world data and interactive evaluation. This shift allows for benchmarking autonomous driving algorithms under realistic conditions, overcoming the limitations of static datasets.

How to Participate

To participate in the competition, both automatic registration and manual verification are required:

Click the “Login with Huggingface” button.
Click the “Register” button and complete the registration form. After this automatic registration step, the “Submission Information” page will become accessible. It provides detailed instructions on how to run local tests and submit your proposal.
Access to “My Submissions” and “New Submission” will be granted after we manually review your registration and authorize your account. This process is typically completed within 24 hours.

💰 Awards

Each competition will have the following awards:

Innovation Award: $9,000
Outstanding Champion: $9,000
Honorable Runner-up: $3,000

Winners will be announced at the Workshop @ ICCV 2025.

🤵 Organizers