Dianwei Chen

CARLA · SUMO · ROS · PRGP

I build simulation and AI tools to make autonomous vehicles safer—especially when the road gets icy, the scenario gets messy, or the AI hits an edge case.

12
Publications
6
Published
6
Preprints
3
Universities

Featured work

VLA · 2026 SALSA — Safe Social Navigation Annotation-free post-training for vision–language–action models; 86% fewer near-collisions on SCAND and real-world deployment. ICDCS · 2026 Network Digital Untwinning Backward optimization of digital twins — IEEE 46th ICDCS 2026, Seoul. Survey · 2026 Digital Twins in ITCS: A Survey Six DT categories (SV-DT … ToT) for intelligent transportation and communication systems. IV Workshops · 2026 INSIGHT Vision–language models for context-aware hazard detection and edge-case AV evaluation — IEEE IV 2026 Workshops.

Latest news

  • Jun 2026 New preprint SALSA on safe social navigation with vision–language–action models. arXiv
  • May 2026 INSIGHT accepted at the 2026 IEEE Intelligent Vehicles Symposium (IV) Workshops. arXiv
  • May 2026 Network Digital Untwinning accepted at IEEE 46th ICDCS 2026, Seoul. arXiv

All news →

About me

All models are wrong, but some are useful. — George Box

I'm Dianwei Chen, a Ph.D. candidate in Civil & Environmental Engineering at the University of Maryland, College Park, advised by Prof. Terry Yang. I'm part of the M-Trail Lab, where we work on cyberinfrastructure and digital twins for transportation. Before UMD, I received my M.S. in Electrical & Computer Engineering from The Ohio State University (with Prof. Keith Redmill and Prof. Umit Özgüner) and my B.E. from Zhejiang University.

My research sits at the intersection of reinforcement learning, digital twins, AV safety evaluation, and vision–language models. I focus on how to test and validate autonomous systems in conditions that are hard to replicate in the real world—winter weather, rare pedestrian behaviors, and safety-critical edge cases. To do that, I build and use high-fidelity simulation (e.g. CARLA–SUMO co-simulation), physics-informed models such as Physics Regularized Gaussian Processes (PRGP), and RL-based behavior models to stress-test AVs and find failure modes before they hit the road.

I also work on the bridge from simulation to deployment: I've contributed to ROS and Autoware integration on real test vehicles (including PCD mapping and trajectory planning) and care deeply about open-source, reproducible tools for the AV and smart-transportation community.

If you're interested in collaboration, internships, or just talking about safe autonomy and digital twins, feel free to reach out.

Research interests

  • Reinforcement Learning for AV control
  • Digital Twins & simulation
  • Autonomous Vehicle Safety
  • Vision Large Language Models
  • Vision-Language-Action (VLA) models
  • CARLA-SUMO co-simulation
  • Physics Regularized Gaussian Process (PRGP)
  • Pedestrian behavior & edge-case AV failures

Timeline

Education and research path.

  1. Undergraduate B.E., Zhejiang University Engineering foundation before graduate study in the U.S.
  2. Graduate · 2023 M.S. in Electrical & Computer Engineering, The Ohio State University Research with Prof. Keith Redmill and Prof. Umit Özgüner; thesis on deep RL and adversarial pedestrian modeling (IEEE IV 2023).
  3. Present Ph.D. in Civil & Environmental Engineering, University of Maryland M-Trail Lab, advised by Prof. Terry Yang — digital twins, AV safety, simulation (CARLA–SUMO), and vision–language models for edge-case evaluation.

Visitors & fun

Live visitor map by Map My Visitors. Below: page views, GitHub, and a random research tip.

—
page views
GitHub stats for DianweiCHEN My GitHub →
Research tip

Loading…