About Me

I am the World Model Tech Lead at AFARI (MEGVII), working on world models for autonomous driving.

I focus on turning cutting-edge research into production systems, enabling applications such as closed-loop simulation, synthetic data generation, and closed-loop training.

Research interests: world models · controllable driving world simulation · synthetic data generation · 3D scene generation · multi-sensor calibration.

News

2026/08 📊 IsoWeatherBench preprint (adverse-weather video-gen benchmark) posted on SSRN.
2026/08 🚀 Released the AFARI World Model: BehaviorWorldGen.
2026/08 📄 Submitted GaussianDWM++ to TPAMI.
2026/06 🎉 PatchScene accepted to CVPR 2026.
2025/05 📐 High-precision stereo calibration published in Applied Optics.
2025/01 📡 Released targetless multi-LiDAR–camera–IMU calibration on arXiv.

Selected Publications

01 · Driving Intelligence

World Models & Scene Generation

4

Controllable, behavior-aware world models and generative 3D scenes for closed-loop autonomous-driving simulation, data generation, and training.

BehaviorWorldGen closed-loop overview
BehaviorWorldGen: Closing the Loop between Action Models and World Simulators via Controllable Behavior-Aware Structured World Generation
Flagship of the AFARI World Model Team — an interaction-consistent closed loop between action models and world simulators.
arXiv 2026 Project Lead · Corresponding
Paper Project
Jiaqi Wang, Zhuo Zhang, Haining Guan, Tingguang Zhou, Haowen Cui, ChuanYe Wang, Zhongyang Zhu, Yulong Zheng, Xuefeng Chen, Zhen Yang, Tianchen Deng, Feiyang Tan, Xiwu Chen, Hangning Zhou, Bo Dai, Lixia Shen, Xiyang Wang, Jiajun Zhu
Closes the loop between action models (VLA / end-to-end planners) and world simulators through controllable, behavior-aware structured world generation, enforcing consistent agent–ego interaction for balanced and reliable closed-loop rollouts.
IsoWeatherBench: An Iso-Severity Benchmark for Adverse-Weather Video Generators with a Physically-Calibrated Visibility Axis
Under Review Project Lead · Corresponding
Paper SSRN
Baocheng Geng, Bo Dai, Jiajun Zhu
An iso-severity benchmark that places adverse-weather video generators on a physically-calibrated visibility axis, revealing that apparent gains often reflect milder generated weather rather than higher quality — evaluated on nuScenes for autonomous-driving perception.
GaussianDWM++: Language-Grounded 3D Gaussian Driving World Model for Unified Scene Understanding, Editing, and Multi-Modal Generation
TPAMI · Under Review Project Lead
Paper arXiv
Tianchen Deng, X. Chen, S. Wu, Q. Chen, Jiajun Zhu, Bo Dai, J. Yang, H. Wang
A language-grounded 3D Gaussian driving world model that unifies scene understanding, editing, and multi-modal generation within a single representation.
PatchScene: Patch-based Voxel Diffusion for Large-Scale Scene Completion
CVPR 2026 Project Lead · Co-first Author
Paper Project
Qingdong Xu*, Jiajun Zhu*, Shilin Zhu, Xinjing He, Chao Lu, Huanran Wang, Jiyao Zhang
A patch-based voxel diffusion model that scales generative completion to large-scale 3D scenes for autonomous driving.
02 · Sensing Foundations

3D Sensing, Calibration & Reconstruction

3

Geometric calibration, high-precision 3D measurement, and active next-best-view planning for reliable perception, scanning, and 3D reconstruction across cameras, LiDARs, and IMUs.

NBV-Diffusion: Generative Next-Best-View Planning for Robot-Assisted 3D Scanning
RCIM · Under Review
Wenqin Liu, Jiajun Zhu, et al.
A generative diffusion framework for next-best-view planning in robot-assisted 3D scanning, predicting informative viewpoints to achieve complete, high-quality 3D reconstructions.
High-Precision Stereo Calibration for 3D Measurement via Speckle-Enhanced Circular Control Point Matching
Applied Optics 2025
Paper Optica
W. Liu, Jiajun Zhu, W. Li, L. Yang, S. Hu, Z. Zhang
A high-precision stereo camera calibration method for 3D measurement using speckle-enhanced circular control-point matching. Applied Optics 64(13), 3561–3570.
Targetless Intrinsic and Extrinsic Calibration of Multiple LiDARs and Cameras with IMU using Continuous-Time Estimation
arXiv 2025
Paper arXiv
Y. Lv, Y. Zhang, C. Lu, Jiajun Zhu, S. Wu
Targetless intrinsic and extrinsic calibration of multiple LiDARs and cameras together with an IMU, formulated through continuous-time estimation.

Experience

MEGVII

World Model Tech Lead

Led world models for autonomous driving — closed-loop simulation, synthetic data generation, and closed-loop training.

2023 – Present

Honors & Awards

  • Excellent Postgraduate, Xi'an Jiaotong University, 2022/2023
  • National Scholarship, China, 2017