Zhennan Wang 王振楠

With a dual background in robotics and AI, I pursue Physical AI, focusing on embodied multimodal world models, latent-space world models, unified generation and understanding, autonomous-driving world models, and end-to-end planning.

I am currently the Foundation Model Lead at RoboScience (since April 2026), working on embodied world models. Previously, I worked on autonomous-driving world models at XPENG (2025–2026), end-to-end autonomous driving at Haomo.AI (2023–2025), and foundational computer vision and AI for Science as a postdoctoral researcher at Pengcheng Laboratory (2021–2023).

I received my Ph.D. from Shenzhen University (2016–2020), advised by Prof. Chen Xu, where I studied deep learning and computer vision, particularly neural network regularization and visual recognition. Prior to that, I earned my master’s degree from the China Academy of Machinery Science and Technology (2011–2014), advised by Academician Zhongde Shan, focusing on digital control of robotic systems.

My research has appeared at venues including ICCV, NeurIPS, CVPR, ECCV, IJCV, and Nature Machine Intelligence. I have also served as a reviewer for NeurIPS, ICLR, CVPR, ICCV, and ECCV, among others.

Selected Projects

GenSight: future video with aligned segmentation, depth, optical flow and 4D robot trajectoriesGENSIGHT / 2026 ↗ PROJECT SITE

01 EMBODIED WORLD MODEL

GenSight

Seeing and Imagining as One

A unified world model bringing video generation and dense 4D perception into a single pretrained video diffusion transformer. RGB video, depth, embodiment segmentation, and first-frame-relative optical flow share a common representation, backbone, and velocity-prediction space.

The same model supports future video generation with aligned dense predictions and single-step perception from observed videos.

RoboScienceProject lead2026
Figure 2 of the X-World paper, showing the multi-camera video-generation architectureX-WORLD / 2026 PAPER FIGURE 2

02 AUTONOMOUS DRIVING WORLD MODEL

X-World

Controllable Multi-Camera World Modeling

A controllable multi-camera generative world model for autonomous driving. Starting from synchronized multi-view video history, X-World generates temporally coherent future driving videos conditioned on ego actions, surrounding traffic agents, static road elements, camera parameters, and appearance-level text prompts.

Its view-temporal attention and decoupled condition injection enable controllable and consistent simulation for closed-loop evaluation, online reinforcement learning, and scalable driving-data generation.

XPENGMain contributor2026
CogAD original paper Figure 1: hierarchical human perception and planningCOGAD / 2026 PAPER FIGURE 1

03 END-TO-END AUTONOMOUS DRIVING

CogAD

Cognitive-Hierarchy Guided End-to-End Autonomous Driving

An end-to-end planning framework inspired by human cognition. Its dual hierarchical design first reasons from global context to important local objects, then predicts driving intent before multi-modal trajectories.

Evaluated on nuScenes and Bench2Drive, with strong results in complex and long-tail scenarios. CogAD was also applied to Haomo.AI’s logistics-vehicle business.

Haomo.AIFirst contributor2025
Model architecture from the related Nature Machine Intelligence E2VD paperPENGCHENG · SHENNONG / 2022 NMI PAPER FIGURE

04 AI FOR SCIENCE

Pengcheng · Shennong

Simulating Viral Evolution with AI

Research on large-scale protein language models and AI-based simulation for anticipating high-risk SARS-CoV-2 variants, conducted with the Pengcheng Laboratory and collaborators.

The team's paper Running Ahead of Evolution was named an ACM Gordon Bell COVID-19 Special Prize Finalist (2022). I was one of the main contributors to the project. Related follow-up research on virus variation driver prediction was published in Nature Machine Intelligence (2025).

Pengcheng LaboratoryMain contributor2022 finalist

Selected Publications

2026
Research method figure

Seeing and Imagining as One: Unified Video Generation and 4D Perception for Embodied World Modeling

Haitang Feng*, Zhuoman Liu*, Zhennan Wang†, Sheng Wang†, et al.

arXiv preprint · GenSight · † Project leader

2026
Research method figure

X-World: Controllable Ego-Centric Multi-Camera World Models for Scalable End-to-End Driving

Chaoda Zheng, Sean Li, Jinhao Deng, Zhennan Wang, et al.

arXiv preprint · XPENG

2026
Research method figure

CogAD: Cognitive-Hierarchy Guided End-to-End Autonomous Driving

Zhennan Wang, Jianing Teng, Canqun Xiang, et al.

CVPR 2026 Workshop on Autonomous Driving (WAD) · First author

2025
Figure 2 from the A-FPL journal paper

Adaptive Fuzzy Positive Learning for Annotation-Scarce Semantic Segmentation

Pengchong Qiao, Yu Wang, Chang Liu, et al., Zhennan Wang, et al.

International Journal of Computer Vision (IJCV)

2025
Research method figure

A Unified Evolution-Driven Deep Learning Framework for Virus Variation Driver Prediction

Zhiwei Nie, Xudong Liu, Jie Chen, Zhennan Wang, et al.

Nature Machine Intelligence

2023
Research method figure

LaPE: Layer-adaptive Position Embedding for Vision Transformers with Independent Layer Normalization

Runyi Yu*, Zhennan Wang*, Yinhuai Wang*, et al.

ICCV 2023 · Equal contribution

2023
ACSeg original architecture figure from an author’s research site

ACSeg: Adaptive Conceptualization for Unsupervised Semantic Segmentation

Kehan Li, Zhennan Wang, Zesen Cheng, et al.

CVPR 2023

2023
Research method figure

Fuzzy Positive Learning for Semi-Supervised Semantic Segmentation

Pengchong Qiao, Zhidan Wei, Yu Wang, Zhennan Wang, et al.

CVPR 2023

2023
MCVP workflow faithfully adapted from Figure 2 of the Running Ahead of Evolution paper

Running Ahead of Evolution: AI-Based Simulation for Predicting Future High-Risk SARS-CoV-2 Variants

Jie Chen, Zhiwei Nie, Yu Wang, et al., Zhennan Wang, et al.

International Journal of High Performance Computing Applications · SC22 Gordon Bell COVID-19 finalist work

2022
Locality guidance method

Locality Guidance for Improving Vision Transformers on Tiny Datasets

Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, Jie Chen

ECCV 2022

2020
Research method figure

MMA Regularization: Decorrelating Weights of Neural Networks by Maximizing the Minimal Angles

Zhennan Wang and collaborators

NeurIPS 2020 · First author

2019
Projection and rejection diagram adapted from Figure 1 of PR Product

PR Product: A Substitute for Inner Product in Neural Networks

Zhennan Wang, Wenbin Zou, Chen Xu

ICCV 2019 · Oral presentation · First author

CogAD demonstration

Research demonstration · nuScenes / CARLA video supplied by the author