Yiyang Zou (邹奕杨)

Short Bio

My name is Yiyang Zou (邹奕杨), I am currently a second-year Ms. student at the School of Computer Science, Wuhan University, where I am fortunate to be advised by Prof. Yu Wu. I received my Bachelor's degree in HongYi Honor Class, also at the School of Computer Science in Wuhan University.
My research focuses on Autonomous Driving, Reinforcement Learning and Embodied AI.

News

2026.06: I am invited as a reviewer for CICAI 2026.
2026.05: I started an internship at Tencent as a Game Reinforcement Learning AI Algorithm Developer.
2026.03: I started a research internship at HKU OpenDriveLab, working on embodied manipulation.
2026.04: One paper has been accepted by CVPR 2026 as Highlight.
2025.09: One paper has been accepted by NeurIPS 2025.

Internship

Algorithm Intern · DiDi Chuxing 2025.01 — 2025.06

Driving-Accident Anticipation · Out-of-Cabin Perception Group, Alot Dept.

3D-CNN Transformer Video Risk Forecasting

Built a YOLOv8-based automated data pipeline for hazardous-driving videos (detection-assisted labeling, key-event frame mining, risk-segment alignment), substantially cutting annotation cost and improving training-data quality.
Designed a 3D-CNN snippet encoder + Transformer temporal decoder on mmaction2 for multi-step future risk probability modeling; explored multi-backbone fusion, composite losses and multi-task learning, achieving stable gains on AUC and TTA on the internal benchmark.
Co-authored Accident Anticipation via Temporal Occurrence Prediction, accepted by NeurIPS 2025.

Research Intern · HKU OpenDriveLab 2026.03 — 2026.04

Embodied Manipulation — Value-Augmented Policy Training for Deformable Tasks

π_0.5 Flow-Matching Offline RL

Built a Value-Augmented Training Framework for Deformable Embodied Manipulation on top of π_0.5, jointly modeling action generation and scene-level value estimation for deformable tasks such as cloth folding.
Proposed a SigLIP-feature-based top-k similar-frame retrieval & score write-back mechanism, providing dense offline advantage labels for value learning and preference-guided generation.
Without an explicit value function, the framework matches π_0.6* performance on real-robot cloth-folding tasks.

Algorithm Intern · Tencent 2026.05 — present

Game Reinforcement Learning AI Algorithm Developer

PPO Game AI Reinforcement Learning

Selected Publications

RiskProp: Collision-Anchored Self-Supervised Risk Propagation for Early Accident Anticipation
Yiyang Zou^†, Tianhao Zhao^†, ..., Yu Wu*
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), CCF-A, 2026
[Paper][Code]

Accident Anticipation via Temporal Occurrence Prediction
Tianhao Zhao, Yiyang Zou, Peilun Xiao, ..., Yutian Lin*
Conference on Neural Information Processing Systems (NeurIPS), CCF-A, 2025
[Paper]

Selected Awards and Honors

Lei Jun Fund for Computer Innovation and Development, 2024
(雷军计算机创新与发展资助基金)

Bronze Medal, Hubei Provincial Site of the 2023 China Collegiate Programming Contest (CCPC), 2023
(2023年中国大学生程序设计竞赛湖北赛区铜奖)

First Prize in the 13,14th National Mathematics Competition for College Students, 2022,2023
(第十三、十四届全国大学生数学竞赛一等奖)