I am Likui Zhang (张立魁), a Ph.D. student (combined Master-Ph.D. program) at the School of Computer Science, Sun Yat-sen University, advised by Prof. Liang Lin and Prof. Xiaodan Liang in HCP Lab. Previously, I received my B.Eng. in Intelligent Science and Technology from Sun Yat-sen University (School of Intelligent Engineering).

My research focuses on Embodied AI, World Model and 3D Vision, specifically on data simulation, Vision-Language-Action (VLA) models, World Models(WM) and 3D Gaussian Splatting. I am always open to research collaborations and interested in exploring related opportunities. Feel free to reach out via email: zhanglk9@mail2.sysu.edu.cn.

🔥 News

• 2026.05 🔥 Released JoyAI-RA 0.1, a foundation model for robotic autonomy (JD Explore Academy).
• 2026.02 🎉 AtomicVLA accepted to CVPR 2026!
• 2025.06 🎉 RoboPearls accepted to ICCV 2025!

📝 Publications

Technical Report

JoyAI-RA 0.1: A Foundation Model for Robotic Autonomy

Joy Future Academy, JD

Core Contributor, Technical Report

Project Paper

TL;DR: A foundation model for robotic autonomy developed at JD Explore Academy, covering architecture design, action head design, VLM-subtask planning, and RFT training.

CVPR 2026

AtomicVLA: Unlocking the Potential of Atomic Skill Learning in Robots

Likui Zhang, Tao Tang, Zhihao Zhan, Xiuwei Chen, Zisheng Chen, Jianhua Han, Jiangtong Zhu, Pei Xu, Hang Xu, Hefeng Wu, Liang Lin^†, Xiaodan Liang^†

First Author, CVPR 2026

Project Paper Code

TL;DR: A unified planning-and-execution VLA framework with Skill-Guided Mixture-of-Experts that unlocks atomic skill learning, achieving 10% improvement on LIBERO-LONG and 18.3% on real-world long-horizon tasks.

Under Review

E0: Enhancing Generalization and Fine-Grained Control in VLA Models via Tweedie Discrete Diffusion

Zhihao Zhan, Jiaying Zhou, Likui Zhang, Qinhan Lv, Hao Liu, Jusheng Zhang, Weizheng Li, Ziliang Chen, Tianshui Chen, Keze Wang, Liang Lin, Guangrun Wang^†

Under Review

Project Paper Code

TL;DR: A VLA model using tweedie discrete diffusion that formulates action generation as iterative denoising over quantized action tokens, achieving state-of-the-art across 14 environments with 10.7% average improvement.

ICCV 2025

RoboPearls: Editable Video Simulation for Robot Manipulation

Tao Tang*, Likui Zhang*, Youpeng Wen, Kaidong Zhang, Jia-Wang Bian, Xia Zhou, Tianyi Yan, Kun Zhan, Peng Jia, Hefeng Wu, Liang Lin^†, Xiaodan Liang^†

Co-first Author, ICCV 2025

Project Paper

TL;DR: A semantically enhanced editable 4DGS simulation framework for high-quality controllable robot manipulation video simulation, enabling automated video generation via natural language instructions.

Under Review

Human-Adapter: Personalized Multimodal 3D Human Generation with Gaussian Splatting

Likui Zhang, Zhibin Liu, Hefeng Wu, Zhenyu Xie, Ruomei Wang, Xiaodan Liang, Liang Lin^†

First Author, Under Review

TL;DR: A 3D human generation model that produces high-fidelity personalized 3D humans from a single face image, body parameters, and text description via adaptive Decoupled Cross-Attention and two-stage detail optimization.

📖 Educations

• 2024.09 - 2029.06

  Ph.D. in Computer Science and Technology, Sun Yat-sen University, Guangzhou.
  School of Computer Science | Advisors: Prof. Liang Lin & Prof. Xiaodan Liang
  Research: Embodied AI, Simulation Reconstruction, VLA Models
• 2020.09 - 2024.06

  B.Eng. in Intelligent Science and Technology, Sun Yat-sen University, Shenzhen.
  School of Intelligent Engineering
  National Encouragement Scholarship, University Second-class Scholarship

💻 Internships

• 2025.12 - 2026.05

  Embodied AI Algorithm Intern, JD Explore Academy, Beijing.
  Embodied Foundation Model Department — Designed JoyAI-RA 0.1 architecture, RFT training, and AtomicRL framework for long-horizon tasks.
• 2025.08 - 2025.12

  AI Algorithm Intern, Huawei IAS BU (2030 Lab), Shenzhen.
  R&D Management Department — Designed AtomicVLA framework and contributed to multimodal model EVA with D-GSPO training.

🏆 Awards & Patents

• National 2nd Prize, "XingZhi Cup" National AI Innovation Application Competition (Autonomous Driving Trajectory Prediction Track)
• Provincial 1st Prize, 27th China Robotics and Artificial Intelligence Competition
• National Invention Patent (under substantive examination): An Editable Robot Manipulation Simulation System and Method