Hi, welcome to my homepage.

Biography

I am an M.S. student in Applied Mathematics at the School of Mathematics and Statistics, Xi'an Jiaotong University (XJTU). My research interests include World Action Models, Reinforcement Learning, and Embodied Intelligence.

I have worked on embodied reinforcement learning and world-model navigation at the Institute of Automation, Chinese Academy of Sciences (CASIA). At Meituan, I worked on large-scale pretraining for multimodal trajectory generation with Flow Matching, followed by post-training with Energy-Field guidance and reinforcement learning. At LinkerBot, I worked on DexLeWM post-training to adapt pretrained VLA/WAM capabilities to coordinated dexterous manipulation.

I am currently working on autonomous driving safety at Tuojing AI, jointly with MMLab, The University of Hong Kong, where I study generative world models and counterfactual causal reasoning.

I am interested in building unified world representations for robots. I welcome research discussions and academic collaboration; feel free to reach me by e-mail.

I am currently seeking research internship opportunities. If you think my background could be a good fit for your team or project, please feel free to contact me by e-mail.

News

[2026/05/30] Our paper DriveAnchor: Progressive Anchor-based Flow Learning for Autonomous Driving Planning is available on arXiv.
[2026/05/07] Our paper HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning is available on arXiv, with code released.
[2026/01/06] Our paper Spatial Sampling-based Passivity and Synchronization of Multiweighted Coupled Reaction-diffusion Neural Networks was accepted by CNSNS.
[2025/09/01] Our paper Adaptive Passivity-based Synchronization of Spatiotemporal Neural Networks with Multi-weighted Coupling under Spatially Point Measurements was accepted by CISC 2025.

Selected Publications

arXiv 2026
DriveAnchor: Progressive Anchor-based Flow Learning for Autonomous Driving Planning Limin Yan*†, Haoyun Tang*, Yutao Qiu, Hongqing Liu, Haoyu Xu‡ * Equal contribution. † Lead corresponding author. ‡ Co-corresponding author. arXiv:2606.00519, 2026
arXiv alphaXiv 15
```
@article{yan2026driveanchor,
  title={DriveAnchor: Progressive Anchor-based Flow Learning for Autonomous Driving Planning},
  author={Yan, Limin and Tang, Haoyun and Qiu, Yutao and Liu, Hongqing and Xu, Haoyu},
  journal={arXiv preprint arXiv:2606.00519},
  year={2026}
}
```

arXiv 2026

HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning Haoyun Tang*, Haodong Cui*, Keyao Xu, Kun Wang†, Zhandong Mei† * Equal contribution. † Corresponding authors. arXiv:2605.05951, 2026

arXiv alphaXiv 13 Code 7

@article{tang2026hamworld,
  title={HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning},
  author={Tang, Haoyun and Cui, Haodong and Xu, Keyao and Wang, Kun and Mei, Zhandong},
  journal={arXiv preprint arXiv:2605.05951},
  year={2026}
}

CNSNS 2026

Spatial Sampling-based Passivity and Synchronization of Multiweighted Coupled Reaction-diffusion Neural Networks Haodong Cui, Haoyun Tang, Mingyu Ma, Cheng Hu†, Tingting Shi † Corresponding author. Communications in Nonlinear Science and Numerical Simulation, 2026

PDF DOI

@article{cui2026spatial,
  title={Spatial Sampling-based Passivity and Synchronization of Multiweighted Coupled Reaction-diffusion Neural Networks},
  author={Cui, Haodong and Tang, Haoyun and Ma, Mingyu and Hu, Cheng and Shi, Tingting},
  journal={Communications in Nonlinear Science and Numerical Simulation},
  volume={157},
  pages={109666},
  year={2026},
  doi={10.1016/j.cnsns.2026.109666}
}

CISC 2025

Adaptive Passivity-based Synchronization of Spatiotemporal Neural Networks with Multi-weighted Coupling under Spatially Point Measurements Haoyun Tang, Haodong Cui, Mingyu Ma, Cheng Hu†, Tingting Shi † Corresponding author. Chinese Intelligent Systems Conference (CISC 2025), 2026

PDF DOI

@inproceedings{tang2026adaptive,
  title={Adaptive Passivity-based Synchronization of Spatiotemporal Neural Networks with Multi-weighted Coupling under Spatially Point Measurements},
  author={Tang, Haoyun and Cui, Haodong and Ma, Mingyu and Hu, Cheng and Shi, Tingting},
  booktitle={Proceedings of 2025 Chinese Intelligent Systems Conference},
  pages={266--276},
  year={2026},
  publisher={Springer Nature Singapore},
  doi={10.1007/978-981-95-6813-0_26}
}

Education

Xi'an Jiaotong University 2026 – Present

M.S. Student School of Mathematics and Statistics

M.S. in Applied Mathematics. Research Focus: World Action Models, Reinforcement Learning, and Embodied Intelligence.
Xinjiang University 2022.09 – 2026.06

Undergraduate Student School of Mathematics and System Science

B.S. in Mathematics and Applied Mathematics. University-level Outstanding Graduate. Academic Standing: Top 3%, recipient of three national-level awards and seven provincial-level awards, with two published papers.

Research & Industry Experience

Tuojing AI 2026.09 – Present (Remote)

Autonomous Driving Safety Research Intern Joint with MMLab @ HKU

Working on autonomous driving safety with generative world models and counterfactual causal reasoning, building on CounterScene to generate realistic yet adversarial safety-critical scenarios.
LinkerBot 2026.06 – 2026.09 (On-site)

Embodied AI Algorithm Intern Evo Lab · Beijing

Developing DexLeWM for multi-task fine manipulation with a 7-DoF arm and 16-DoF dexterous hand by adapting pretrained Being-H0.5-style arm-gripper VLA/WAM capabilities. Built action adapters, dimension masks, and decoupled Arm/Hand Experts with arm-to-hand attention and event gating. Aligned action representations with effect latents from future visual observations through Action–Future Effect Alignment and posterior–prior distillation, and incorporated multi-horizon chunks and episodic memory for robust real-robot control.
Meituan 2026.03 – 2026.06 (On-site)

Autonomous Driving Algorithm Intern Autonomous Delivery Business Unit · Beijing

Worked on multimodal trajectory generation for autonomous delivery vehicle motion planning with Flow Matching, Energy Field modeling, and reinforcement learning. Contributed to Guide Flow Decoder design, FM pretraining on 120M trajectories, and FM + EF post-training on 7.5M samples. Proposed Hybrid KNN + epsilon-ball sampling to mitigate mode collapse, tuned dynamics parameters, optimized Anchor sampling and inference, and participated in closed-loop evaluation and deployment. The deployed model converges within 4 inference steps and supports intersection straight-driving, left-turn, and lane-change scenarios.
Institute of Automation, Chinese Academy of Sciences 2025.11 – 2026.02 (On-site)

Reinforcement Learning Algorithm Intern Advisor: Prof. Zhiqiang Pu · Beijing

Optimized MARL algorithms for AirSim-based 3D pursuit-evasion tasks using a MAPPO pipeline for 3v1 and 5v2 settings with up to 80% success rate; built a DreamerV3 world-model framework for vision-only navigation with ArUco-guided target reaching and obstacle avoidance; and deployed embodied RL systems on RK3588 and ROS, including SAC-based autonomous obstacle avoidance on LingAo ground vehicles and a mocap-connected real-time perception-reasoning-control loop for Owl2Mini drones.

Selected Awards & Honors

Meritorious Winner, Mathematical Contest in Modeling (MCM), 2025 [PDF]
Provincial Special Prize, China Undergraduate Mathematical Contest in Modeling (CUMCM), 2023
Provincial First Prize, National College Student Mathematics Competition, 2024
Provincial Second Prize, National College Student Market Research and Analysis Competition, 2024
Provincial Second Prize, National Undergraduate Statistical Modeling Competition, 2024
National Second Prize, Asia and Pacific Mathematical Contest in Modeling (APMCM), 2024
National Third Prize, HuaShu Cup Mathematical Modeling Competition, 2024
Jinlongyu Academic Scholarship, 2025

Selected Projects

Car Obstacle Avoidance Navigation (DreamV3) [Demo] Keywords: DreamV3, Autonomous Navigation, Obstacle Avoidance, Reinforcement Learning Built a small-car navigation project using DreamV3-style world model reinforcement learning for obstacle avoidance and autonomous route following in a structured driving environment.
UAV Obstacle Avoidance Navigation (SAC) [Demo] Keywords: UAV, SAC, Obstacle Avoidance, Reinforcement Learning Built a UAV obstacle avoidance navigation experiment using Soft Actor-Critic to learn collision-aware flight behavior in a structured simulation environment.
Road2Gen-Drive (MetaDrive IL+RL / GAIL) [GitHub] Keywords: MetaDrive, Behavior Cloning, PPO, GAIL, Generalization Built end-to-end autonomous driving experiments in MetaDrive with two independent pipelines: IL+RL (BC pretraining + PPO fine-tuning) and GAIL for adversarial imitation learning and generalization evaluation.
Multi-UAV Collaborative Surrounding (MAPPO + AirSim) [GitHub] Keywords: AirSim, MAPPO, Multi-UAV, Curriculum Learning Built a 3v1 multi-UAV cooperative surrounding environment in AirSim and a MAPPO training pipeline with curriculum learning and tailored reward design.

Hobbies & Interests

Outside research, I enjoy suspense and science-fiction films. I also play basketball as a starting power forward, with two college tournament championships, and keep active through running, badminton, swimming, and hiking.