Zhouheng Li

Zhouheng Li

Ph.D. student @ZJU.

Zhouheng Li

I am currently a 4th-year Ph.D. student in the College of Control Science and Engineering at Zhejiang University , Hangzhou, China, under the supervision of Prof. Lei Xie and Prof. Hongye Su. I was previously advised on energy model-based composition by Prof. Yilun Du from Harvard University. I am currently collaborating closely on car racing-related research with Dr. Mattia Piccinini and Dr. Baha Zarrouki from the AVS Lab, Technical University of Munich (TUM).
I am currently visiting the AutoMan Lab at Nanyang Technological University (NTU) since May 2026, focusing on VLA-enhanced multi-agent game planning for autonomous racing, under the supervision of Prof. Chen Lv.
✨ Enjoyments of life: 🎾 tennis, 🎹 piano, 🎲 board games (Seven Wonders: Duel, Splendor, etc), hiking, ping-pong, traveling. 🚀 I am on the job market now! Please contact me if you have any relevant positions or opportunities (📨 zh_li@zju.edu.cn).

🎯 Research

My research focuses on developing embodied robots that can make intelligent decisions and plan effectively in highly dynamic scenarios, particularly when operating at the limits of handling (📝 Publications). To this end, I actively integrate physics-informed generative models with model-based approaches, such as Model Predictive Control (MPC) and Model Predictive Path Integral (MPPI). Currently, I am especially interested in planning strategies for autonomous car racing and drone racing, with a focus on the following directions:
⭐ Game Planning for Competitive Multi-Agent Interaction: Game theory provides a principled framework for modeling competition over limited resources, yet key challenges remain in computational efficiency and safe strategy transitions. My research leverages generative models as priors to improve warm-starting and adaptation, enabling stable and robust motion planning in complex multi-agent environments.
⭐ Physics-Informed World Models and VLAs: A key limitation of current world models and VLAs is that their predictions and decisions often lack essential physical consistency. Integrating physics-informed neural networks (PINNs) into world models and VLAs can improve prediction consistency and action feasibility, enabling more robust long-horizon behavior.
⭐ Multimodal Generative Models: Integrating diverse sensory modalities while balancing multiple objectives remains challenging for robotic systems. This direction focuses on leveraging energy-based models and classifier-free guidance to enable flexible composition of image and state modalities at inference time, supporting efficient adaptation and strong few-shot and zero-shot generalization.
If any of these topics caught your interest, feel free to drop me emails. I enjoy collaborating on interesting projects and making amazing things happen together!

✨ Highlights

Demos showcasing the effectiveness of my related work across distinct challenges: (i) vision-based drone racing with neural signed distance fields; (ii) car racing with velocity-enhanced MPCC; (iii) energy-based model composition at inference time for safe planning; and (iv) physics-informed BEV world modeling for consistent long-horizon prediction. Click the buttons below to explore the details of each project!

Vision-based Drone Racing

Slow Motion replay
Orientation Mismatch replay
Position Mismatch replay
Real-time vision-guided drone racing with neural SDFs. Real-world trajectories on a circular track with perturbed gates, where the quadrotor navigates successfully using only onboard depth perception.
Web
PDF

Velocity-Enhanced Car Racing

racing
Implementation of the proposed EVO-MPCC planner on the Roboracer platform. The local planner achieves feasible and aggressive cornering through a sharp U-turn at 11.5 km/h.
Web
PDF
Code

Energy-based Diffusion Navigation

Static Model replay
Dynamic Model replay
Composed Result replay
This energy-parameterized diffusion composition guarantees collision-free trajectory planning in unseen scenes. Energy-parameterized diffusion learns a conservative energy field, enabling flexible integration of multiple constraints and generalization to previously unseen environments without retraining.
Web
PDF
MPC
ItCA

Physics-Informed BEV World Model

Baseline with Eat up (DIAMOND) replay
Baseline with Eat up (DIAMOND) replay
Soft Mask with Consistency replay
Soft Mask with Consistency replay
The Physics-Informed World Model is a lightweight generative model that predicts future images conditioned on the current scene and control inputs. By incorporating the relative distance between the ego vehicle and surrounding vehicles via a soft mask as an additional conditioning signal, it effectively captures inter-vehicle interactions, thereby enabling forecasting with strong existential and temporal consistency in dynamic environments.
Web
PDF
Code

🔥 News

Older news

📝 Selected Publications

    * means equal contribution.

  • EVO-MPCC: Enhanced Velocity Optimization with Learning-Based Auto-Tuning for Real-Time Vehicle Trajectory Planning

    Zhouheng Li, Bei Zhou, Mattia Piccinini, Cheng Hu, Baha Zarrouki, Rahul Mangharam, Lei Xie

    Code

  • Vision-Guided MPPI for Agile Drone Racing: Navigating Arbitrary Gate Poses via Neural Signed Distance Fields

    Fangguo Zhao, Hanbing Zhang, Zhouheng Li, Xin Guan, Shuo Li

  • Rapid and Safe Trajectory Planning over Diverse Scenes through Diffusion Composition

    Wule Mao*, Zhouheng Li*, Yunhao Luo*, Yilun Du, Lei Xie

    Code

  • A Data-Driven Aggressive Autonomous Racing Framework Utilizing Local Trajectory Planning with Velocity Prediction

    Zhouheng Li, Bei Zhou, Cheng Hu, Lei Xie, Hongye Su

    Code

  • Reduce Lap Time for Autonomous Racing with Curvature-Integrated MPCC Local Trajectory Planning Method

    Zhouheng Li, Lei Xie, Cheng Hu, Hongye Su

    Code

  • A rapid iterative trajectory planning method for automated parking through differential flatness

    Zhouheng Li, Lei Xie, Cheng Hu, Hongye Su

    Code

  • An Overtaking Trajectory Planning Framework Based on Spatio-temporal Topology and Reachable Set Analysis Ensuring Time Efficiency

    Wule Mao, Zhouheng Li, Entao Sun, Lei Xie, Hongye Su

  • An aggressive cornering framework for autonomous vehicles combining trajectory planning and drift control

    Wangjia Weng, Cheng Hu, Zhouheng Li, Hongye Su, Lei Xie

  • Adaptive Learning-based Model Predictive Control Strategy for Drift Vehicles

    Bei Zhou, Cheng Hu, Jun Zeng, Zhouheng Li, Johannes Betz, Lei Xie, Hongye Su

  • Enhancing Physical Consistency in Lightweight World Models

    Dingrui Wang*, Zhexiao Sun*, Zhouheng Li, Cheng Wang, Youlun Peng, Hongyuan Ye, Baha Zarrouki, Wei Li, Mattia Piccinini, Lei Xie, Johannes Betz

    Code

  • A Learning-based Planning and Control Framework for Inertia Drift Vehicles

    Bei Zhou, Zhouheng Li, Lei Xie, Hongye Su, Johannes Betz

  • Learning to Race in Extreme Turning Scene with Active Exploration and Gaussian Process Regression-based MPC

    Guoqiang Wu*, Cheng Hu*, Wangjia Weng, Zhouheng Li, Yonghao Fu, Lei Xie, Hongye Su

🏁 Competitions


💡 Services

  • The IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS’ 26).
  • IEEE International Conference on Robotics and Automation (ICRA’ 26).
  • Robotics and Autonomous Systems.
  • IEEE Robotics and Automation Letters.

🤗 So Glad You're Here!

I truly appreciate everyone who takes the time to visit my homepage :)