Robotics researcher

LI ZHOUJIAN

MSc Robotics

National University of Singapore

Portrait of LI ZHOUJIAN

I study learning-based locomotion and perception for autonomous robots. My work spans humanoid and quadruped control, human motion retargeting, and autonomous navigation.

Research interests

View research

Email GitHub LinkedIn
Publications

Research output

Publications

Preprints & conference papers

arXiv preprint2025Featured work

DVDP: An End-to-End Policy for Mobile Robot Visual Docking with RGB-D Perception

H. Min†, Z. Li†, Y. Yang, J. Chen, S. Yuan

† Equal contribution

arXiv:2509.13024

An end-to-end RGB-D policy for mobile robot docking, with real-world deployment on SCOUT Mini.

  • Visual docking
  • RGB-D perception
  • End-to-end learning
Read on arXiv PDF 45.8 MB
DVDP architecture: RGB-D inputs, encoder, cross-attention, and trajectory decoder View figure
RGB-D perception → docking trajectory
Conference paper2025

A Reinforcement Learning-Driven Algorithm for Rapid Path Replanning of Robot Navigation in Indoor Uncertain Discrete Environments

H. Min†, Z. Li†, W. Chi

† Equal contribution

CCDC 2025 · 37th Chinese Control and Decision Conference

Reinforcement learning for efficient path replanning and adaptation to uncertain indoor environments.

  • Reinforcement learning
  • Path planning
  • Navigation
IEEE Xplore PDF 1.6 MB
Network architecture for robot path planning from a static map and start and goal positions View figure
Learning to replan in uncertain environments

From methods to systems

Research projects

Locomotion, motion imitation & navigation

PASSAGE: Scaling Scene-Aligned Motion Learning for Perceptive Humanoid Traversal in Cluttered Environments

Jun 2026 - Sep 2026

Egocentric perception · Whole-body motion planning · Perceptive tracking

A closed-loop system for humanoid traversal through cluttered spaces, connecting scene perception, whole-body motion generation, and control. Learned from roughly 100 hours of human demonstrations, it coordinates ducking, stepping over obstacles, and local route selection using egocentric sensing and onboard inference.

Cluttered-space traversal

Real world · Traversing a confined course with blocks, overhead constraints, and low barriers.

Real world

4 clips

Simulation

2 clips

For more details, visit the project page or read the paper on arXiv .

PACE: Perceptive humanoid locomotion

May 2026 - Sep 2026

Perceptive and Compliant Expert Locomotion

A depth-guided locomotion framework for the Unitree G1, combining learned foothold estimation, recurrent terrain memory, and a mixture-of-experts policy to traverse stairs, gaps, and raised platforms.

  • Unitree G1
  • 29 DoF
  • 4 policy experts
  • 50 Hz control
Outdoor stair ascentClimbing an outdoor staircase with depth-guided foot placement.

Hardware demonstrations

6 clips

Foothold visualization in simulation

Unitree G1 descending stairs in simulation, with foothold markers visualized on the terrain.
Method & implementation
  • Predictive perception. A mixture-density variational foothold estimator predicts contact locations, while an attention-based GRU retains local terrain geometry across depth frames.
  • Expert policy. Four softly gated experts are trained with PPO and terrain-conditioned adversarial motion priors. Walking style is distilled from a previously trained flat-ground policy.
  • Sim-to-real transfer. Calibrated depth-sensor models, firmware-informed actuator gain randomization, and compliance-oriented reward shaping address measured hardware mismatches.
  • Deployment. A shared sensor pipeline and a fused recurrent ONNX graph connect depth, IMU, and joint-encoder inputs to 29 joint-position targets at 50 Hz.

Current scope: real-robot demonstrations on structured terrain. The technical report identifies a remaining descent-energy gap between simulation and hardware; broader repeated-trial evaluation remains future work.

Video-to-robot motion imitation

May 2026 - Jun 2026

ViMoRe + MoTrace · Unitree G1

A connected workflow that reconstructs human motion from monocular video, retargets it to a 29-DoF humanoid, and trains a whole-body tracking policy in simulation.

  1. Monocular video
  2. SMPL-X motion
  3. G1 reference
  4. RL tracking
Tennis swingReconstruction & tracking
Human motion reconstructionVideo overlay & world-frame motion · GVHMR
G1 motion trackingLearned policy rollout · Simulation
Basketball danceReconstruction & tracking
Human motion reconstructionVideo overlay & world-frame motion · GVHMR
G1 motion trackingLearned policy rollout · Simulation
Method & implementation
ViMoRe
Integrated GVHMR reconstruction with GMR retargeting to produce G1 joint-motion references. Validated tennis, basketball dance, and break-dance clips, with clip-start trimming to handle initial-pose inverse-kinematics traps.
MoTrace
Converted references into training data with quaternion validation and forward-kinematics replay. Adapted BeyondMimic workflows to local motion files for training, checkpoint evaluation, and ONNX export.
Demonstrated
Per-motion tracking policies for tennis and basketball dance in simulation. Motion reconstruction currently targets static-camera input videos.

Built on GVHMR, GMR, and BeyondMimic.

Multi-goal mapping & navigation

Jan 2026 - Mar 2026

ROS / Gazebo · Dynamic indoor environments

A mobile robot navigation stack for sequential goal search in a simulated factory, integrating 3D mapping, localization, target detection, and global–local planning.

End-to-end navigation in simulationFull multi-goal navigation workflow, with the Gazebo scene, camera observations, and RViz navigation view shown together.
Mapping
Built 3D point-cloud maps and 2D occupancy grids, with AMCL and EKF-based localization to reduce long-term pose drift.
Navigation
Combined YOLO target detection with A* / TEB planning and layered costmaps for autonomous search and navigation across multiple goals.
  • ROS
  • Gazebo
  • AMCL / EKF
  • A* / TEB
  • YOLO

Single-leg handstand for quadruped robots

Aug 2025 - Nov 2025

Unitree Go1 · MuJoCo / MJX · Brax PPO

Learning to balance a quadruped on a single front foot. A custom Go1 task combines support-foot balance and foot-lift rewards with two-stage PPO training to develop and refine the skill in simulation.

Single-front-leg balanceLearned policy rollout · MuJoCo simulation · 0:10
Method & implementation
  • Task design. Extended the two-leg handstand environment with center-of-mass alignment over the support foot, free-foot lift and contact rewards, and pose regularization for the non-supporting legs.
  • Two-stage training. Learned the skill with PPO, then fine-tuned with energy, joint-acceleration, and action-rate penalties alongside domain randomization.
  • Sim-to-sim validation. Exported policy weights and observation normalization for standalone MuJoCo playback at 50 Hz. The validation setup supports standing, two-leg handstand, single-leg handstand, and crouched initial poses.
View repository

Built on MuJoCo Playground.

Autonomous mapping & navigation

Sep 2020 - Sep 2023

Ecovacs University–Enterprise Joint Project

Enabling an intelligent vehicle to map its surroundings and navigate around pedestrians.

Contribution
Combined pedestrian detection and tracking, trajectory prediction, and ROS-based path planning.
Outcome
Built a navigation pipeline with autonomous mapping, A* path planning, and OpenCV-based pedestrian posture recognition.
  • ROS
  • YOLOv4
  • OpenCV
  • A*
Technical details
  • Used convolutional neural networks to detect and track pedestrians in video.
  • Predicted pedestrian trajectories for real-time obstacle avoidance.
  • Worked with YOLOv4 for perception and ROS for mapping and autonomous path planning.

Offline reinforcement learning for cleaning robots

Sep 2022 - Feb 2023

Conservative Q-Learning · ROS Noetic & Gazebo

Learning floor-cleaning robot navigation from simulated and annotated trajectories.

Contribution
Created domestic simulation environments, generated A* trajectories, and annotated paths in RViz for offline learning.
Outcome
Built a Gazebo path-training dataset and tuned the conservative coefficient α in the CQL model.
  • Offline RL
  • CQL
  • Gazebo
  • RViz
Technical details
  • Implemented an offline reinforcement learning workflow using Conservative Q-Learning in ROS Noetic.
  • Constructed multiple domestic environments in Gazebo and generated path trajectories with A*.
  • Used RViz to inspect robot trajectories and manually annotate optimal paths.
  • Used the annotated data as a supervisory signal for model training.

Ultrasonic sensing for challenging surfaces

May 2023 - Aug 2023

Transparent & reflective object recognition

Recognizing transparent and highly reflective objects through ultrasonic sensor data.

Contribution
Modeled sensors in URDF and Gazebo, collected ROS data, and fitted Gaussian models using maximum likelihood estimation.
Outcome
Improved object recognition through Gaussian modeling and evaluated accuracy, recall, and F1 score.
  • ROS
  • URDF
  • Gaussian modeling
  • Sensor simulation
Technical details
  • Implemented URDF-based sensor emulation in Gazebo to approximate real-world sensor physics.
  • Adjusted Gazebo object properties to evaluate reflective and transparent surfaces.
  • Streamlined ultrasonic data collection, constructed Gaussian models, and estimated their parameters.
  • Evaluated recognition accuracy, recall, and F1 score to guide model tuning.

Academic background

Education

Aug 2025 - Jan 2027

MSc Robotics

National University of Singapore

Learning-based locomotion, motion retargeting and whole-body control, SLAM, and autonomous navigation.

Selected coursework

Robot Kinematics · Robot Dynamics and Control · Machine Learning

More coursework

Autonomous Robot Navigation · Robot Vision and Artificial Intelligence

Sep 2020 - Jun 2024

BSc Intelligent Manufacturing

Soochow University

3.5/4.0GPA · Top 15%

86/100Average score

  • First-Class Innovation Scholarship Top 2%
  • Second-Class Academic Excellence Award Top 13%

Industry experience

Internships

May 2026 - Sep 2026

Humanoid Locomotion Algorithm Intern

Galbot

  • Developed reinforcement-learning locomotion policies and supported real-robot deployment, from simulation training to on-board inference and control.
  • Combined proprioception and terrain observations with AMP motion priors, terrain curricula, and domain randomization for walking and stair traversal.
  • Integrated LiDAR-based elevation mapping on Jetson AGX with an exported ONNX policy in a 30 Hz control loop.

Feb 2025 - Jun 2025

Algorithm Intern, Science & Technology Development Department

Peking University Shanghai Lin-gang International Science and Technology Innovation Center

  • Deployed large language models locally for on-premises inference.
  • Fine-tuned large language models for safety alignment.

Jun 2023 - Aug 2023

Assistant Navigation Algorithms Engineer

Suzhou Suxiang Robot Intelligent Equipment

  • Built simulation maps in Ubuntu.
  • Worked on local path planning through A* algorithm optimization.

Tools & methods

Technical Skills

Learning & control

  • PPO / AMP / CQL
  • Mixture-of-experts policies
  • Reward design
  • Domain randomization
  • LLM safety alignment fine-tuning
  • PyTorch / JAX / Brax

Related projects

PACE locomotion Go1 balance control Offline RL navigation

Perception & motion

  • Depth perception & GRU memory
  • SMPL-X motion reconstruction
  • IK-based motion retargeting
  • Whole-body motion tracking
  • OpenCV / YOLOv4
  • SLAM / AMCL / EKF
  • Global & local path planning

Related projects

PASSAGE traversal ViMoRe + MoTrace Multi-goal navigation Mapping & navigation

Simulation & deployment

  • Isaac Sim / Isaac Lab
  • MuJoCo / MJX
  • ROS / Gazebo / RViz
  • URDF sensor modeling
  • ONNX policy deployment
  • On-premises LLM deployment
  • Sim-to-sim / sim-to-real

Related work

Go1 single-leg handstand Galbot deployment PKU-IICSH LLM deployment Ultrasonic sensing

Programming & tools

  • Python
  • C / C++
  • MATLAB / Simulink
  • Ubuntu

Languages

Chinese Native

English Proficient

Awards & honors

2023

  • National First Prize

    RAICOM Robotics Developer Competition

  • National Third Prize

    16th China College Students Computer Design

  • Provincial Second Prize

    RAICOM Jiangsu Regional Contest

2022

  • National Second Prize

    24th China Robotics and AI Competition

  • Provincial First Prize

    13th Jiangsu Province Robotics Competition

  • Second Prize

    16th Siemens Cup Intelligent Manufacturing

  • Provincial Third Prize

    10th National Optoelectronics Design Competition

Get in touch

Let’s talk robotics.

Open to research collaborations, opportunities, and conversations about robotics and AI.

zhoujianli01@u.nus.edu
GitHub LinkedIn