News

• [2026.08] 🤖 I started my master's degree at UC Berkeley and my internship at BAIR.

• [2026.06] 🏆 Our team won First Place in the Flexible Manipulation Track at ICRA 2026 WBCD Challenge, Vienna, Austria.

• [2026.03] 🎉 I joined JD Explore Academy as an Embodied AI Algorithm Intern, Beijing, China.

• [2025.07] 🎉 I was honored to receive First Prize in the Embodied AI Bootcamp organized by HKCRC, Hong Kong, China.

• [2025.04] 🏆 Our team won Third Prize in the DJI Technology RoboMaster Competition 2025, Shandong, China.

• [2024.12] 🎉 I joined the Institute for AI Industry Research, Tsinghua University as an Intern, Beijing, China.

• [2024.06] 🏆 Our team won Second Prize in RoboCup@Home 2024, Fujian, China.

Research Experience

My research focuses on skill learning for robotic dexterous manipulation and policy post-training with reinforcement learning.

Conducted research on 3D/4D world models and vision-language-action models for embodied AI.

Developed LiDAR–inertial–visual SLAM methods and explored tactile sensing and thermal imaging for robot learning.

Publications

* : Equal contribution; †: Corresponding author(s)

We propose a framework that uses reinforcement learning in simulation to transform low-quality human-video priors into executable robot demonstrations. We evaluate our approach on four bimanual dexterous manipulation tasks using Dexmate dual arms and Sharpa dexterous hands.

We introduce PRISM, a simple, task-agnostic framework. Attached to a frozen JEPA encoder, a lightweight MLP predicts a Gaussian action prior. At planning time, a closed-form Product-of-Gaussians update integrates this prior to guide sampling based on confidence.

We introduce DexNinja, a Real2Sim2Real data engine for robust bimanual dexterous food cutting. DexNinja generates physically consistent data with deformable cutting simulation, and trains a stage-conditioned vision-tactile ACT policy with decoupled arm/hand actions and adaptive action chunking.

We propose PL-LIT, a general LiDAR-Inertial-Thermal SLAM system. PL-LIT combines an online photometric calibration module with a deep neural network for point-line feature extraction. For state estimation, we design a tightly coupled LiDAR-Inertial-Thermal formulation within an Error-State Iterated Kalman Filter. In addition, PL-LIT builds a probabilistic thermal-intensity voxel map.

This paper proposes a consistent and tightly-coupled visual-inertial-ranging odometry system based on the Lie group.

This paper proposes a robust online calibration method for UWB anchors in Visual-Inertial Navigation Systems.

Research Projects

Our team won first place in the Flexible Manipulation Track at ICRA 2026 WBCD Challenge, held in Vienna, Austria. Our approach is based on fine-tuning a pre-trained vision-language-action (VLA) model with reinforcement learning. Ultimately, the model achieved a high success rate on the deformable/flexible clothing tasks.

This project explored two approaches to post-training robotic models. First, we used the DAgger framework with positive and negative labels to post-train a VLA model. Second, we used sparse task-success rewards and an off-policy actor-critic framework to fine-tune a policy in simulation.

Developed an intelligent home service robot using a Kinova Gen2 arm and an Azure Kinect camera on ROS Noetic, with functions including guest reception and guidance, voice interaction, object recognition, and room cleaning. Used Gemini, SAM2, and GraspNet to estimate object pose and grasp objects. Controlled the arm using Actionlib and MoveIt. Applied FAST-LIO2 with Livox MID-360 for indoor mapping and localization.

Building on KungfuBot, we used inverse kinematics retargeting and reinforcement learning to enable humanoid robots to imitate highly dynamic behaviors (e.g., kungfu, dancing). Developed sim2sim in MuJoCo and sim2real with Unitree SDK, successfully deploying the framework on the Unitree G1 robot.

Used PnP to calculate the 3D position of the armor plate, and used Kalman filtering for tracking. The robot used four M3508 motors for the chassis and two M6020 motors for the turret, controlled via CAN communication and PID algorithms for current, speed, and position loops. Used feedforward PID to improve the gimbal drive and achieve gravity compensation.

ICRA 2024 RoboMaster Sim2real Challenge: In the organizers' Docker-based Habitat-Sim environment, we used ArUco markers and solvePnP to estimate box poses. We deployed an extended Kalman filter with an omnidirectional motion model for state estimation, using IMU data, odometry, and laser scans converted from depth images.

Honours & Awards

• 2026 Beijing Outstanding Undergraduate Graduate
• 2025 First Prize in the Embodied AI Entrepreneurship Bootcamp organised by Hong Kong Center for Construction Robotics
• 2025 Third Prize in DJI Technology RoboMaster University League Competition 2025, China, Shandong Province
• 2024 Xiaomi Scholarship
• 2024 Second Prize in RoboCup@Home Competition (leader)
• 2024 Second Prize in China iCAN Innovation Contest (leader)
• 2023 Beijing University of Chemical Technology Excellent Student Leader
• 2021 First Prize in DJI Technology RoboMaster Youth East China League Competition

Leadership & Activities

• Leader of the SIE Robotics Club
• Signed Photographer for 500px and Visual China
• Leader of the Comprehensive Office, BUCT Student Union
• Member of BUCT SIE Basketball Team
• Teaching Assistant of Robotic Workshop at BUCT

Website template adapted from here.