Zhengxiao Han

I am a Master Student (MSR Program) at Northwestern Univeristy.

I spent a gap year as a Research Assistant at PKU-Agibot (智元机器人) Lab, Peking University, under the supervision of Prof. Hao Dong. Before that, I spent two years at DISCOVER Lab, Tsinghua University as a Research Intern, under the supervision of Prof. Hao Zhao and Prof. Guyue Zhou.

I received my Bachelor's degree in 2023 at BUCT, where I used to lead the SIE Robotics Club, a student organization focusing on general-purpose service robots.

Proudly, I am a core contributor to Mini Pupper, an open-source quadruped robot.

I am actively seeking Co-op intern positions in Robotics for Summer 2025.

LinkedIn  /  GitHub  /  X  /  Google Scholar  /  Email

🔥 News

• [2024.09] 🎓 I started Master in Robotics (MSR) program at Northwestern University.

• [2023.09] 💼 I started my gap year as a research assistant at PKU-Agibot (智元机器人) Lab, Peking University.

• [2023.07] 📚 Two papers get accepted to CICAI 2023, with a Best Demo Award.

• [2023.07] 📚 One paper gets accepted to ICCV 2023.

• [2023.06] 🎓 Get my B.Eng. in Mechanical Engineering at Beijing Univeristy of Chemical Technology.

• [2022.10] 🦾 Improved Mini Pupper 2 and launched it on Kickstarter.

• [2022.06] 📚 Published an AWS Robotics Blog in deploying applcation changes from cluod-based simulation to real robot hardware.

• [2022.05] 🏆 We get a Third Prize in ICRA 2022 RoboMaster University Sim2Real Challenge.

• [2021.10] 💼 I started my roboitcs journey as an intern at Institute for Al Industry Research (AIR), Tsinghua University.

• [2021.09] 🦾 Our open-source quadruped robot Mini Pupper is launched on Kickstarter, and crowd-funded about $500, 000.

• [2021.05] 🏆 We get a Second Prize in RoboCup@Home 2021, China.

📚 Publications

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

We developed an intelligent desktop operating robot designed to assist humans in their daily lives by comprehending natural language with large language models and performing a variety of desktop-related tasks.

Cooperated with Baidu Apollo and MARS Lab at Tsinghua University. We propose to introduce a tailored Model Predictive Control (MPC) module as a rescue into the state-of-the art interactive trajectory prediction model M2I. Notably, our method can effectively address the Out-of-Domain (OOD) problem in long-term simulation by enforcing a flexible regularization that admits the replayed data, while still enjoying the diversity of data-driven predictions.

We present a large-scale interactive trajectory prediction dataset named INT2 for INTeractive trajectory prediction at INTersections. INT2 includes 612,000 scenes, each lasting 1 minute, containing up to 10,200 hours of data. We benchmark the best open-sourced interactive trajectory prediction method on INT2 and Waymo Open Dataset, under in-domain and cross-domain settings.

🦾 Projects

I have always been working on Robotics.

I generated a synthetic dataset by Utilizing Isaac Sim and OmniObject3d. By extracting the model's pointcloud, I used RANSAC to fit the planes, then calculate the normal vectors, and finally used convex hulls to find areas that can place other objects. With this pipeline, I generated more than 300,000 data samples including RGB, Depth, and Segmentation information for training a grasping policy. The whole data generation process was done by me independently.

My MSR journey at Northwestern University kicked off with an intense two-week hackathon. We firstly reviewed BFS, DFS and RRT path planning algorithms. The most complex challenge was designing a robotic grasping system. I used HSV filtering to create a mask for detecting a purple pen, and by leveraging the camera's intrinsic matrix and aligned depth image, I calculated the pen's 3D position. To handle camera movements, I used an ArUco tag to dynamically calculate the transformation between the camera and robotic arm, making my system robust to any accidental camera shifts. At last I can generate a 6D grasping pose for the robotic arm.

Constructed a 7-DOF humanoid robotic arm with dexterous hands equipped with tactile sensors for data collection, and interfaced it with ROS Control and MoveIt. All work was done by me independently. We will work on encoding visual, tactile (not vision-based) and joint signals for imitation learning and deploy learned policies on real robots.

Achived hardware design of a new 6-DOF robotic arm, and interfaced all hardware under ROS Control and MoveIt. Interfaced the robotic arm with GSNet and achived a grasping demo. All work was done by me independently.

Achived hardware design of a 6-DOF robotic arm and a 2-finger adaptive parallel gripper (same capability as Robotiq 2f-85 Gripper). Constructed a robotic arm SDK with CMake, and interfaced the lib with ROS and Isaac Sim.

Inspired by ALOHA, we proposed an improved bimanual demonstration system. Constructed the SDK using C++ and CMake. Utilized KDL to solve kinematic and dynamic problems, then achived gravity compensation based on inverse dynamics. Designed a parallel two-finger gripper based on rack-gear structures.

Developed all the ROS 2 (Humble) software suite including Locomotion, SLAM, Navigation, and CV functions. Deployed Cartographer with an Extended Kalman Filter (EKF) to fuse IMU data and LiDAR odometry, enhancing the robostness and accuracy of Mini Pupper's localization system.

Utilized Pure Pursuit algorithm for multi-robot formation control, which was tested in both real world and Isaac Sim. Utilized an EKF-based 2D LiDAR localization system as the odometry. Utilized KDL for solving kinematic problems of our own 6-DOF robotic arm, and combined it with the odometry, achieving Chiken-Head Mode. Combined all the functions above, and constructed a multi-robot cooperative system to carry a box in Isaac Sim.

Deployed an Extended Kalman Filter (EKF) alongside an omnidirectional motion model for state estimation using sensor data, including IMU and odometry. Utilized A* algorithm for global path planning and Timed Elastic Band (TEB) algorithm for local path planning. Utilized ArUco library for detecting boxes' poses.

Inspired by Stanford's open-source quadruped robot Pupper, I designed my own mechanical hardware, mainly improved its leg structures. After that, I was contacted by Mini Pupper's team and joined them. We developed the first product together. I independently developed all the ROS software suite including Locomotion, SLAM, Navigation, and CV functions. We crowd-funded $500,000 on Kickstarter when I was a junior.

Miscellaneous

Design and source code from Jon Barron's website.