Kaixin Chai

Hi, I'm Kaixin. I graduated with a Bachelor's degree in Engineering from Xi'an Jiaotong University in 2022. After graduation, I began my journey in robotics.

Currently, I am a PhD student at KAIST, Seoul, advised by Joseph J. Lim. My research interests focus on mobile robot interaction with environments, particularly skill learning and data sources. Recently, I have been exploring methods for learning interaction skills from human videos.

Before that, I worked on robot navigation, control, and perception under the guidance of Fei Gao. I am deeply grateful for the guidance from my mentors and the companionship from my labmates throughout my research journey.

CV  /  Bilibili  /  Github  /  Email

Project

Below are some of the projects I have been involved in, listed in chronological order. Research-oriented projects are highlighted with a yellow background, while engineering projects are presented with a white background.

Visual perception provides the foundation for humanoid robots to perform intricate interactions in complex environments. This project introduces a new paradigm to learn visuomotor whole-body skills for humanoid robots from human video demonstrations.

Develop N2M, a transition module that optimizes robot positioning for mobile manipulation tasks, substantially enhancing task success rates through ego-centric observation-based guidance that generalizes across diverse environments with high data efficiency.

Developing an efficient trajectory optimization method for tractor-trailer robots that handles complex kinematics and deformable structures while achieving superior planning speed and path quality. Validated in real-world transportation tasks.

Develop LEMON-Mapping, a loop-enhanced framework for multi-robot point cloud fusion that resolves map inconsistencies through robust loop processing and spatial bundle adjustment.

Consistent mapping and localization are crucial for multi-robot tasks. In this project, We developed a lightweight multi-robot localization framework, comprising: GICP module for inter-frame estimation; BTC descriptors for loop detection; GTSAM for constructing loop factor graphs. The system supports intra-loop and inter-loop closures. To address communication instability, a prior loop mechanism enables single robots to close loops with a pre-existing map. This system can efficiently run on Orin-NX using only two cores.

Build the hardware for a mobile manipulator and develop Inverse Kinematics code to control the arm's movements. Execute long-horizon tasks using pre-trained skills. Additionally, I adjust the robot's initial position at the beginning of each sub-task to better align observations with the skills' training data distribution.

Set up a mathematical model for ground effect prediction based on flown field simulation and experiments, making the drones fly safer and more stable near the ground.

Enable a wheel-legged robot equipped with LiDAR and cameras to identify and follow specific targets.

Try to improve LiDAR-based localization accuracy and stability via motion planning. We derive a new metric for localization reliability through perturbation analysis and apply the metric to motion planning, which enables the robot to avoid degraded areas in advance.

Deploy obstacle avoidance algorithms to enable the drone to fly safely in complex environments.

We propose a planning framework for wheeled robot movement on uneven terrain, which is efficient and allows robots to move safely.

Design a Helium-Assisted drone for floating advertisements in shopping malls. The buoyancy generated by the helium balloon offsets 95% of the weight, extending the flight time from five minutes to one hour.

Demonstrate a novel energy generation system that maintains the turbine at optimal power conversion efficiency in changing water flow.

Design a new thermal simulation device to generate an even temperature field. To obtain better control strategies, we model the whole heat transfer process and utilize deep reinforcement learning technology.

Analyze the cluster effect of Savonius wind turbines using the variable-speed control method.

We designed a system for a metal processing plant to monitor the production quality of products. With the help of industrial cameras and convolutional neural networks, we achieved a recognition recall rate of 94%.

Thanks for the website template offered by Jon Barron.