I am a master student at Columbia University. My major is Mechanical Engineering. I started working at Creative Machine Lab in the 21-fall semester. From Sep 2021 to August 2022, I worked with Yuhang Hu on the AI Face project, advised by Professor Hod Lipson. Currently, I am working on applying Neural Field models to robot arm 3d morphology learning. Before coming to Columbia University, I graduated from Huazhong University of Science and Technology.
Eduaction:
Columbia, MS in Mechanical Engineering, GPA 3.99 / 4.3
Huazhong U of Science and Technology, BS in Mechanical Engineering, GPA 3.72 / 4.0
Research Interests: robotics, machine learning.
Yuhang Hu, Boyuan Chen, Jiong Lin, Yunzhe Wang, Yingke Wang, Cameron Mehlman, Hod Lipson
Human-Robot Facial Co-expression, In Preparation at Science Robotics.
1. Hardware (based on previous generations): using Solidworks and Blender for the structure design, using 3d printing materials and aluminum profiles for the construction.
2. Coding: Using Ros2, control different modules in parallel, including cameras, learning models, eyeballs, eyelids, mouth and jaw, and neck. I also wrote a tracking and attention demo. The whole system was coded on Nvidia Jetson Xavier.
3. Learning models: Training the learning model mapping from facial image to robot motor control
1. Forward model (self model): Training the model to predict the next state with current state and action. Then, use the self-model and different reward function to select different gaits.
2. Training data sampling: Testing different data sampling methods to improve training efficiency.
Description: Robotics studio (MECE4611) project. Designed a parallel legged robot. Controled with liber Potato. 1. Set up simulation in Mujoco. 2. Training a Model to represent the inverse kinematics. 3. Design a elliptical gait and find the motor positions using the IK model. 4. Simulate the elliptical gait in Mujoco.
Description: Robotics studio (MECE4611) project. Designed a parallel legged robot. Controled with liber Potato.
Description: Robot learning course (MECE6616) project. Using different methods for the 2-dof and 3-dof arm's torque control task: 1. Model predictive control, 2. Deep Q-learning, implemented with pytorch. 3. PPO, using Openai Gym and Stable-Baselines3.
Description: Evolution algorithm (MECS4510) final project. Using evolution algorithm to train a group of soft robots walking. Co-evolution method has also been used to evolve the robot shape and the controller at the same time. The mass-spring physical simulator was coded using c++ and OpenGL.
Description: Travel sales man problem, 1000 points, using genetic algorithm.
Description: Using genetic programming to solve the symbolic regression problem (find the math expression that fits the given data points).
Description: In this project, we compared four different learning methods, including DNN, CNN, and Res-net, training the handwriting recognition models. The datasets we used are the EMNIST(letters included) and the MNIST(only digits)
1. Hardware: Designing a parallel sturctured robot manipulator that has three degrees of freedom and a suction cup.
2. Coding: The manipulator was controlled through G-code. Using OpenCV for the Chess pieces movements detection. Using Alpha-Beta Pruning algorithm for the AI Chess playing.