Course Projects

Nonlinear Optimization course independent project.

 

Control Lyapunov function (CLF) can give exponential stabilizing performance for a controller, a typical way to design such a controller in a minimum effort manner is known as min-norm controller. The application can stablize the movement of a complicated system, reaching to a goal at finite time etc. In a robot system where the control input is discretized, a smooth trajectory is not guaranteed when the updating time step is large.

​

In this project, inspired by the solving process of trust region method, a controller that can smooth the zig-zag trajectory of the min-norm controller is proposed. A case study on range-based homing problem is performed and it demonstrate that the proposed controller can converge much faster and smoother.

Reinforcement Learning course project in a group of three.

​

In this project,  we propose a model-free method to solve this path planning problem under a partially observable environment or even a dynamic environment. We show that through reinforcement learning, we can get a policy which can make the robot navigate under dynamic environment in a collision free way without building a map. We shown that without any hand design policy, by using the state-of-art deep reinforcement learning, we are able to train an end-to-end path planing algorithm that is robust to obstacle number, target size, and map size in either dynamic or static environment. We also provide a flexible reinforcement simulation environment that can be widely used for further development.

Image Processing course project in a group of four.

​

By using OpenCV and Python, we designed a cartoonizer which is capable of converting a real-life image to painting texture, a face morpher to produce a video showing the face morphing process from a custom image to the template image, a comic character converter which can convert an uploaded image to a chosen comic-character-like, a face exchanger to exchange the face in custom image with the one on the stage.

Deep Learning course project in a group of three.

​

The project explores how convolutional neural networks (ConvNet) can be used to identify series of digits in natural images from the Street View House Numbers data set. It is a challenging problem due to variability of the length of digits and the need of digits segmentation. We settled the problem by image cropping according to the annotated bounding boxes and adding multiple parallel softmax layers in the model. Finally we achieved 92.5% test accuracy with local contrast normalization technique.