Tianyi Gu

Coursework Projects

• In CS880 Introduction to Mobile Robotics, my final project studies the problem of autonomous mapping. We implement the frontier-based exploration algorithm combined with the occupancy grid mapping technique that enables a Turtlebot robot to build a map autonomously for an unknown environment. We applied the theory of Bayesian inference to update an occupancy map. We also used the frontier based exploration algorithm to navigate the robot to unexplored areas on the map. The experiment results show that the robot can map the environment fully autonomously, both in simulation and in real-world settings. Our source code is here. We also took a video that records the process of a Turtlebot mapping a corner area in Kingsbury.

• In CS980 Topics in Reinforcement Learning, my final project study the problem of dynamic obstacles avoidance for mobile robots. We studied two deterministic approaches which use heuristic search techniques, and four stochastic approaches which use reinforcement learning techniques. We proved these two type of approach are mathematically different. The experiment results show that deterministic approaches are not only faster but also robuster than stochastic approaches. But stochastic approaches still applicable for certain problem scenarios. All of the source codes can be found here.

• In CS980 Planning for Robots, my final project design two control algorithm: sampling based model-predictive control (SBMPC) and bisection search based model-predictive control (BBMPC). The algorithms are implemented as the controller for a real-time planning system in ROS to enable a Pioneer robot to move quickly in environments with dynamic obstacles. The behaviors of both algorithms are demonstrated through straight and curve line following experiments from simulation and real world environments. We also discussed several issues of the real-time planning system. All of the source codes can be found HERE.

• In CS830 Introduction to Artificial Intelligence, my final project present a new anytime motion planning approach called B-SST. B-SST first runs BEAST, an effort-aided planner, to find a first solution as quickly as possible, then switch to another motion tree growth process called SST-with-cost-pruning, which adopt both idea from SST and cost pruning algorithms. We first introduce several related work that we build upon. Then B-SST is described in detail. Results with a variety of vehicles and environments showed that B-SST is competitive compared to A-BEAST and other successful anytime planners. We also discussed a more sophisticated idea on how to create a better anytime motion planner in the end. All of the source codes can be found here.

• In CS880 Introduction to Information Retrieval, my final project study the task of helping an automated player win a computer game by reading a strategic user's guide designed for human players. In complex computer games such as Star Craft, War Craft, and Civilization, finding a winning strategy is challenging even for humans. Therefore, human players typically rely on manuals and guides. Recently, researchers have tried to using such textual information to train an automated playe. Our goal, is to better understand the retrieval model used in their paper in terms of algorithms and metrics from the field of information retrieval. Our results provide some evidence that inverse document frequency out performs recurrent neural networks at assisting human players, and could be used as a baseline for evaluating retrieval models used in playing games like this. All of the source codes can be found here.

• In CS880 Probabilistic AI and Machine Learning, my final project compared the accuracy and reliability of several different classifiers for recognizing handwritten digit, training on MNIST dataset. Classifiers such as K-Nearest Neighbors, Decision Tree, and Random Forest are applied to the problem, and it turns out they both have their pros and cons. All of the source codes can be found here.

• In CS980 Topics in Multi-Agent and Multi-Robot Systems, my final project was an MDP based approach to slove the Quay Crane Scheduling Problem under Uncertainty in Container Terminals. The number of time-conflict tasks for yard cranes (YC) in yard operation is considered as one of the important measures for the evaluation of the level of fluency for quay crane scheduling by terminal experts. In this project, an MDP model for QCSP is developed. A UCT algorithm is designed to solve the model.