Inverse Reinforcement Learning for Inferring Human-Centered Costmaps [report]
- Course: Computational Human Robot Interaction, Lecturer: Prof. Stefanos Nikolaidis, USC
- A scalable method to collect information on task-basedpersonal space.
- An algorithm for inferring a task-based costmap from adataset of motion capture trajectories of two humans.
Master Thesis: Optimal Navigation System for a Mobile Robot to Execute Dynamical Multiple Social Tasks [pdf]
- Advisor: Dr. Li-Chen Fu, NTUEE
- A navigation system integrated with perception module and decision making system can handle dynamical social tasks during human robot interaction.
- A model that formulates the household environment into discrete graph and task into time-decaying reward function can transfer complex social task planning problem to mathematical optimization problem.
- A task planning algorithm using relative low memory consumption in time such that the robot can deal with scenarios when newly-added social tasks are called from human.
Multi-Agent Path Finding with Simple Temporal Network with Uncertainty (STNU) [slides]
- Course: Coordinate Robotics, Lecturer: Prof. Nora Ayanian, USC
- Apply STNU to MAPF to deal with uncertainty to deal with uncertainty movements during plan execution.
Real-time Human, Object, Pose Detection for Service Robots [link]
- Course: GPU Programming, Lecturer: Prof. Wei-Chao Chen, NTUEE
- Developed a system for robot to execute commands given by human gesture and voice.
Face Recognition with Squeezed Deep Learning Model [poster]
- Course: Deep Learn for Computer Vision, Lecturer: Prof. Yu-Chiang Frank Wang, NTUEE
- Developed a face recognition network with squeezed model combing InceptionNet, Inception-ResNet, and ShuffleNet.
- Trained and tetsted on CelebA Dataset, comparing with VGG-16 with the testing accuracy about 0.6, the testing accuracy of our model is 0.65 with ten times fewer parameters.
Making Latte Art through ITRI Robotic Arm
- Course: Robotics, Lecturer: Prof. Li-Chen Fu, NTUEE
- Manipulated robot arm to generate latte art given a certain pattern while recognizing objects like cups, stir bar with top view camera.
Machine Translation for Native Taiwanese Language on Local Drama Conversation
- Course: Machine Learning, Lecturer: Prof. Hung-Yi Lee, NTUEE
- Developed a word-embedded model trained on Conversation Native Language Dataset with 0.51 accuracy, while the best model on Kaggle is 0.55.
Solving Marble Maze with Universal Robot UR5 [link]
- Course: Introduction to Robotics, Lecturer: Prof. Ruzena Bajcsy, Department of EECS, UC Berkeley
- Encoded the corner of the maze and solve the adequate path for the marble.