Vision-based Robotic Chess Player

Abstract

A collaborative robot is a type of robot built to interact with humans in a shared workspace physically. This thesis will introduce a system that manipulates a collaborative robot to play a chess game with a human. This system heavily relies on computer vision to plan the robot's action. The thesis will present a workflow of pose estimation about a chessboard's position to the robot base and the process of using transfer learning to develop a pre-trained neural network to detect the chess pieces. The detection of the chess piece is based on images collected in real-time, and one major factor that influences the picture is the lighting condition. Some suggest that data augmentation can improve the neural network's ability against light variation. This thesis compares the performance of neural networks trained with the computer augmented dataset and real collected dataset, and studies the data augmentation effects about the neural network. Results will suggest that the real collected dataset is still better than the computer augmented dataset, and data augmentation can improve the neural network's performance but has limitations.