Embodied object hunt

• Main Author: Yeo, Zhi Hong
• Other Authors: Cham Tat Jen
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence; Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/137914
• Institution: Nanyang Technological University

Embodied Question Answering is an Artificial Intelligence task that includes many sub-tasks. One of which is goal-driven navigation. The objective of this study is to determine the suitability of an algorithm for training an agent to perform goal-driven navigation tasks on the Habitat platform, using Deep Reinforcement Learning and Imitation Learning. Deep Reinforcement Learning, in particular Deep Q-Network is used in conjunction with Imitation Learning, primarily for reward shaping and behaviour cloning, to train the agent. The hyperparameters present in these methods are also fine-tuned to produce the best performing model. All of this will be run on the platform, Habitat. It is a platform that consists of habitat-api, mainly for the back-end logic of Artificial Intelligence tasks, and habitat-sim, mainly for the rendering and simulating of 3D environments and agents. The results of this study can act as a litmus test for using Deep Q-Network with Imitation Learning for goal-driven navigation tasks, as well as future works on navigation tasks related to Embodied Question Answering.