Abstract :
For a long time in human history, media such as books, music, and movies have played an important role in telling stories and conveying information among people. They are linear media, which are consumed passively without user intervention. Along with the proliferation of interactive platforms such as the web, augmented and virtual reality, the shift from linear media to non-linear media such as interactive installations, video games is taking place. Due to the dynamic nature of non-linear media, which allows user intervention in the middle of consuming content, the traditional methods for creating content are too costly and time-consuming to be used in developing non-linear applications. In this talk, I'll focus on how to design human-like virtual agents that can interact with users and easily adopt unseen environments for interactive content generation. More specifically, I'll introduce physically simulated agents and building control policies that can move the agents in a human-like manner by using deep reinforcement learning. If time allows, I'll also briefly introduce a newly released opensource research platform, fairmotion, that supports data loading, processing, modeling, and visualizing of motion data, which were developed for collaboration among content creators, computer graphics, and machine learning researchers.