Recently, there have been several deep reinforcement learning based visual attention models. However, these models are highly sample inefficient and impractical for real-world implementation. In this study, we use Virtual Reality as an intuitive interface for Imitation Learning to supervise a deep neural network based visual attention model. We demonstrate that few minutes of interaction data is sufficient to train the agent to perform a ball gathering task in a 3D environment making it suitable for practical applications.
Nishanth Koganti, Abdul R. A. Ghani, Yusuke Iwasawa, Kotaro Nakayama, Yutaka Matsuo: “Virtual Reality as a User-friendly Interface for Learning from Demonstrations.” Demonstrations Track, Conference on Human Factors in Computing Systems, (CHI). Montreal, Canada, April 21-26, 2018.