Prediction in Reinforcement Learning

This research project aims to build a prediction model that can generate future scenes to inform robot decision making, then use the prediction model as an extra training signal for the reinforcement learning agent, and to improve sample efficiency for real-world robotic tasks.

This research contributes to a key research area in computer vision; namely teaching robots/AI to understand the world through vision and interact with the world correctly as a result. It is important for the robot to understand the physical dynamics of our world in order for them to correctly predict the consequences of their actions. It is also important to model the dynamics through vision when building intelligent agents so that they can have a better understanding of the world.