State Representation Learning in Robotics: Using Prior Knowledge about Physical Interaction

摘要

State representations critically affect the effectiveness of learning in robots. In this paper, we propose a roboticsspecific approach to learning such state representations. Robots accomplish tasks by interacting with the physical world. Physics in turn imposes structure on both the changes in the world and on the way robots can effect these changes. Using prior knowledge about interacting with the physical world, robots can learn state representations that are consistent with physics. We identify five robotic priors and explain how they can be used for representation learning. We demonstrate the effectiveness of this approach in a simulated slot car racing task and a simulated navigation task with distracting moving objects. We show that our method extracts task-relevant state representations from highdimensional observations, even in the presence of task-irrelevant distractions. We also show that the state representations learned by our method greatly improve generalization in reinforcement learning.