Integrated Platform for Understanding Physical Prior & Task Learning

Abstract

Recently, many reinforcement learning algorithms within the field of robotics have demonstrated considerable performance in multiple physical environment tasks. However, their learning patterns are very different from those of humans. Humans develop their prior knowledge about the physical world and utilize it in task learning to learn effectively. On the other hand, in the case of general machine learning algorithms, tasks are performed without prior knowledge, thus creating a difference between humans and robots in their initial stages of learning. In order to reconcile this difference, it is necessary to study the learning and utilization of prior knowledge in reinforcement learning algorithms. To accomplish this, we propose a platform that integrates prior knowledge learning into task learning environments, and then we show configuration and application examples to emphasize the necessity and usability of this platform.