The Maryland Virtual Demonstrator Environment for Robot Imitation Learning

Abstract

Robot imitation learning, where a robot autonomously generates actions required to accomplish a task demonstrated by a human, has emerged as a potential replacement for a more conventional hand-coded approach to programming robots. Many past studies in imitation learning have human demonstrators perform tasks in the real world. However, this approach is generally expensive and requires high-quality image processing and complex human motion understanding. To address this issue, we developed a simulated environment for imitation learning, where visual properties of objects are simplified to lower the barriers of image processing. The user is provided with a graphical user interface (GUI) to demonstrate tasks by manipulating objects in the environment, from which a simulated robot in the same environment can learn. We hypothesize that in many situations, imitation learning can be significantly simplified while being more effective when based solely on objects being manipulated rather than the demonstrator's body and motions. For this reason, the demonstrator in the environment is not embodied, and a demonstration as seen by the robot consists of sequences of object movements. A programming interface in Matlab is provided for researchers and developers to write code that controls the robot's behaviors. An XML interface is also provided to generate objects that form task-specific scenarios. This report describes the features and usages of the software.