A virtual demonstrator environment for robot imitation learning

摘要

To support studies in robot imitation learning, this paper presents a software platform, SMILE (Simulator for Maryland Imitation Learning Environment), specifically targeting tasks in which exact human motions are not critical. We hypothesize that in this class of tasks, object behaviors are far more important than human behaviors, and thus one can significantly reduce complexity by not processing human motions at all. As such, SMILE simulates a virtual environment where a human demonstrator can manipulate objects using GUI controls without body parts being visible to a robot in the same environment. Imitation learning is therefore based on the behaviors of manipulated objects only. A simple Matlab interface for programming a simulated robot is also provided in SMILE, along with an XML interface for initializing objects in the virtual environment. SMILE lowers the barriers for studying robot imitation learning by (1) simplifying learning by making the human demonstrator be a virtual presence and (2) eliminating the immediate need to purchase special equipment for motion capturing.