Imitation learning and behavior generation in a robot team

Abstract

In this paper, we propose a method to apply robotic imitation learning in robot teams. In our method, behavior primitives with task-relevant information are defined as the basic units for robots to complete a task. Each behavior has its own task-relevant affordances. The learned behavior primitives format TAEM based behavior libraries and are stored in a database for robots to share. The motivation of learning, conducted by robots, is goal-oriented and strongly related to the given task. Given a task, a robot analyzes the environment and searches the behavior library to find suitable behaviors to generate a behavior sequence to complete the task. If it thinks that it cannot complete this task, this robot requests other robots for assistance or request a human teacher to demonstrate the required behaviors. The newly learned behaviors will be added into the existing behavior library. We also develop inhibiting properties for robots to evaluate the current behaviors, which enables robots to request collaborations from other robots. The experimental results show the validity our proposed method.