A ROS-integrated architecture to learn manipulation tasks from a single demonstration

In the robot programming by demonstration (PbD) framework, the high-level representation of a skill in a series of action units gives an intuitive method to program and instruct robot behaviors. In this context we present a ROS (Robot Operating System) integrated architecture for learning households manipulation tasks by one observation. The user is observed during the execution of everyday tasks, every action is analyzed and its effect is translated into changes in the environment state. During the observation a strip-like map of the task is built and stored as a sequence of actions. From the map obtained the task can be performed. A planner robustly adapts the execution both to different environment initial conditions and to possible faults, occurring during the operations. We test the capability of the chosen approach to autonomously learn and robustly perform complex tasks, such as setting up and clearing a table in a real kitchen-like environment.