Retry-behavior Emergence for Robot-Motion Learning Without Teaching and Subtask Design

The behavior in robot-motion learning varies greatly depending on the demonstration data used for learning and modalities of a robot. One of the differences is the emergence of untaught retry behavior. The ability to retry is important for executing difficult and uncertain tasks and improving robustness against external disturbances. We consider it important to recognize the progress status of a task, especially the completion status. We examined cases in which task execution is obstructed in an object-picking task, which is the basic task of many tasks, and demonstrated the effectiveness of two approaches: 1) incorporating a modality capable of accurately recognizing the end state and 2) introducing an action into the demonstration that facilitates the grasping state using existing modalities. The results of experiments using a real robot showed that the tactile modality as approach 1 was effective in recognizing whether the object was grasped and that adding a lifting motion after grasping to visually judge as approach 2 was effective. Upon analyzing the internal state of the recurrent neural network used in the learning model, it was revealed that the two approaches significantly contributed to appropriate state transitions.