Fiziksel İnsan-Robot Etkileşiminde Niyet Kestirimi: Derin Öğrenme ile Alt Görev Tanıma

Understanding the human's intention in a physical collaborative task is important for a robot to provide the right type of assistance. We claim that most tasks are made of sequential subtasks and define the human intention as the current desired subtask to be executed. These subtasks have different requirements and a single controller will not be the best way to regulate the interaction. If these subtasks can be recognized in real time, an appropriate controller or controller parameters can be selected. In this work, we propose the idea of subtask recognition, formulate it as a time series classification problem, and develop a deep learning approach to accomplish it. We perform three experiments. The first one (n=10 subjects) verifies the viability of our subtask recognition approach, achieving above 90% test accuracies. The second one (n=5) tests our approach in multiple challenging conditions and shows that it is robust to different human operators, human arm configurations, controller parameters, and environment stiffnesses. The third experiment (n=18) tests our approach in real time. The results provide empirical evidence that the task performance is improved using the proposed approach.