Polymorphic robot learning for dynamic and contact-rich handling of soft-rigid objects

Abstract

In the operation of robots in regular human lives, the capability of object handling is of fundamental importance. Robotic manipulation has gone from handling single rigid body objects with firm grasping to handling soft objects and dealing with slip and contact. Meanwhile, technologies such as robot learning from demonstration has enabled intuitive human-to-robot teaching. This paper discusses a new level of robotic learning-based manipulation. Instead of the single form of learning from demonstration, we propose a polymorphic learning scheme that integrates additional types of robot skill acquiring, including adaptive definition and evaluation. In addition, compared to the current studies of handling pure rigid or soft objects in a pseudo-static manner, our work aims to allow robots to learn to manipulate objects that are partly soft partly rigid, require time-critical dynamic skills and subtle contact control, such as handling tethered tools and even using martial arts instruments. This type of tasks, once successfully robotized, open a variety of new possibilities in robot-human coexistence.