Learning Peg-In-Hole Actions with Flexible Objects

This paper presents a method for learning Peg-In-Hole actions with flexible objects. To learn the actions we parametrize the entire trajectory by a single point and use Kernel Density Estimation to reflect the different variations of the action and the object characteristics. The object is characterized by its elastic behaviour rather than geometric properties. Thereby an action learned for one object can be transferred to a new object that behaves similarly although it might have different elastic properties, dimensions and geometries. To bootstrap the learning mechanism, the system performs simulated actions and utilizes the detailed information obtained from the simulation environment. Subsequently Peg-In-Hole actions are tested successfully on the real life