Learning Joint Stiffness Variations from Demonstration

Humans demonstrate an impressive ability to manipulate fragile objects without damaging them, graciously controlling the position of hands and tools while managing contact forces as necessary for the task. Robot control has been characterized by position control with the aim of producing fast, accurate and repeatable motion. These capabilities have reached a level where robots often outperform human workforce in repetitive assembly tasks. However, for extending the applicability of robotic manipulation outside this class of industrial tasks, pure position control is inadequate. In tasks which involve contact with objects whose position is not known with perfect certainty, the use of a pure position controller can cause unbounded rise of the contact force, ultimately leading to dangerous behaviors such as breakage or instability in the control system. To deal with the problems that arise in contact, a compliant controller can be used instead. However, determining the compliance parameters to suit the task is often difficult, and may require accurate models of the environment and robot. An attractive alternative is to use learning, replacing the need of model-based compliance planning with instructions from a teacher and/or experience. The problem of adapting the compliance to suit the need of the task was treated in a Reinforcement Learning (RL) setting in [1]. In that work, the robot adapted the gains of a feedback controller to maximize task performance while minimizing energy. Another approach based on reinforcement learning was presented in [2] in which reinforcement learning was used for learning task space stiffness and motor coordination for flipping pancakes. In Robot Learning from Demonstration(RLfD) the task model is derived directly from demonstrations without the use of a reward function describing the task as in RL. Calinon et al. [3] proposed an approach to extract stiffness variations from the variability of demonstrated kinematic data. This approach is based on the assumption that stiffness should be directly related to kinematic variability. In our previous work [4] and in this abstract, we relax this assumption by teaching motion and stiffness profiles independently. In this abstract, we contribute by presenting a novel interface for teaching varying joint stiffness, and show how this interface can be integrated in a learning system for teaching stiffness variations. The approach is illustrated by a task of lighting a match using the KUKA Lightweight Robot (KUKA LWR), see Figure 1. Instead, we will focus on the problem of teaching a robot how to vary the stiffness in a task for which it has already learned the positional profile. To this end, we present an interface which adapts the stiffness of the joints based on the Fig. 1: The figure shows the KUKA Lightweight robot lighting a match using a motion and stiffness profile learned from demonstration.