Target-Referred DMPs for Learning Bimanual Tasks from Shared-Autonomy Telemanipulation

Abstract

The Learning from Demonstration (LfD) paradigm allows transferring human skills to robots without the need for explicit programming. To be effective, it requires: (i) a learning technique able to encode and adapt demonstrated skills to different contexts and (ii) an intuitive user interface for task demonstrations. Both aspects become more crucial when dealing with multi-robot coordination. Dynamic Movement Primitives (DMPs) are among the most reliable LfD techniques. However, they might struggle to correctly replicate learned manipulation tasks for a target object with a different orientation from the demonstration. On the user side, telemanipulation solutions can provide an effective interface for demonstration acquisition. Recent shared-autonomy control strategies allow intuitive coordination of multi-robot platforms, but none has been exploited in LfD. In this work, we propose a novel implementation of DMPs, called Target-Referred DMP (TR-DMP), which improves generalization capacities and overcomes the above limitation. Furthermore, we show how to embed a shared-autonomy tele-manipulation strategy in our LfD architecture for an intuitive training and execution of bimanual coordinated tasks. The improved performance is proven through two real case studies.