Description

Robot learning of cloth handling skills is an increasingly relevant task, that arises in the field of assistive robotics. However, it presents numerous challenges, related both to the non-rigid nature of the object to be manipulated, and to the type of real-world datasets used to learn these skills, which usually comprise a small number of demonstrations provided by a human agent. Thus, the use of Bayesian, data-efficient machine learning models, such as Gaussian Processes (GPs), is particularly suitable for this task. Furthermore, GPs provide us with richer information than other approaches, albeit with a high computational cost. If capable of dealing with this issue, however, GPs can transmit valuable information to a stochastic controller tracking the reference signal, as well as its precision requirements. Within this context, this thesis project is aimed at learning to manipulate cloth by combining stochastic models and predictive control, in order to achieve a robust and generalizable behavior that can be adopted in real-world scenarios.

The work is under the scope of the following projects:

• CLOTHILDE: Cloth manipulation learning from demonstration (web)