Description

The study of rigid object robotic manipulation has remarkably advanced recently due to its successful application in industrial assembly lines. While the manipulation of rigid objects has been widely assessed, deformable objects are still manually processed given that handling deformable objects usually requires visual and tactile information, more complex models, prior knowledge and the interaction of several dexterous manipulators. Achieving proper manipulation on flexible objects is a key factor for robotic and industrial development on many fields such as textile and food industries, as well as in household premises.

Unfortunately, the manipulation of objects in real situations raises important issues that prevent the proper execution of the actions theoretically estimated. Indeed, inaccurate sensory perceptions that are affected by noise, carried out in dynamic, inaccessible, indeterministic environment constitute a hostile scenario. In this context, planning requires continuous actions capable of coping with the perception deficiencies and the environment mutability in order to accomplish the tasks assigned.

This master thesis will address planning algorithms for manipulation of deformable objects, acting as a bridge between modelling and planning. The high-level task specifications have to be integrated with the low-level geometry extracted from simplified physical models. We will also take into account the uncertainty of the sensors, movements, and therefore, the general scenario state in order to plan a specific sequence of motion commands. We will explore several techniques of artificial intelligence, learning and planning to suit the particularities present in deformable objects.

The work is under the scope of the following projects:

• PAU: Percepción y acción ante incertidumbre (web)
• APREN: Modelos perceptivos y técnicas de aprendizaje para robots de servicios (web)
• IntellAct: Intelligent observation and execution of Actions and manipulations (web)