Garments are a key element of people’s daily lives, as many domestic tasks -such as laundry-, revolve around them. Performing such tasks, generally dull and repetitive, implies devoting many hours of unpaid labor to them, that could be freed through automation. But automation of such tasks has been traditionally hard due to the deformable nature of garments, that creates additional challenges to the already existing when performing object perception and manipulation. This thesis presents a Robotic System for Garment Perception and Manipulation that intends to address these challenges.
The laundry pipeline as defined in this work is composed by four independent -but sequential- tasks: hanging, unfolding, ironing and folding. The aim of this work is the automation of this pipeline through a robotic system able to work on domestic environments as a robot household companion.
Laundry starts by washing the garments, that then need to be dried, frequently by hanging them. As hanging is a complex task requiring bimanipulation skills and dexterity, a simplified approach is followed in this work as a starting point, by using a deep convolutional neural network and a custom synthetic dataset to study if a robot can predict whether a garment will hang or not when dropped over a hanger, as a first step towards a more complex controller.
After the garment is dry, it has to be unfolded to ease recognition of its garment category for the next steps. The presented model-less unfolding method uses only color and depth information from the garment to determine the grasp and release points of an unfolding action, that is repeated iteratively until the garment is fully spread.
Before storage, wrinkles have to be removed from the garment. For that purpose, a novel ironing method is proposed,nthat uses a custom wrinkle descriptor to locate the most prominent wrinkles and generate a suitable ironing plan. The method does not require a precise control of the light conditions of
the scene, and is able to iron using unmodified ironing tools through a force-feedback-based controller.
Finally, the last step is to fold the garment to store it. One key aspect when folding is to perform the folding operationin a precise manner, as errors will accumulate when several folds are required. A neural folding controller is proposed that uses visual feedback of the current garment shape, extracted through a deep neural network trained with synthetic data, to accurately perform a fold.
All the methods presented to solve each of the laundry pipeline tasks have been validated experimentally on different robotic platforms, including a full-body humanoid robot.