Abstract
In robotics, the area of deformable object manipulation receives far less attention than that of rigid object manipulation. However, many objects in real life are deformable. Research on deformable object manipulation is indispensable to equip robots with full manipulation dexterity. Deformable linear object (DLO) is one type of deformable objects that commonly presents in the industry and households, for instance, electrical cables for power transfer, USB cables for data transfer, or ropes for dragging and lifting equipment. In the context of H2020 VERSATILE, a project focusing on industrial automation using robots, we focus our research on DLO manipulation via visual feedback.One characteristic of deformable object manipulation is that the object shape changes while being manipulated. Consequently, a research direction is to control the shape of the object during manipulation. We tackle the shape control problem by using vision. Initially, we parameterize the shape with Fourier series, estimate and update the interaction matrix online, and finally control the DLO shape.In the subsequent research, instead of using human-defined features for parameterization, we let the robot automatically learn feature vectors from visual data. We propose a method that allows the robot to simultaneously generate a feature vector and the interaction matrix from the same data. Our approach requires minimum data for initialization. Learning and control can be done online in an adaptive manner. We can also apply the method to rigid object manipulation directly without modification.Neither of the two frameworks requires camera calibration, and both are verified with simulation and real robotic experiments.Another area of importance in deformable object manipulation is the utilization of external contacts. The object deformation is defined in a configuration space of infinite dimension. Nonetheless, the inputs from robots are limited. External contacts can and should be used for manipulating deformable objects. We take a practical scenario in the industry -- cable routing with external contacts as the process to automate with our robot. We propose a planning algorithm that allows the robot to use contacts for shaping the cable and achieving the desired cable configuration. Real robotic experiments with different contact placement scenarios further validate the algorithms.