Dense-Descriptor Transporter Networks for Multidimensional Deformable Object Manipulation

With the continuous development of robotics, there are emerging new trends in robotic manipulation of deformable objects, calling for improved adaptability and flexibility to better cope with complex environments and task requirements. Due to the high-dimensional state space and complex dynamics of deformable objects during deformation, it is challenging for neural networks to directly extract object features such as shape, texture and color from images. In this paper, we propose an imitation learning approach that utilizes dense descriptor modules to improve its perception of different deformable objects and deformations. The learned descriptors can extract the features of deformable objects from RGB-D images and map them to the descriptor space, making the representation of objects more compact and robust. Furthermore, our proposed approach demonstrates better competitiveness in simulation environments featuring fabrics, ropes, or bags when compared to the baseline methods.