Guiding Robotic Cloth Grasping in Darkness: Infrared Semantic Segmentation and Grasping Position Selection

Abstract

Robotic cloth grasping is a key component in many robotic cloth manipulation scenarios, such as automated wardrobe management, clothing laundering, and assisted dressing. Due to the deformability and large surface of cloth, which distinguishes it from conventional rigid targets, most current studies adopt the paradigm of first perceiving the cloth area and then estimating the grasping position in the area. However, current methods overlook the perception sensitivity to illumination variations, leading to performance degradation in low-light conditions or completely dark scenes. In particular, current visual perception methods that fuse RGB and infrared images or are based on low-light enhancement still rely on valid RGB input. Consequently, they cannot be applied to scenes where the lighting becomes extremely low or even completely dark. In this paper, to address the above challenges, we propose a robotic cloth grasping framework based on infrared perception, which includes an infrared semantic segmentation model and a grasping position selection strategy. The infrared semantic segmentation model employs RGB-guided learning to extract features corresponding to RGB information from infrared images. Only infrared images are used as input during inference, enabling robust cloth perception in complete darkness. The grasping position selection strategy is through depth-derived wrinkle analysis via geometric processing. It eliminates dependency on annotated data or task-specific training, remaining invariant to scene lighting conditions. Building on the proposed methods, we developed a robotic cloth grasping system for testing. Extensive evaluation and baseline comparison experiments as well as ablation studies were performed to confirm the effectiveness and superiority of our method.