Instance-Level Coarse-to-Fine High-Precision Grasping in Cluttered Environments

Grasping is usually the initial stage of robotic manipulation tasks. High-precision grasping can reduce the uncertainty of the target and is beneficial for completing downstream manipulation tasks. This article proposes a coarse-to-fine grasping pose estimation scheme for cluttered environments, which can achieve submillimeter grasping accuracy using only consumer-level cameras. In the fine estimation stage, a cascaded end-to-end grasping pose prediction model is designed. We propose a new regularization method based on the semantic segmentation priors to avoid the overfitting problem. Also, an object-level data augmentation method is adopted to adapt the model to cluttered environments. With this method, the model trained with the data collected under a pure background can be generalized to cluttered environments. A large variety of typical experiments are conducted to validate our algorithm, including insertion tasks, screwing tasks, unlocking tasks, and door-opening tasks.