Underwater Autonomous Grasping Strategy with Constrained Visual Information

When the robot arm grasps across media, continuously grasping and placing the objects inevitably causes water flow fluctuations, thus interfering with subsequent visual quality. This results in the inability to correctly feedback object information, ultimately leading to grasping failure. To tackle this problem, we propose a generalized grasping model based on constrained visual information (VC-Grasp). Specifically, we leverage RGB-D information combined with the physical logic of action primitives to achieve effective movements. During the operation, we thoroughly consider the relative position changes between objects, thereby attaining true zero-collision. Eventually, we predict the optimal action sequences for the upcoming period to complete the task. Additionally, we propose a novel sim2real framework, which significantly enhances the generalization capability of our model. Extensive experiments in physical environments demonstrate that our system outperforms other methods in terms of grasping success rate and time efficiency under constrained visual information conditions.