Light-Weight Convolutional Neural Networks for Generative Robotic Grasping

Grasp planning with high-performance and efficiency in unstructured environments is a critical problem that intelligent robots need to solve urgently to complete picking tasks. To solve this problem, a quantized grasp quality generative neural network is proposed to generate pixel-level grasps. A light-weight convolutional neural network is first used to generate initial grasp configurations. Aiming at the degradation of grasping performance caused by the light-weight design, a decoupled grasp quality network is designed to generate a pixel-wise grasp quality. To improve the robustness and generalization of the model, an adaptive filtering method is proposed to filter the grasp configurations. Then, an ellipse fitting-based grasp pose optimization method is proposed to obtain the final grasp configuration. On the public Cornell dataset, the state-of-the-art grasp accuracy of 98.9% and the efficiency of 15 ms are achieved, with approximately 564× fewer parameters compared to comparable performance. In the grasping experiments on the robotic platform, grasp success rates of 99% and 93.4% can be achieved in single object and cluttered scenes, respectively. Finally, the model is deployed on an embedded artificial intelligence (AI) computing device to verify its practicality.