Robotic Grasp Detection Method Based on A Novel Multi-Scale Multi-Task Mutual-Learning Architecture in Neural Network

Robotic grasping is challenging due to considerable variability in objects' shapes, positions, and orientations. Therefore, we put forward a novel multi-scale multitask mutual-learning (MSTL) architecture as a module in the grasp detection network. The proposed architecture enhances the network's learning capability by facilitating the exchange of complementary features among subtasks under different scales. Furthermore, the attention mechanism applied to it makes the exchange more purposeful and selective. In our experiments, the proposed architecture demonstrated superior performance, improving the prediction accuracy from 92.9% to 95.8% on the Cornell dataset, exceeding other state-of-the-art methods, and producing the highest rate of success in real-world grasping.