Increasing the Robustness of Random Bin Picking by Avoiding Grasps of Entangled Workpieces

In bin picking applications, a robot often picks workpieces that have a complex geometry. This complex geometry can cause entanglements between workpieces resulting in failed grips. This paper presents a machine learning approach to avoid these situations and therefore improves the calculation of suitable grips. Using the depth map of the workpieces and their surrounding neighborhood, a convolutional neural network, which is trained on simulated data, predicts whether an entanglement is present. This information is used to select and calculate the most reliable grip. By avoiding such entangled workpiece situations the robustness of random bin picking increases.