Prediction of the configuration of objects in a bin based on synthetic sensor data

摘要

Bin picking describes a robot system that picks unsorted objects from a bin based on sensor data. The object recognition is one of the main difficulties when calculating the grasping pose. In the worst case, a faulty object recognition leads to a system standstill. This work introduces a framework concept for improving the object recognition by predicting the configuration of all the objects in the bin based on an initial scan in form of a point cloud. This is done by taking advantage of an online simulation, with the goal of reducing the total number of scan cycles, which leads to a reduced risk of failure and improves the overall performance of a bin picking system.