If machine learning methods are to be used in industry, robots will have to learn faster than before how to grasp components safely. In the Deep Grasping research project, scientists from Fraunhofer IPA and the University of Stuttgart are therefore developing a virtual learning environment. The neural networks are trained in this environment and then transferred to the real robot.
Thanks to artificial neural networks, robots are able to draw conclusions from practical gripping experiments. In other words, they learn from experience and, as time goes by, they improve their bin-picking skills. However, this form of machine learning has a crucial weakness: learning times are too long. For this reason, it is still uneconomical for industry to perform handling tasks with the aid of artificial intelligence. In order to drastically shorten the time needed by robots to learn to grip components safely, scientists from Fraunhofer IPA and the University of Stuttgart are currently developing a virtual learning environment in the Deep Grasping research project. In the future, robots will train their neural networks and exchange their experiences before they are put into operation. Bin-picking will no longer be trained in practice but only simulated. That saves time. The pre-trained neural networks are then transferred to the real robot.