Robotic dexterity was one of MIT Technology Review's 10 Emerging Technologies in 2019, selected by Bill Gates. This is an essential area, given that the machines are still unskilled and are usually only capable of doing the same task over and over again. If we move them to a different location on the assembly line or ask them to grab a different object than usual, they get confused.
Juan Aparicio, the Spaniard responsible for Advanced Manufacturing Automation at Siemens Corporate Technology, is working to make industrial robots skilled enough to do human tasks in environments such as factories, warehouses, and farms. Their purpose is to grab different objects, even if they have never seen them and manipulate them to carry out different tasks. Thus, he has created a system, installed on a platform, called the Neuronal Processing Unit (NPU) which works with deep learning to calculate in milliseconds what movements it must make in order to grab an object. Thanks to this technology, Aparicio has become one of the winners of Innovators Under 35 Europe from MIT Technology Review.
His goal is not to take jobs away, but to empower employees, increase collaboration between them and robots, and make factories more autonomous. The researcher details, "There are robots that can win at chess. And when I say robots I mean artificial intelligence. Nevertheless, there is no robot capable of getting up, going to a shelf to pick up the board and start moving the pieces." he goes on to add: "There is a tremendous need, not only in industry, but also in agriculture, services, warehouses, etc.," he lists. Where he currently resides in California (USA), "there are farms that have to close because there are no people willing or able to do the job."
To understand how to grasp the object in front of it in each case, the system uses DexNet, an algorithm co-created with the University of Berkeley in California linked to a database of 3D objects. DexNet combines artificial intelligence with geometric models to calculate all possible grips.
Once the object is grasped, the next task is handling it using reinforcement learning. "The secret lies in guiding the robot with a series of stimuli or rewards (like you would give a dog when it does things right, for example) until it gets the job done." The system uses models and simulations to accelerate learning.
The machine is also capable of alerting humans when they cannot finish a task. "Any autonomous or intelligent system, just like a human, is going to burn out. It's normal for the robot to try to grab something and release it or it won't see how to come up with a solution," explains Aparicio.
Another advantage of his system is that robots can connect to the cloud. Therefore, among Aparicio's future plans include democratizing the algorithm ("that anyone can send a request and in turn receive the points needed to make the robot perform the grip") and industrializing it, for which they are already talking to several companies. The team had the opportunity to demonstrate its progress at Hannover Messe (Germany), the most important industrial fair in the world.
Nicholas Zylberglajt, Co-founder and President of the European Young Innovators Forum and a member of the Innovators Under 35 Europe jury, believes that taking unknown objects "repeatedly, reliably, and intelligently" is one of the bottlenecks in warehouse automation. For this reason, he believes that Aparicio's proposal "is closer" to providing a solution for multinationals such as Amazon, ABB, and Alibaba.
By José Manuel Blanco
Translation: Brian Bostwick