Recognizing waste as waste and distinguishing goods as sortable items: Swisslog and Kuka aim to simplify robot programming and the correct item selection of food and other products in large warehouses using artificial intelligence. Remarkable results are possible – but other things could probably be done more simply.
“Programming a robot for a task is still largely reserved for experts,” says Roland Ritter, Platform Program Manager Simulation at Kuka. “That’s precisely why we’re working on an AI chatbot that translates a simple voice command into programming code.” From the command “Grab the components one after the other and place them on the table in a U-shape,” the AI model then generates the code that makes the robot do exactly that.
Ritter is part of a team developing an AI chatbot for the Kuka Robot Language, and together with Swisslog's Head of Development, Niklas Goddemeier, an AI expert for robot-based order picking in the food, pharmaceutical or electronics sectors, he will answer questions from twenty participants in a virtual press conference.
From a Swiss perspective, the focus is primarily on data analysis and predictive maintenance, AI algorithms for warehouse optimization, reducing operating costs, and eliminating staffing shortages. When it comes to implementing instructions to the machine, says Ritter, it seems to be more common than not that "the first action is not what you expect." On the one hand, the effective use of AI promises that grippers and robotic arms shouldn't require "teach-in" to perform their task—for example, picking the correct items from a box containing various products. On the other hand, even here, implementing the verbal instruction definitely requires training to understand what is meant and what the robot is supposed to grasp.
A simple example: The device is supposed to remove cubes from a box, place them on a piece of paper without prior training, and then, as a test, move all the cubes to the left. "I could also write it in English in the menu window," says Ritter during the demonstration, which is displayed graphically on the screen and via a "live" camera. He could even at this stage—similar to "Alexa"—transfer the result, along with the actual sequence code, to the actual control system of the robot arm.
Photos: Swisslog/Kuka
Niklas Goddemeier, joining remotely from the Swisslog office in Dortmund, demonstrates "Item Pick" as a prime example of the learning process. On average, Swisslog customers such as Rewe, DM, or companies in the pharmaceutical industry have tens of thousands of different products in their assortments, packaged in bags, boxes, or even without outer packaging. "Every day, these diverse items have to be picked, meaning they have to be assembled for a customer or delivery order – and ideally, without errors," he says.
Not all possibilities can be pre-programmed. The key concepts here, in all their experimental spirit, are to "make the solution space manageable," to "increase the grasping success rate," and to clearly separate materials and the desired item from "waste." Errors should be corrected automatically within the context. Inside the box, there's visible chaos. Shampoo bottles, screwdrivers, a piece of cardboard—all to test the machine. The robot dutifully selects the appropriate attachment for its gripper and sorts the material into different compartments. The piece of cardboard is lifted by a suction gripper and placed in a wastebasket next to the work surface. Even building pallets no longer seems to pose a problem.
One participant asks whether what's being demonstrated is "already in practice," or "are you still practicing?" Would spontaneously formulated sentences—that is, sentences spoken "off the cuff"—also be possible? Ritter has to admit that everything is still a work in progress. At least initially, the AI actually has to learn how and in what form questions are asked or can be asked. "We're in the process of teaching the syntax and training language models." This will be interesting when you're not just dealing with a robot, but also want to immediately start up a conveyor system. Lead angles and positions have to be programmed in.
A few years ago, a manual "teach-in" was still performed with the robot arm held to the hand. The AI, so to speak, first needs to "ascertain" its place before it understands what it's doing. However, this allows it to be used more flexibly and develop more comprehensive scenarios on the object. Currently, all of this still takes place in a simulated environment. In the meantime, a digital twin steps in to check whether the AI-generated robot program is error-free.
www.swisslog.ch / www.kuka.com
