Artificial intelligence (AI) may not be a panacea for the current requirements in the packaging industry, but when used wisely, it can greatly simplify and accelerate processes. As a driver of innovation, packaging machine manufacturer Schubert has been exploring the potential of AI for years and has already developed exciting new technologies for robot-assisted packaging solutions. Its customers, mainly manufacturers of consumer goods, benefit from greater efficiency, lower costs, easier handling, and more flexibility.
Along the supply chain – from the raw material to the consumer product ready for sale – many options are currently being discussed for using AI to improve processes. One approach is the robot-based packaging system, as robots are usually controlled by programmed algorithms that can be supported or even completely replaced by AI in the future, depending on specific requirements. This is where Crailsheim-based packaging machine manufacturer Schubert set out years ago with its research and development of AI applications. Ralf Schubert, Managing Partner of Gerhard Schubert GmbH, is firmly anchored in machine technologies and is responsible for the technical direction of the family business. For him, the topic of artificial intelligence is clearly defined: “I see opportunities for the use of AI wherever the programming of algorithms is too complex, or the algorithms are too slow.” The Schubert tog.519 cobot is one example of how AI based on neural networks ensures simple handling and high flexibility for customers.
AI used in the Schubert cobot for quickly picking up unsorted items
The cobot is designed for high-performance pick & place applications with lightweight products all the way through to bin picking, i.e. picking unsorted items from a crate. At up to 90 cycles per minute, the tog.519 picks a wide variety of sorted or unsorted products or packaging materials and places them in any conceivable destination – without requiring any programming. Bin picking is by far the most complex pick & place task, as unsorted products always behave differently. Each time the cobot reaches into the bin, it has to work out which product is in the top position and can be picked up best.
Ralf Schubert explains how this works: “We use an AI-supported image processing system that we developed ourselves as the basis for controlling the tog 519. It recognises both the products as they are picked up and the environment in which they are to be placed. An expensive 3D camera is not needed for this; a standard 2D camera is perfectly adequate. Even for bin picking, because here the AI simply generates the 3D images required for the cobot from the camera images supplied.” This allows the cobot to permanently ‘see’ what it has to do through the generated images. This virtually eliminates the need for a teach-in or start-up process in the event of a format change or a new task at a different location. Once set up, the Schubert cobot gets to work immediately.
The cobot’s neural network is so extensively trained that the robot can immediately process new products from the same product group. These can include pouches (sachets, flowpacks, stand-up pouches, sealed pouches, etc.) in various sizes or bottles in different shapes and materials. “We train one network per product class,” explains Ralf Schubert. The Schubert team configured the tolerance of the AI in such a way that variable sizes, materials, surfaces, or thicknesses are accepted within a product group. “If similar processes were programmed using conventional algorithms, even a different colour would be problematic with otherwise identical packaging,” says the Managing Director, adding: “Bin picking is therefore one of the applications in the packaging process that cannot be implemented in practice without AI.”
Training with a generated image data set
Training neural networks, i.e. machine learning, still takes time. In general, a distinction is made between supervised learning and reinforcement learning. Both types of training are essentially based on a large data set of images, but they differ in the way they are carried out. “With supervised learning, the result of the desired action is known and precisely definable,” explains Ralf Schubert, “but not with reinforcement learning”. Therefore, each image from the data set for supervised learning has to be provided in advance with labels that contain a clear yes or no, whereas a significantly smaller number of images is sufficient for reinforcement learning, which also do not have to be labelled. In this case, the network more or less trains itself with a repeated cycle using a reward system, but it needs to be able to perceive its environment. This means that cameras and various sensors have to be integrated into the robot, which can provide feedback on the current positions of the robot, packaging, and product.
Good training always calls for a large number of images that are generated artificially in advance. Ralf Schubert describes how this works for a cobot network: “For a new product group, for example, we work with 50,000 images that we generate over the course of two days. The network is then trained in-house for a further two days. If extensions to this product group need to be trained later, for example an entirely different packaging surface that is not recognised, our customers can do this themselves on site with a few images.”
More flexibility in production thanks to quick format changes
The trained network in the Schubert cobot demonstrates how quickly new formats can be introduced with different products without having to be reprogrammed each time. Image processing is key to this. It ensures that other products, different formats, or new packaging tasks can be processed immediately without additional effort. Thanks to AI, production changeovers can therefore be carried out extremely efficiently and without long downtimes – be it in terms of the size and shape of the products or the format and material of the packaging. Customers who have to constantly adapt their pick & place processes to market-specific requirements can therefore achieve a very high level of flexibility in their production with several cobots. They can either be used individually at different process steps and locations or set up as a line.
Increasing efficiency and saving costs with optimised robot paths
At Schubert, AI not only helps the cobot ‘see’, but it also helps the robots in the packaging machines follow better paths. More precisely, to more organic motion sequences, called Schubert Motion. Behind this is a small, highly skilled Schubert team in Dresden, who has set the development of the motion software as its first milestone and already achieved it in the F2 robot with series production. With the AI-controlled software, the movements of the packaging robots can be optimised for speed, significant vibration reduction and even energy efficiency. Ralf Schubert works closely with the Dresden team: “Schubert Motion enables us to generate the robot paths with the help of AI and improve them accordingly. Compared to programmed robot paths by humans, these motion sequences are up to 20 per cent faster, more economical in terms of energy consumption and gentler on the mechanics, which are subject to less strain.” For customers, this has advantages on several levels. The packaging speed increases, the energy consumption in the machine decreases and the components involved in the packaging system are protected. All of this has a positive impact on sustainability in the packaging process and also reduces operating costs. At the same time, low-vibration operation reduces the background noise in production, which makes working more pleasant.
The future of AI lies in simplifying technology
At Schubert, AI has so far been used primarily in robot control, where it noticeably ensures easier handling, faster and more flexible processes, greater energy efficiency and, ultimately, lower costs. But it is unlikely to stop there. Ralf Schubert looks ahead and concludes: “Simplicity is and will remain the key to the packaging technology of the future. What could be extremely interesting in this context is AI for programming, problem solving and knowledge management – similar to Chat GPT, but as Schubert GPT. One day, it might even be possible to talk directly to the packaging machine, as you would with a bot.”
