Remmert and Fraunhofer IOSB-INA automate long-goods handling with AI robotics
Fully automated singulation of unsorted long goods without offline programming – presented at the Remmert Experience Day
In the application shown, PickFLEX retrieves packaged materials from a cassette and places them into another cassette, which is fed to the picking station via a conveyor system.
The first Remmert Experience Day in September also marked the grand opening of the new Experience Center, which was built specifically for customer presentations.
Dr. Ing. Sebastian Wrede teaches in the field of cognitive robotics at the CoR-Lab at Bielefeld University and is a senior researcher at the Fraunhofer IOSB-INA Institute in Lemgo.
Remmert Product Manager Fabian Gries explains the PickFLEX robotics concept to the audience
Picking long goods is one of the last largely manual steps in otherwise highly automated production environments. Aluminium profiles can weigh up to 30 kilograms depending on their length. Removing individual bars from a bundle is physically demanding and usually takes two people – miscounts during shipping are a frequent consequence. Together with Fraunhofer IOSB-INA, Remmert has developed PickFLEX, a solution that automates exactly this process. It was presented at the Remmert Experience Day on 3 September 2026. “In many companies there is still a gap between storage and production that nobody has closed – precisely where the work is hardest,” says Fabian Gries, Product Manager at Remmert. “PickFLEX closes that gap. And because the system learns rather than being programmed, it grows with our customers’ requirements.”
Picking without pre-sorting
Unlike conventional robotics applications, PickFLEX does not require pre-sorted material at fixed coordinates. Cameras and AI-based image processing capture the actual position of the parts in the cassette, and the industrial robot picks accordingly. New materials are taught to the system rather than programmed into it.
Dr.-Ing. Sebastian Wrede, Senior Scientist Robotics at Fraunhofer IOSB-INA, explained how demanding this is in technical terms: long goods are highly variable and are stored in only semi-structured form. Reflective, glossy or dark surfaces make 3D detection difficult, yet millimetre accuracy is required across objects that are metres long. Added to this are tilted and twisted parts, which make safe gripping points hard to determine, as well as large collision volumes, deflection and vibration as material is moved through the cell.
Four building blocks of the architecture
The jointly developed control architecture comprises AI perception with retrainable vision models, full scene reconstruction including cassettes and dividers, autonomous decision-making on pick and place positions, and online path planning. All movements are calculated at runtime, so manual teaching is no longer necessary. Development took place in simulation environments such as KUKA.Sim and NVIDIA Isaac Sim. Gripper tests were run physically at SmartFactoryOWL. Integration via ROS 2, KUKA.RSI and OPC UA keeps the system vendor-neutral.
Scalable and ready for integration into the storage portfolio
Vacuum or magnetic grippers allow PickFLEX to be adapted to different materials, with payloads ranging from approximately 50 to 300 kilograms. The system can be combined with Remmert storage systems. Typical applications include order picking prior to shipping and material supply for saws or window manufacturing machines. Remmert expects a fast return on investment in a multi-shift operation, as the system takes over the work of around one and a half people per shift.
Bringing the latest trends into practice
For Sebastian Wrede, the project is a good example of how different technical trends come together in a practical application: first, the shift from rigid rules to AI models that learn perception, grasping and motion from data. Second, training data from a digital twin validates the system before the real machine comes into play.
Generalist vision-language-action models that transfer capabilities between tasks will, according to Wrede, remain a research topic for industrial precision requirements for some time. Continuous learning from experience and human feedback is also promising: in Fraunhofer research, around 20 human corrections raised the success rate of an assembly task from 20 to almost 100 per cent. This approach could in future help handle jammed long goods, for example. His conclusion: AI robotics does not replace conventional automation but complements models, control technology and engineering expertise with perception, adaptation and the ability to learn.
Beyond the PickFLEX robot, Remmert already uses digital twin technology to emulate the control logic of complex storage systems. This significantly reduces commissioning times for high-bay warehouses, among other benefits.