Automated palletizing and depalletizing with artificial intelligence

April 14, 2021

Especially in production and logistics, efficiency along the entire value chain is an essential factor in competition. Reliable automation of manual and time-intensive processes is crucial for a modern and efficient factory and warehouse.

Traditional palletizing solutions - inflexible, complicated, expensive

Traditional palletizing robots either function purely statically or have to be programmed from scratch in a very complex way. In this case, algorithms are developed by experts by hand, which often requires a lot of know-how and time. Often complex palletizing tasks, such as chaotic sorting, difficult geometries or mixed pallets, cannot be realized at all or only with great difficulty using classic approaches. All this leads to high costs and to the fact that automation requirements cannot be met.

Solving complex palletizing tasks efficiently with Vision AI

With AI-based image processing, complex palletizing tasks can be automated reliably and quickly during ongoing operation. To enable orientation in three-dimensional space during the gripping process by the robot, 3D imaging methods such as time-of-flight or stereo vision camera systems are used. With Data Spree's Deep Learning DS software, the software logic can be implemented efficiently and easily in the background.

Figure 1 - Left: False color image of depth information and gray scale image, right: 3D point cloud and object surface of the detected packages

The first process step is to capture images of the objects to be palletized. Then the objects are assigned to classes, for example package type A, package type B and package type C. This assignment is called data annotation or labeling.

The AI then iteratively trains the recognition and correct assignment, as well as the position, size and orientation of the objects. Here, the AI works similarly to the human brain and learns to recognize, assign and localize objects based on the image data - without the need to manually pre-define specific object features. With Deep Learning DS, you can quickly and easily perform this "learning process" yourself. In addition, Data Spree also offers the complete process up to productive integration into the facility as a service.

This method thus allows the most diverse and complex palletizing tasks to be implemented quickly and easily - and without a single line of programming code. Automation processes can thus be implemented very efficiently and robustly. A ready-to-use prototype can thus be created in just a few hours and expanded into a productive solution within a very short time. Data Spree's fast AI models additionally ensure real-time capability in high-frequency production and logistics operations. Another advantage is the flexibility of the learning system. If products, product properties or objects change due to production or logistics changes, the AI can simply be "fed" with new images and retrained. In this way, it is possible to react quickly and effectively to changes in production or logistics without having to start from scratch or buy and implement a new solution.

Quick and easy AI implementation and execution

The trained AI model can be individually integrated into any customer application through the open ONNX standard format. Data Spree's own execution environment Inference DS also offers a simple graphical user interface in which the AI model can be quickly executed via drag-and-drop principle on there spective hardware, such as smart camera or industrial PC. This saves integration time - and above all costs.  

This example shows a chaotic palette structure with different objects of various size, dimension, geometry and orientation. On the left side you can seethe output false color image, which is composed of depth information and grayscale image. On the right, the 3D point cloud with markers on the detected objects and the InfluxDB dashboard fully integrated via the ADLINK Data River to track the set of objects. This sample application achieves execution times of less than 30ms, making it excellent for fast palletizing processes.  Using the Inference DS robot plugin, the object coordinates can be easily output and sent to the appropriate controller. Thus, the customized palletizing application is not only quickly completed, but also quickly and easily integrated.

Let's talk about your project.
Adopt one of our ready-to-use solutions, start creating your deep-learning-based system, or tell us your requirements to let us build an individual application for you.