Andy Wilson, Editor
In automation and handling engineering, feeding of work pieces between different stages of production can be improved by automation. In many applications, parts are stored unordered in bins and need to be separated and placed on conveyor belts specially adapted to the geometric properties of the parts. Consequently, any variations in the shape of work pieces require re-tooling of the production line, a process that is time consuming and expensive. Such tasks are exacerbated when large or heavy parts need to be picked randomly from bins and placed on production lines.
Vision-guided robotic systems can be used to automate this task. Although various 2-D image-processing algorithms can be used to recognize parts, these methods are not well suited for recognition and localization of randomly stored parts. At Fraunhofer IPA (Stuttgart, Germany; www.ipa.fraunhofer.de), algorithms have been developed to recognize and localize objects based on the analysis of 3-D data. Using such methods, many different kinds of objects can be picked and placed, only requiring an adjustment or replacement of the robot’s gripper.
A prototype vision-guided robotic system developed by Fraunhofer IPA uses object recognition software based on the best fit of geometric primitives in measured point clouds to pick and place parts from a bin.
Unlike other types of 3-D vision systems, the Fraunhofer IPA software tool is based on the best-fit of geometric primitives in measured point clouds. Geometric primitives are shapes like planes, spheres, cylinders, and cones. Using best-fit principles to fit such shapes to measurement points, the separate geometric primitives within the objects considered are recognized in order to localize the object as a whole.
“This approach is chosen because mechanical parts generally consist of objects that comprise planar, cylindrical, or conical features,” says Jens Kuehnle of Fraunhofer IPA. “However, our algorithms can not only cope with cylinders and cones but also with more complex parts which may even contain free-form surfaces.”
To demonstrate the capability of the PC-based software, Kuehnle and his colleagues have developed a specially designed robot cell to handle forged parts that are stored unordered in a bin. To image the box, a structured light sensor with a scanning area of 250 × 600 mm and a depth resolution of 0.2 mm is used to capture depth profiles. To obtain a depth map, the sensor is mounted on a linear axis and moved horizontally to scan the box.
The forged part is first localized and then a collision check is performed to ensure that the object chosen can be picked without colliding with the surrounding scene. After a pick point is computed, the coordinates are transferred to a robot that grips and places the chosen object in an ordered position outside the bin. According to Kuehnle, localizing the position of an individual work piece takes approximately 0.5 s with a positional accuracy of ±0.5 mm.
The objects that can be picked using the Fraunhofer IPA 3-D object recognition and localization algorithms typically contain a dominant geometric primitive feature. However, work in progress will in future allow more complex work pieces such as crankshafts with free-form surfaces to be picked and placed. Furthermore, Kuehnle and his colleagues are developing tools that can be used to automatically teach the software how to handle unknown parts. The software is currently available to companies interested in developing automated handling systems.