The core functionality of this service is the recognition and localization of one or more pre-trained object classes in point clouds captured with a 3D camera. Here, it is assumed that the objects are resting on one or more planar layers (e.g., a pallet, cardboard separators, or a conveyor belt). The achievable accuracy is sufficient for direct loading of a machine, without additional mechanical centering/alignment aids. 

Some processes do not lend themselves to imaging with a 3D sensor, either because the parts in question exihibit unfavorable reflectance, or the relatively long image acquisition time is prohitive (under a given cycle time). For these cases, we offer a service for localizing items on a planar surface in 2D images. The service learns from each image taken, minimizing the influence of ambient light, which is all the more important without the active illumination typically found in 3D sensors.

Like any other sensor, in order to maintain sufficient accuracy, cameras must undergo regular calibration to compensate for changing environmental conditions (vibrations, temperature fluctuations, etc.). This is often done with calibration software provided by the camera vendor which is unable to exchange data directly with the robot cell thus making the process error-prone and time-consuming. Our API enables complete integration of all common calibration procedures (including extrinsic, intrinsic and hand-eye calibration) into the automated process, for almost any type of robot and camera.

The precise picking of an object directly from a box of bulk goods (“bin picking”), e.g., for assembly or machine loading, is advantageous for several reasons. There is no need to keep object-specific carriers on hand, no manual loading of the carriers is necessary, and the transport density remains unchanged.