AI Model Types

There are two types of application modules

  • K-Pick
  • U-Pick

Solutions are all based on a first stage consisting of image segmentation. In U-Pick such segmentation is performed assuming all objects belong to the same generic category (object). In K-Pick instances are classified into the categories defined by the user. In K-Pick texture the identification of all instances in the pick scene is made based on texture (i.e., RGB) attributes whereas in K-Pick geometry the identification is based on geometric attributes derived from the point clouds captured by the camera.

Unlike U-Pick, K-Pick builds upon knowledge of the objects involved in the pick application. Information about the objects is provided by the user by means of the scan process: the user places the object in a particular position over a scanning mat in order for the robot, with camera in hand, to perform a scanning trajectory around it. From this scan two data types can be obtained:

  • RGB images
  • Point clouds

Instance segmentation models are learnt from scan-based RGB images (in the case of the solution K-pick texture). Point clouds are used to build 2.5D models that can then be used for pose estimation of the items in the pick scene.

K-Pick is suitable in contexts where we are dealing with a few known objects. As such:

  • Objects are available to the user and scanning them is not particularly costly (in the sense that we are dealing with a small number of objects).

U-Pick is suitable in contexts where:

  • In pick: objects / items to be picked are not known or, if they are, there is a large number of them, making the scan process time consuming and the training process very hard (in the case that we are dealing with many references). Robot pick pose is estimated based on the point-cloud surface of the object.
  • In place: place is not oriented (i.e., undefined). Place position and orientation of the object is only approximate.

Unlike K-Pick, the U-Pick approach builds on a pre-trained model.

Depending on the needs of the process you should use one or the other. U-Pick for cases when the objects to be picked don't need a specific place orientation, the objects don't have to be classified and the pick point can be automatically determined by the system

Card: U-pick Card: K-Pick

Here are some common examples for each type of module

U-Pick:

  • Emptying a box
  • Moving items from conveyor

K-Pick:

  • Sorting items
  • Picking from a specific point
  • Oriented place