From belt picking to bin packing

We face the problem of computer-vision aided robot grasping of objects with more or less random positions. This field is of vital importance in the further progress in flexible automation of industrial processes, since conventional methods using fixtures and/or vibration bowls are expensive and inflexible. We study various types of disorder: A) visually isolated objects lying in distinct resting modes on a flat homogenous conveyer belt, B) partially occluded objects lying in distinct resting modes on a flat homogenous conveyer belt, C) visually separated objects, unrestricted object-camera pose, and fully surrounded by background, D) partially occluded objects, unrestricted relative orientation, but with a sizeable fraction of their contour detectable using foreground-background separation, E) partially occluded objects with unrestricted pose and no help from foreground-background separation. The cases A), B), and - to some extend - D) are encountered in belt picking, while case E) is true bin picking. Since physical storage of products and components in industry is based on deep containers with many layers of somewhat disordered objects, the belt-picking concept is only the first step for achieving flexible, unsupervised parts feeding. We have developed and tested a generic, fast, and easily trainable system for the cases A) and B). The system is unique because it handles the perspective effects exactly so there is no restriction concerning object dimensions relative to the distance to the camera. We report on a strategy to be used in treating case C) using the principles developed for the cases A-B). We discuss possible strategies to be employed when going all the way to cases of D) and E).