Automatic picking of randomly distributed objects from a bin has been denoted the "Holy Grail" in the world of robot automation. A particular property of the bin-picking scenario (in contrast to most other industrial robot applications) is that grasp errors are allowed to occur: Usually bin-pickers fill feeding stations such that occasional errors do not disturb the main process, since the feeding station has a buffer of a number of objects which it provides to the assembly process. Grasping errors are usually detected by haptic feedback. We intend to utilize the huge amount of grasp data generated in industrial bin-picking for grasp learning. This aim is achieved by using the novel concept of grasp densities (Detry et al., 2010). Grasp densities can describe the full variety of grasps that apply to specific objects using specific grippers. They represent the likelihood of grasp success in terms of object-relative gripper pose, can be learned from empirical experience, and allow the automatic choice of optimal grasps in a given scene context (object pose, workspace constraints, etc.). We will show grasp densities extracted from empirical data in a real industrial bin picking context, constituting (to our knowledge) one of the first examples of grasp learning in industrial robotics.
Main Research Area:
IEEE ICRA 2011 Workshop on: Uncertainty in Automation