On Transferability of grasp-affordances in data-driven grasping

It has become a common practice to use simulation to generate large databases
of good grasps for grasp planning in robotics research. However, the existence of a generic simulation context that enables generation of high quality grasps that can be used in several different contexts such as bin-picking or picking objects from a table, has to our knowledge not been discussed in the literature. In this paper we investigate how well the quality of grasps simulated in a commonly used ”generic” context transfer to a specific context where the object is placed on a table. We generate a large database of grasp hypothesis for several objects, which we then evaluate in different dynamic simulation contexts eg. free float (no gravity, no obstacles), standing on table and lying on table. We present a comparison on the intersection of the grasp outcome space across the different contexts and quantitatively show that to generate reliable grasp databases, it is often required to use context specific simulation.