Spatial constraint identification of parts in SE3 for action optimization

In this paper we present a method to structure contextual knowledge in spatial regions/manifolds that may be used in action selection for industrial robotic systems. The contextual knowledge is build on relatively few prior task executions and it may be derived from either teleoperation or previous action executions. We argue that our contextual representation is able to improve the execution speed of individual actions and demonstrate this on a specific time-consuming action of object detection and pose estimation. Our contextual knowledge representation is especially suited for industrial environments where repetitive tasks such as bin-and belt picking are plentiful. We present how we classify and detect the contextual information from prior task executions and demonstrate the performance gain on a real industrial pick-and-place problem.