Robotic Understanding of Scene Contents and Spatial Constraints

The aim of this paper is to create a model which is able to be used to accurately identify objects as well as spacial relationships in a dynamic environment. This paper proposed methods to train a deep learning model which recognizes unique objects and positions of key items in an environment. The model requires a low amount of images compared to others and also can recognize multiple objects in the same frame due to the utilization of region proposal networks. Methods are also discussed to find the position of recognized objects which can be used for picking up recognized items with a robotic arm. The system utilizes logic operations to be able to deduct how different objects relate to each other in regard to their placement from one another based off of the localization technique. The paper discusses how to create spacial relationships specifically.