A probabilistic framework to detect suitable grasping regions on objects

This work relies on a probabilistic framework to search for suitable grasping regions on objects. In this approach, the object model is acquired based on occupancy grid representation that deals with the sensor uncertainty allowing later the decomposition of the object global shape into components. Through mixture distribution-based representation we achieve the object segmentation where the outputs are the point cloud clustering. Each object component is matched to a geometrical primitive. The advantage of representing object components into geometrical primitives is due to the simplification and approximation of the shape that facilitates the search for suitable object region for grasping given a context. Human demonstrations of predefined grasp are recorded and then overlaid on the object surface given by the probabilistic volumetric map to find the contact points of stable grasps. By observing the human choice during the object grasping, we perform the learning phase. Bayesian theory is used to identify a potential object region for grasping in a specific context when the artificial system faces a new object that is taken as a familiar object due to the primitives approximation into known components.