Even though grasping has been an interesting and crucial topic in robotics for a long time, robots still have great difficulty in picking up arbitrary objects when they face open, unknown environments or under uncontrolled conditions. Solving these issues would make using robots fit for many repetitive scenarios such as checking and restock supermarket aisles, tidying up households, dispatching mail orders at distribution centres, collecting ripe fruits, or doing ecological pest control by selectively removing bugs. To become reality, robots have to learn to grasp as reliable as humans. Many people have tried to discover the tricks people unconsciously employ when they cannot rely on their perception, and transfer these insights to robots. As a result, a new research and development conducted with a new vision to this problem introducing a novel approach for training a robot how to grasp, using topographic features of the objects, especially developed for grasping without requiring a prior knowledge of the objects, called Height Accumulated Features (HAF) and Symmetry Height Accumulated Features (SHAF). This method abstracts topographic information from perceived surfaces of objects hence enables to learn how to grasp them, even if they are unknown or on a heap of other objects. An important and inseparable part of the core of such an approach is the decision part, where the perceived scenes are interpreted into abstract models and need to be classified, to help the rest of the core to make a conclusion and find the best suitable points to grasp an object. This thesis proposes improvements and optimisations on this part of the core, in three different stages, extending machine learning training instances, introducing new feature set definitions, to be used as either substitutions or complements to HAF and SHAF, and selecting the best classification algorithm, to be employed for the grasping using topographic features. Three new feature set definitions, Circular Feature Definition (CF), Differential Intra-Circular Feature Definition (DICF) and Bell-Circular Feature Definition (BCF) try to summarize the scene in an abstract form with enormously less data dimensions, compared the combination of HAF-SHAF, not only to maintain and even improve the accuracy of the outcome, but also to increase efficiency by reducing the computation complexity. Furthermore, the extended data set is used in order to put the new features, as well as HAF and SHAF into test, with different classification algorithms including Support Vector Machines, Decision Trees, k Nearest Neighbors and Random Forest, in order to find the best combination of (feature set definition, classification algorithm) for the problem of grasping using topographic features. Finally, the results of the experiments are presented and compared to provide more insight into the performance of each setting.