Research about tasks that robotic maids should be able to perform is an emerging research area such as cooking and cleaning. Among them, manipulation of clothes is one of the hardest tasks due to the fact that textile is highly deformable and it is hard to model a good grasping point on them. In literature there are certain algorithms depending on 3D information of the cloth but most of them are not robust. Among them, Fast Integral Normal 3D (FINDDD) descriptors is a promising way for finding grasping points around previously detected collar of a polo shirt but, it fails to select the position in the best way possible. In this work, an improvement to the FINDDD descriptor selected grasping points, using vector autoregression model is proposed. Previously calculated FINDDD scores will be used as features and the system will be able to find a grasping position which is better to grasp a polo shirt. With better grasping positions, the robot will be able to perform further tasks such as folding, hanging to a hook etc. To train the system, a training set is formed from the samples of the simplest placement of the polo shirt. Then, the trained system is tested on 7 different positions and placements of the polo shirt. To form the training set, we have developed a new interaction method based on the use of a sensor that captures natural hand movements. With use of this method, the robot is teleoperated using natural hand movements instead of direct manipulation or haptic devices as usage of natural hand movements enhances the user experience and eases the data collection process. The system is evaluated based on the distance between the user given grasp positions and the system output grasp positions. Further evaluation is done by checking whether the new grasping positions indeed better positions by analysing the real world images. Even though, the system works well for the spread open cases, the results obtained are prone to placement changes especially when wrinkles are introduced.