Learning grasp stability based on haptic data

Grasping is an essential skill for a general purpose service\ua0robot, working in an industrial or home-like environment. The\ua0classical work in robotic grasping assumes that the object\ua0parameters such as pose, shape, weight and material properties\ua0are known. If precise knowledge of these is available, grasp\ua0planning using analytical approaches, such as form or force\ua0closure, may be enough for successful grasp execution. However, in unstructured environments the information is usually\ua0uncertain, which presents a great challenge for the current\ua0state-of-the-art work in this area.\ua0Sensors can be used to alleviate the problem of uncertainty.\ua0To determine the shape and pose of an object, vision has been\ua0commonly used. However, the accuracy of vision is limited\ua0and small errors in object pose are frequent even for known\ua0objects. It is not uncommon that even these small errors cause\ua0failures in grasping. These failures are also difficult to prevent\ua0at the grasp planning stage. This problem is magnified when\ua0also the object models are acquired on-line using vision or\ua0other similar sensors. While the tactile and finger force sensors\ua0can be used to reduce this problem, a grasp may fail even when\ua0all fingers have adequate contact forces and the hand pose is\ua0not dramatically different from the planned one.\ua0The main contribution of this paper is to show that it\ua0is possible to infer knowledge about grasp stability using\ua0information from tactile sensors while grasping an object\ua0before being further manipulated. This is very useful, because\ua0if failures can be detected, objects can be regrasped before\ua0trying to lift them. However, the relationship between tactile\ua0measurements and grasp stability is embodiment specific and\ua0very complex. For this reason, we propose to use machine\ua0earning techniques for the inference