No abstract

h i g h l i g h t s• A probabilistic approach for task-specific category based grasping is proposed.• The grasp stability is maximized probabilistically over shape uncertainty.• The approach integrates information over all training objects for better generalization.• The technique can cope with a sparser training set than most data-driven methods.• Only incomplete point clouds obtained from a single RGB-D image are needed. a b s t r a c tThe problem of finding stable grasps has been widely studied in robotics. However, in many applications the resulting grasps should not only be stable but also applicable for a particular task. Task-specific grasps are closely linked to object categories so that objects in a same category can be often used to perform the same task. This paper presents a probabilistic approach for task-specific stable grasping of objects with shape variations inside the category. An optimal grasp is found as a grasp that is maximally likely to be task compatible and stable taking into account shape uncertainty in a probabilistic context. The method requires only partial models of new objects for grasp generation and only few models and example grasps are used during the training stage. The experiments show that the approach can use multiple models to generalize to new objects in that it outperforms grasping based on the closest model. The method is shown to generate stable grasps for new objects belonging to the same class as well as for similar in shape objects of different categories.

No abstract

No abstract

Among all senses the sense of touch is the only one without which humans are not able to manipulate objects. Similarly, tactile sense is invaluable for robotic manipulation in uncertain environments. It is however not thoroughly understood to what extent properties of the robot environment can be inferred from the tactile sense. This paper presents a novel approach that allows to study how much information a robot can optimally learn from a single tactile exploration attempt. Our method makes use of a simulator as an internal memory for the robot. The evaluation is based on assessing how much information error minimization between predicted and actual sensor readings can provide about the environment. This paper focuses on evaluating geometric parameters in a transportation task. Experiments performed with a set of objects with various shapes indicate that a single exploration action is not guaranteed to provide much information for all uncertain factors if the attempt is not originally planned for information gain in mind. Moreover, the information gain for different attributes varies significantly depending on the object geometry.