Task-specific grasping of simiiar objects by probabiiistic fusion of vision and tactiie measurements

This paper presents a probabilistic approach for task-specific grasping of novel objects from a known category. RGB-D imaging is used to establish an initial estimate of the target object´s shape and pose, which is used to plan an optimal grasp over the uncertain estimate. Tactile information is then used for incrementally improving the estimate and sequentially replanning better grasps. The resulting grasp is maximally likely to be task compatible and stable taking into account shape uncertainty in a probabilistic context. Experimental results in simulation and on a real platform show that tactile information can be used for improving the stability of grasps for objects which belong to a known category even if they vary considerably in shape.