Enhancing Adaptive Grasping Through a Simple Sensor-Based Reflex Mechanism

This paper presents an approach to achieve adaptive grasp of unknown objects whose position is only approximately known via point-cloud data. We exploit the adaptability of a soft robotic hand which can autonomously conform to the shape of a grasped object if properly approached. Once a grasp approach has been preliminarily planned based only on rough estimates of the object position, the hand is shaped to a pregrasp configuration. Before closing the hand, a sensor-based algorithm is applied that corrects the relative hand-object posture so as to enhance the probability that the object is uniformly approached by all fingers, thus avoiding undesired premature contacts. The algorithm minimizes the distance between the hand's fingerpads and the object by continuously controlling both the wrist pose and orientation and the hand closure. Experimental studies with a Kuka-LWR arm and a Pisa/IIT Softhand illustrate the benefit of the developed technique and the improvement in grasping performance with respect to open-loop execution of grasps planned on the basis of prior RGB-D cues only.