Enhancing a robot gripper with haptic perception for risk mitigation in physical human robot interaction

Abstract

Utilising a two finger robot gripper in physical human robot interaction bears the risk of clamping fingers in the gripper. In this paper, we formulate a new grasp strategy which aborts grasps if a human body part is grasped instead of a workpiece. The strategy integrates a pressure-based haptic exploratory procedure seamlessly into the grasp process. It uses force and deformation data gathered in the exploratory procedure to distinguish human body parts from workpieces. We compare a support vector machine (SVM) and a random forest classifier for this task. The validation of the grasp strategy is carried out by grasping experiments with a two finger gripper in which a dummy hand and real human hands are used. Using this strategy grasps can be aborted without exceeding the maximum permissible grasp force for collisions with humans. The SVM classifier achieves an accuracy of 99.06% and a recall of 99.997% on our experimental data. Classification only takes 3.65 ms on embedded hardware. The SVM outperforms the random forest classifier.