A Fast and Robust Deep Learning Approach for Hand Object Grasping Confirmation

Abstract

One of the most important skills for service robots is object manipulation, which is still a challenging task. Since object manipulation is a hard task, it is relevant to know if an object was successfully grasped, avoiding future wrong decisions. Object grasp confirmation is commonly solved by using robotic sensors (infrared, pressure, etc.), but, in many cases, these sensors are not available for all robots. In contrast, depth and RGB sensor are present in almost all service robots. In this work a novel computer vision based method oriented to hand object grasp confirmation is proposed, which uses a deep learning network trained with depth maps. In order to measure the performance of the proposed method, experiments were performed using a single-arm manipulator service robot for both known and unknown objects. Experimental results show that the proposed approach correctly identifies 99% of both classes (object grasped or not grasped) with known objects and $$92\%$$ with unknown objects. The grasping confirmation method was added to the Storing Groceries task, for RoboCup@Home competition, improving its time performance.