Reflectance Estimation for Pre-grasping Distance Measurement using RGB and Proximity Sensing

Gustavo Alfonso Garcia Ricardez, Zihan Yue, Yosuke Suzuki, Tadahiro Taniguchi

Abstract: Before grasping an object, robots should determine where is best to grasp it to increase the picking success rate. To search for feasible grasping points, placing proximity sensors based on the reflection of light at the tip of the gripper's fingers can be used to obtain the distance to objects (pre-grasping distance). However, the proximity information output by the sensor is sensitive to the object's reflectance, which leads to an error in the pre-grasping distance. In this study, we use neural network regression to estimate the reflectance of objects based on RGB images so that the pre-grasping distance error is reduced, assuming that the reflectance statistically depends on object category regarding daily objects found in the service robot environments. We created a dataset with images of 40 objects for which we measured the reflectance. Then, we trained two regressors based on well-known architectures. In the experiments, we compare the accuracy in the distance estimation for our proposed method and the conventional fixed reflectance method with both known and unseen objects. The proposed method significantly reduces the pre-grasping distance error for objects with various reflectances compared to the conventional method. The results show that our proposed method can achieve a distance estimation error under 2 [mm] at distances of 20 [mm] or less.