Grasp2Hardness: fuzzy hardness inference of cylindrical objects for grasp force adjustment of force sensor-less robots

Service robots frequently operate various cylindrical objects with unknown physical properties, which demands the grippers of robots being equipped with force sensors to control grasp force. But force sensors are unnecessary and expensive for imprecise grasp force control for most operations in domestic environment. So as a substitute, this paper introduced the fuzzy hardness (FH) for imprecise grasp force evaluation. In addition, a method to infer the FH of objects was proposed, through vision and supervised learning. In this method, the deformation of objects related to the close degree of gripper was treated as a key variable and measured via visual methods. Based on the measured deformation data, long short-term memory network (LSTM) was introduced to conduct supervised learning synchronously. Then, several predicted deformation curves can be obtained through these LSTM blocks. Subsequently, the FH of objects would be clear when the errors between measured data and the predicted ones were calculated from the curves. The verification experiments showed that the maximum inference accuracy can reach 100% on TPU(80A) with 2 mm wall thickness. Moreover, after FH being applied, the deformation of TPU(80A) objects with 2 mm wall thickness decreased approximately 84.4% compared with using classical method. And all these results indicate that the FH inference method can be applied to adjust the grasp force for force sensor-less robots.