Object stiffness recognition with descriptors given by an FPGA-based tactile sensor

Object recognition by robots is commonly based on shape, texture, and stiffness. A few approaches that involve the use of different sensors have been followed to estimate the stiffness. Since tactile sensors register the force distribution along the contact surface and are commonly required in manipulative tasks, it would be desirable to obtain as much information as possible from them. Specifically, the raw moments of the tactile image provided by a tactile sensor can be used as descriptors to classify deformable objects that are squeezed by a gripper or artificial hand. The provided moments are used as inputs of a classifier. A smart tactile sensor with electronics based on an FPGA can implement the circuitry to compute the moments at a high rate. This reduces the workload of a central processor as well as the traffic through the communication buses, and allows real time operation. This paper shows experiments where a set of objects are classified with this procedure and a smart sensor. Two nearest neighbor classifiers, one non-supervised and the other supervised, are applied and an 86.7% accuracy is achieved.