CyBarKeeper: Design and implementation of a high-resolution haptic sensor system for a robotic gripper using embedded machine learning algorithms

Abstract:
The aim of the project presented in this paper was to embed high-resolution stress sensors in the gripping surfaces of a robot gripper to derive information about the quality/stability and property of the grip using the data thus obtained during the gripping process. Based on the sensor information, the robot's control software can decide whether and how the posi-tion of the gripper relative to the object to be gripped needs to be adapted. ASICs for surface stress measurement were embedded in an elastic silicone mold and mounted on a commercial robot gripper. The ASIC readout was realized using an FPGA development board, for which software was developed to configure the system and display the recorded data. To interpret sensor data, gripping recordings were used to design and train a neural network for the classification of gripping states. Although the neural network was quantized to 8-bit and compressed by more than 75 percent to fit on the FPGA platform, it achieved a classification accuracy of 93 percent for recognising three different grasping classes.