Torque control for grasping by learning experience and tactile feedback

Abstract

In robotic hand grasp control, torque output adjustment is one of most important problem. Grasping force optimization based on Coulomb friction cone needs complex calculation and exact modeling. In view of human grasp, precise manipulation depends on much more experience rather than concrete object model. We propose a torque control approach to output and adjust the optimal torque automatically based on experience learning and fusion sensing including tactile reading, joint angle change, and torque feedback. Firstly, pre-grasp can help to recognize and classify what kind of object it is and whether the robot has grasp experience for this object. If it is an unknown grasp, the output torque will be optimized and adjusted by updating the experience database through incremental learning. Experiments indicate that with the incremental learning, our robotic hand can easily calculate the optimal output torque and finally grasp novel object stably and safely.