Rotational Direction Detection Using Tactile Sensor and External Camera

Abstract

The data-driven grasping methods have become popular recently and it is not necessary to analytically understand the physics of an object. However, when a robot arm lifts an object (especially a heavy object) with a gripper, in order to make a stable grasp, the grasping torque between the robot fingers and the object should be considered. This torque is related to the displacement between current grasping position and the object's Center-of-Mass (CoM). High grasping torque may exceed the hardware strength limits and lead to serious robot damage, or destroy the object. Inspired by human grasping behavior, this paper presents a novel yet intuitive approach. Through analyzing a sequence of images recorded by an external camera and a GelSight sensor, a robot can detect the rotational direction of an object without any pre-knowledge of the object's physical parameters. To evaluate the performance of our proposed method, we test 8 unseen objects in 1,128 grasps. A rotational direction detection accuracy of 88.28% is achieved. Instead of measuring the grasping torque with a torque sensor, by using the detected rotational direction as feedback, a robot can continuously optimize and update its grasping strategy, and perform re-grasping manipulation to approach the object's CoM and decrease the grasping torque. Our method can enable a closed-loop vision-tactile-based feedback control to apply a stable grasp with proper grasping position and torque.