Effects of Force-Torque and Tactile Haptic Modalities on Classifying the Success of Robot Manipulation Tasks

Yukyu Chan, Hungchen Yu, Rebecca P. Khurshid

Abstract

We investigate which haptic sensing modalities, or combination of haptic sensing modalities, best enable a robot to determine whether it successfully completed a manipulation task. In this paper, we consider haptic sensing modalities obtained from a wrist-mounted force-torque sensor and three types of fingertip sensors: a pair of FlexiForce force-sensing resistors, a pair of NumaTac sensors, and a pair of BioTac sensors. For each type of fingertip sensor, we simultaneously record force-torque and fingertip tactile data as the robot attempted to complete two manipulation tasks—a picking task and a scooping task—two-hundred times each. We leverage the resulting dataset to train and test a classification method using forty-one different haptic feature combinations, obtained from exhaustive combinations of individual modalities of the force- torque sensor and fingertip sensors. Our results show that the classification method’s ability to distinguish between successful and unsuccessful task attempts depends on both the type of manipulation task and the subset of haptic modalities used to train and test the classification method.

Link to paper

Robot Setup

The figure above is the setup of the experiment. We use a UR5 Robot (Universal Robots) with an attached two-finger gripper (Robotiq: 2F-85).

Sensor setup

The sensors setup we used includes: A two-fingered Robotiq Gripper with FlexiForce sensor (Left), NumaTac sensor (Middle), and BioTac sensor (right) mounted on each finger.

Picking task

The trajectory of the robot’s motion during the picking task is shown above. During this task, the robot starts above the object with its gripper open, moves down so that the object is between the fingers of the open gripper, closes its gripper, and returns to the starting position. The manipulated object was a soft wall door stop (National Hardware), which is a semi-sphere made from a flexible polymer. This object was chosen because it was difficult, but not impossible, for the robot to pick up, ensuring that there would be a good balance between successful and unsuccessful pick attempts. A trial was counted as a failure if the robot failed to pick up the object or if the robot dropped the object before returning to the starting position.

Scooping Task

The trajectory of the robot’s motion during the scooping task is shown above. During this task, the robot holds a scoop and moves through a pre- programmed trajectory to scoop pasta from a box, as shown in the figure above. During this task, the robot holds on to the scoop with a pinch grasp. However, the pressure readings for both the NumaTac and BioTac saturated when the robot grasped the scoop hard enough so it would not rotate in the robot’s hand around the axis connecting the robot’s two fingers. To address this problem, we added a ‘third finger’ to the robot’s hand to stabilize the scoop. The different types of possible failures are 1) The scoop hits the bottom of the box, causing an emergency stop due to excessive torque applied to the robot’s joints; 2) The scoop slips in the robot’s hand; and 3) No pasta is scooped up. Failures are caused by the variations in the position of the scoop in the robot’s hand and the distribution of pasta in the box.

Published Paper