Single-grasp deformable object discrimination: the effect of gripper morphology, sensing modalities, and action parameters

TL;DR

Tactile sensors can provide a key advantage even if recognition is based on stiffness rather than shape, and generalization across embodiment or grip configurations very difficult.

Abstract

In haptic object discrimination, the effect of gripper embodiment, action parameters, and sensory channels has not been systematically studied. We used two anthropomorphic hands and two 2-finger grippers to grasp two sets of deformable objects. On the object classification task, we found: (i) among classifiers, SVM on sensory features and LSTM on raw time series performed best across all grippers; (ii) faster compression speeds degraded performance; (iii) generalization to different grasping configurations was limited; transfer to different compression speeds worked well for the Barrett Hand only. Visualization of the feature spaces using PCA showed that gripper morphology and action parameters were the main source of variance, making generalization across embodiment or grip configurations very difficult. On the highly challenging dataset consisting of polyurethane foams alone, only the Barrett Hand achieved excellent performance. Tactile sensors can thus provide a key advantage even if recognition is based on stiffness rather than shape. The data set with 24,000 measurements is publicly available.

Figures (8)

• Figure 1: Physical objects datasets. (a) Ordinary Objects set approximately spread out on the elasticity and contact width axes (reference values for this object set are not available). (b) Polyurethane Foams set.
• Figure 2: Robot hands and grippers. Two parallel jaw / two-finger grippers were employed: Robotiq 2F-85 and OnRobot RG6, with gripper position and effort (current/force) feedback. The qb SoftHand has five fingers but only one motor and its position and current as feedback. Two action configurations (a1: minimizing contact with thumb, a2: maximizing contact with thumb) are shown. The Barrett Hand has three fingers that can be rotated around the wrist, 96 tactile sensors, 3 fingertip torque sensors, and 8 joint encoders. We used two finger configurations (a1: opposing fingers, a2: lateral fingers).
• Figure 3: Experimental setup illustration. (a) qb SoftHand action1; (b) qb SoftHand action2; (c-d) Robotiq 2F-85; (e-f) OnRobot RG6; (g) Barrett Hand action1; (h) Barrett Hand action2.
• Figure 4: Visualization of raw data for individual grippers. The effect of action parameters and the grasped objects is illustrated. For the qb SoftHand (a), Robotiq 2F-85 (b), and OnRobot RG6 (c) gripper, we present the normalized current and normalized position during the grasping of three different objects (yellowsponge, yellowcube, and whitedie) using two different velocity settings. For the Barrett Hand (d)-(f), we show the progression of tactile, position, and effort sensors during grasping for two different finger configurations with a joint velocity of $v = 0.6$ rad/s. The yellowsponge and yellowcube are made of the same material, while the yellowcube and whitedie have roughly the same size.
• Figure 5: Classification accuracy -- grippers, classifiers, and action parameters. Subfigures (a)--(d) Objects set; (e)--(g) Polyurethane foams set. Rows represent different classifiers, and columns represent action parameters. Chance level performance: Objects set: 11%; Foams set: 5%.