Tactile object class and internal state recognition for mobile manipulation

Abstract

Abstract-Tactile information is valuable in determining properties of objects that are inaccessible from visual perception. In this work, we present a tactile perception strategy that allows any mobile robot with tactile sensors in its gripper to measure a set of generic tactile features while grasping an object. We propose a hybrid velocity-force controller, that grasps an object safely and reveals at the same time its deformation properties. As an application, we show that a robot can use these features to distinguish the open/closed and fill state of bottles and cans -purely from tactile sensing -from a small training set. To prove that this is a hard recognition problem, we also conducted a comperative study with 17 human test subjects. We found that the recognition rate of the human subjects were comparable to our robotic gripper.

Online spatio-temporal Gaussian process experts with application to tactile classification

“…Moreover, to better understand how human subjects classify objects by touch with true-label feedback, we conducted experiments with fifteen human-subjects and derived some insights as to the strategies used. Our results add to recent findings of a similar experiment but with the traditional offline-training methodology [3], [4].…”

“…We tested STORK-GP on the one-step prediction task using three well-known benchmark problems i.e. the Mackey-Glass, Henon and Lorenz dynamical systems 3 . The objective was to predict the next state x (t+1) given x (t) .…”

“…A closely-related work is a recent study by Chitta et al [3], [4] which classified objects based on the high-frequency response using a decision tree. Similarly, we classified objects using tactile feedback during a grasping motion (without lifting the object).…”

“…We believe such an approach is more applicable to robots operating in the real-world where both novel and familiar items would be frequently encountered. Therefore, instead of using offline classifiers (such as C4.5 decision trees used in [3], [22] or neural networks [23]), we developed an online generative model that that does not require the construction of an extensive feature vector (such as the bag-offeatures approach used by Schneider et al [24]) and provides a probability distribution over object classes.…”

Object manifold learning with action features for active tactile object recognition

“…Alternatively, the differential geometry based feature [10] and the force-distance profiles based feature [11] for tactile sensor data could be used as the object parameters. Since these features are strongly related to the physical quantities such as a surface shape and hardness of the object, the suitable features need to be selected for the task a priori.…”