Experimental study on model- vs. learning-based slip detection

Vision and proprioception are traditional sources of information for robotic grasping, but they are insufficient to achieve a stable grasp without slippage or without applying an excessive force on the object. Tactile sensors can aid in this problem by providing spatial and temporal data on the contact between fingertips and object. In this work, tactile fingertip sensors are used to detect slippage through two separate methods: the first, using principles inspired by human tactile sensing, and the second, by using a convolutional neural network trained with suitably labeled test samples. To perform a fair comparison of the methods, two evaluations are performed using a test bench and a pick-and-place robotic application. Results show promising use of the model-based method to avoid translational slippage, as it was able to consistently keep objects from slipping without overloading the grasp. Limitations of both model- and learning-based approaches are identified and discussed.

I. Introduction

Slip detection and slip prediction (or incipient slip detection) are techniques still under study and development. Different tactile sensors and data processing methods have been employed in research to precisely detect slippage between the robot's fingers and grasped object [1]. Although many different technologies can be used to create tactile sensors, most of them work through detection of vibrations (using IMUs, for instance), pressure sensing (capacitive, piezoresistive) or optical tracking of deformations. A review of tactile technologies, including their advantages and disadvantages on different tasks, is provided in [2], while Table I summarizes the features for commercially available tactile sensors. Different sensors excel at different tasks, and multimodal sensors increase the robustness of the systems. Processing of the tactile signals follows two large trends, “model-based” approaches based mainly on frequency-domain analysis of the data, or “learning-based” approaches that employ machine learning techniques to interpret “tactile images” formed by sensor arrays, or to extract high-level features from low-level tactile information.