Design and development of a robust vision-based tactile sensor

Abstract

For robots to perform advanced manipulation of objects, touch is a critical source of information, and a highquality tactile sensor is essential. Image-based optical tactile sensors, and its inheritances, which have soft touch interfaces, can provide high-resolution tactile images of the contact geometry, contact pressure, and slip conditions. However, due to the lack of robustness provided by the current tactile sensors, the ability to grasp hard or sharp objects is minimal. In this work, we propose an image-based optical tactile sensor and overcome the above limitation of poor robustness by introducing a latex layer on the touch interface. We use a combination of silicone elastomer covered with a latex material and an acrylic sheet to support the silicone elastomer. A camera placed at the bottom of the sensor housing captures the deformation of the elastomer surface illuminated by an inner light. To evaluate the performance, we carried out a series of experiments. First, we evaluated the mechanical characteristics of the silicone elastomer with three types of coating, namely latex membrane, metallic coating, and no coating. The proposed latex membrane clearly outperformed the other two in terms of robustness. Second, we carried out the force-displacement experiments quantitatively to further study the sensitivity and robustness. Last, we validated the sensor performance in terms of its spatial resolution by applying the VGG-19 neural network for classifying touch patterns captured by the sensor. Overall, the proposed sensor achieved the desired robustness, sensitivity, and spatial resolution performance.

PnuTac: A Vision-Based Pneumatic Tactile Sensor for Slip Detection and Object Classification

“…Padmanabha et al [11] performed tactile state estimation for a representative insertion task using OmniTact, a thumb-shaped tactile sensor paired with a ResNet-based neural network. A modified VBTS with a robust sensing surface capable of manipulating hard and sharp objects, as well as performing object classification, was proposed in [22]. Sferrazza et al [23] propose a system based on green fluorescent particles that are randomly distributed within a soft elastic opaque coating, and deep learning is employed to extract information related to the force applied to the sensor's surface.…”