CNN-Based Methods for Object Recognition With High-Resolution Tactile Sensors

Abstract

Novel high-resolution pressure-sensor arrays allow treating pressure readings as standard images. Computer vision algorithms and methods such as Convolutional Neural Networks (CNN) can be used to identify contact objects. In this paper, a high-resolution tactile sensor has been attached to a robotic endeffector to identify contacted objects. Two CNN-based approaches have been employed to classify pressure images. These methods include a transfer learning approach using a pre-trained CNN on an RGB-images dataset and a custom-made CNN (TactNet) trained from scratch with tactile information. The transfer learning approach can be carried out by retraining the classification layers of the network or replacing these layers with an SVM. Overall, 11 configurations based on these methods have been tested: 8 transfer learning-based, and 3 TactNet-based. Moreover, a study of the performance of the methods and a comparative discussion with the current state-of-the-art on tactile object recognition is presented.

Smart Tactile Sensing Systems Based on Embedded CNN Implementations

“…In Alameh et al's research [3], transfer learning was used to classify touch modalities obtained through a small 4 × 4 piezoresistive sensory array, by transforming tensorial data into images and then using different CNN models trained on ImageNet [13]. In Gandarias et al's research [5], they used a light CNN based (only three convolutional layers inside) on AlexNet, to identify 22 objects using their pressure map, collected from a 28 × 50 tactile sensory array. Other works include those in [17][18][19].…”

“…Targeting the classification of tactile data, the use of the dataset collected in [5] was considered. Tactile data were collected by a high resolution (1400 pressure taxels) tactile array, which was attached to the 6 DOF robotic arm AUBO Our-i5 [5]. A set of piezoresistive tactile sensors was distributed with a density of 27.6 taxels/cm 2 , forming a matrix of 28 rows by 50 columns.…”

“…Due to computational and memory limitations in the embedded application, a light CNN model was required to perform classification tasks with high accuracy and fewer parameters. In this work, we chose to use one of the models implemented in [5] as a base model to classify the objects in the aforementioned dataset. Among all the implemented networks, we chose to use the custom network TacNet4 because it was the best network that fit the embedded application (fewer parameters with high accuracy [5]).…”

“…In this work, we chose to use one of the models implemented in [5] as a base model to classify the objects in the aforementioned dataset. Among all the implemented networks, we chose to use the custom network TacNet4 because it was the best network that fit the embedded application (fewer parameters with high accuracy [5]). The model was based on AlexNet, which is usually used in computer vision for object classification [23].…”

“…It proposes an optimized CNN model, adopted from Gandarias et al's research [5], based on reduced data, which demonstrates the ability to provide comparable results in terms of accuracy, i.e., 90.88%, with reduced hardware complexity. • It presents efficient implementations of the CNN model on different hardware platforms for embedded tactile data processing.…”