Baghaei Naeini, Fariborz (2020) Neuromorphic vision-based tactile sensor for robotic grasp. (PhD thesis), Kingston University, .
Abstract
Tactile sensors are developed to mimic human sense of touch in robotics. The touch sense is essential for machines to interact with environment. Several approaches have been studied to obtain rich information from the contact point to correct robot’s actions and acquire further information about the objects. Vision-based tactile sensors aim to extract tactile information by observing the contact point between the robot’s hand and environment and applying computer vision algorithms. In this thesis, a novel class of vision-based tactile sensors is proposed, "Neuromorphic Vision-Based Tactile Sensor" to estimate the contact force and classify materials in a grasp. This novel approach utilises a neuromorphic vision sensor to capture intensity changes (events) in the contact point. The triggered events represent changes in the contact force at each pixel in microseconds. The proposed sensor has a high temporal resolution and dynamic range which are suitable for high-speed robotic applications. Initially, a general framework is demonstrated to show the sensor operations. Furthermore, the relationship between events and the contact force is presented. Afterwards, methods based on Time-Delay Neural Networks (TDNN), Gaussian Process (GP) and Deep Neural Networks (DNN) are developed to estimate the contact force and classify objects material from the accumulation of events. The results indicate a low mean squared error of 0.17N against a force sensor for the force estimation using TDNN. Moreover, the objects materials are classified with 79.12% accuracy which is 30% higher compared to piezoresistive force sensors. This is followed by an approach to preserve spatio-temporal information during the learning process. Therefore, the triggered events are framed (event-frames) within a time window to preserve spatial information. Afterwards, multiple types of Long Short-Term Memory (LSTM) networks with convolutional layers are developed to estimate the contact force for objects with different size. The results are validated against a force sensor and achieve a mean squared error of less than 0.1N. Finally, algorithmic augmentation techniques are investigated to improve the networks accuracy for a wider range of force. Image-based and time-series augmentation methods are developed to generate artificial samples for training the network. A novel time-domain approach Temporal Event Shifting (TES) is proposed to augment events by preserving the spatial information of events. The results are validated on real experiments which indicate that time-domain and hybrid augmentation methods improve the networks’ accuracy significantly considering an object with a different size.
Actions (Repository Editors)
 |
Item Control Page |