Author Identifiers

Shehan Caldera
ORCID: 0000-0002-3157-9145

Date of Award


Degree Type


Degree Name

Master of Engineering Science


School of Engineering

First Advisor

Associate Professor Alexander Rassau

Second Advisor

Dr Douglas Chai


Recent advancements in Deep Learning have accelerated the capabilities of robotic systems in terms of visual perception, object manipulation, automated navigation, and human-robot collaboration. The capability of a robotic system to manipulate objects in unstructured environments is becoming an increasingly necessary skill. Due to the dynamic nature of these environments, traditional methods, that require expert human knowledge, fail to adapt automatically. After reviewing the relevant literature a method was proposed to utilise deep transfer learning techniques to detect object grasps from coloured depth images. A grasp describes how a robotic end-effector can be arranged to securely grasp an object and successfully lift it without slippage. In this study, a ResNet-50 convolutional neural network (CNN) model is trained on the Cornell grasp dataset. The training was completed within 30 hours using a workstation PC with accelerated GPU support via an NVIDIA Titan X. The trained grasp detection model was further evaluated with a Baxter research robot and a Microsoft Kinect-v2 and a successful grasp detection accuracy of 93.91% was achieved on a diverse set of novel objects. Physical grasping trials were conducted on a set of 8 different objects. The overall system achieves an average grasp success rate of 65.0% while performing the grasp detection in under 25 milliseconds. The results analysis concluded that the objects with reasonably straight edges and moderately pronounced heights above the table are easily detected and grasped by the system.