<p>Robot grasping is a critical area of research in robotics, which aims to develop algorithms and strategies that enable robots towards reliable grasping and manipulation of objects in various environments. The ability to grasp and handle objects dexterously has wide-ranging applications in manufacturing, logistics, and healthcare. Recent advances in deep learning have paved the way for learning-based approaches toward autonomous grasping and intelligent manipulation in structured and unstructured environments. However, majority of the existing methods are designed for two-finger robotic grippers and cannot be extended towards application on three-finger robotic grippers. Three-finger grippers can perform a wide variety of grasps, and thereby require more nuanced grasp representation as compared to existing approaches. In this work, we propose a novel method towards representing underactuated three-finger grasps. A new grasp data set is also presented, specifically designed to train and evaluate three-finger robotic grasps using the proposed novel representation. A state-of-the-art deep learning architecture was trained on the dataset towards developing an end-to-end grasp planning pipeline. The proposed pipeline uses color and depth images as inputs to predict the ideal grasp pose for a three-finger robotic gripper to grasp the target object. The trained network was validated through simulation and hardware experiments, and the results were compared with an existing state-of-the-art grasp planning approach to highlight the potential of this new technique for grasp representation and planning.</p>