RGB-D Based Deep Learning Methods for Robotic Perception and Grasping

Abstract

The recent advancements in robotic perception have vested robotic grasping with learning capabilities. During the past decade, empirical methods on grasp detection have been preempted by the data-driven methods highlighting the potential of deep learning and computer vision. Majority of the studies are focused on 2D information from RGB and depth cameras as the input. The ability to learn on 3D point clouds has been a breakthrough which overcome crucial challenges in 2D learning methods. Point cloud based learning endows robustness to occlusions, lack of texture and propagation of projection errors, in grasp detection algorithms. In modern state of the art, 3D model-based learning methods feature 6DoF pose estimation and instant semantic segmentation while other methods utilize deep generative methods. In this context, this thesis studies on synthetic dataset generation, implementation of algorithm and benchmarking grasp manipulation of point cloud based deep learning methods. Two grasp detection pipelines featuring each of the model-based and model-free approaches are implemented and comparatively analyzed in this thesis. The analysis provides an overview on 6DoF grasp pose detection on cluttered multi-object scenes and implicates on adaptation to real-world grasp manipulation tasks such as robotic assembly.