MonoEye: Exploring Monocular and Multi-View Neural Reconstruction Pipelines for Robotic Grasping
Monocular vision-based grasping presents a promising yet inherently challenging problem in robotics. While depth sensors and stereo cameras have traditionally been employed to facilitate grasp planning, recent advances in deep learning—particularly in monocular depth estimation, 3D scene reconstruction, and vision-based affordance prediction—have opened new avenues of research. This thesis investigates novel pipelines that leverage state-of-the-art monocular depth models alongside neural splatting techniques to enable robust robotic grasping from a single image, even under challenging conditions such as transparency, reflectivity, or scene clutter. We conduct a comprehensive evaluation of contemporary methods, propose a pipeline that utilizes Gaussian splatting for rapid and accurate 3D reconstruction, and validate its performance on both simulated and real-world benchmarks. Our experiments demonstrate that a well-designed monocular pipeline can closely match the performance of more expensive depth or multi-camera systems, thereby enabling more cost-effective and flexible robotic perception solutions.
Advisor: Professor Constantinos Chamzas (WPI)
Committee: Professor Nitin Sanket (WPI) and Professor Berk Calli (WPI)
Zoom: Zoom meeting has ended.