6-DoF Grasp Learning in Partially Observable Cluttered Scenes

Abstract

The key element of the efficient interaction of an intelligent robot with its im- mediate environment is object manipulation - a task that current data-driven methods reshape into various methods aimed at object localization, classification, segmentation, and grasp pose estimation. This work is concerned with the grasp pose estimation, namely with the implications of 6-DoF grasp pose estimation for partially visible cluttered scenes. In this thesis, two methods are proposed to address the problem of collision management of the grasp proposals and the full target scene due to the partial visibility and cluttered nature of a scene. The first explores the possibility of embedding input data with differential geometrical shape information, namely the modified mean curvature measure, to improve the qualitative results of grasp estimation. The second method proposes a supervisor network architecture termed Collision-GraspNet that classifies grasp proposals with respect to collision with the scene, including its occluded parts, and improves the invalid proposals via iterative pose sampling. The first proposed approach is tested on the Contact-GraspNet model and compared with GraspNet architecture baseline performance. In its turn, Collision-GraspNet is compared with an analytical proposal filtering approach employed by GraspNet, and evaluated in three stages using various datasets. Grasp supervisor architecture Collision-GraspNet outperformed the respective analytical approach and showed high confidence threshold flexibility. However, curvature-embedded data failed to improve upon the baseline model performance.