Automated depth dataset generation with integrated quality metrics for robotic manipulation

Abstract

This work introduces a fully automatic and adaptable pipeline for synthetic depth dataset generation that later on can be used for the training of deep learning algorithms for robotic manipulation. From any available set of 3D object model meshes, the pipeline outputs rendered depth image data with labeled grasp candidates represented as a set of grasping points relative to the camera with associated metrics. The proposed pipeline allows adaptability in various characteristics such as the input dataset of objects, the sampling method, the gripper type, or the grasp evaluation metrics to allow the generation of a more customized, task-oriented collection of labeled grasps relevant for different robotic applications. The implementation is done using Blender’s Python API workspace, avoiding the use of multiple software tools or libraries, reducing the pipeline complexity, facilitating the extensibility, and providing benefits in terms of visualization and debugging.

Peer Reviewed

Postprint (author's final draft)