Scalable Learning of 6-DoF Object and Robotic Grasp Poses

Abstract

This dissertation addresses the problem of 6-DoF Object Pose and 6-DoF Grasp Pose estimation from visual sensor data, which is crucial for tasks such as robotic manipulation and Augmented Reality. We present novel learning-based methods that are fast, reliable, and scalable concerning training data, test environments, and target objects. Instead of relying on pose annotated data, we train our models in simulation which provides and abundant source of variably steerable data with exact 3D annotations.