Robotik

Using Uncertainty Quantification in 6DoF Pose Estimation to improve Gripper Performance in Robotic Applications

Abstract: Implementing robotic manipulation robustly in a real-world setting requires accurate and robust 6-DoF object pose estimations. One of the critical chal-lenges of implementing this task using deep neural networks is their inac-curacies when dealing with inputs outside the training domain. Therefore efficient and robust uncertainty quantification in pose estimators is critical. We implement a 6-DoF object pose estimation application that solves a children toy using a robot equipped with a parallel gripper. The task of the toy is to pick up known objects of different geometric shapes and insert them into a matching shape in a container. The openings in the container are only a few millimeters larger than the corresponding object. Due to this slim margin for error, we require high-quality pose predictions. Using the application as an example, the gripper performance is optimized through uncertainty quantification. We estimate predictive uncertainty using ensembles. Three different ensemble generation methods based on changing the members weights, changing the members architectures, and changing the training data sources of the members are being tested. In multiple experi-ments, we are able to eliminate false positive detections completely. Using the ensembles, we can detect the objects in 87% of the attempts with an ADD error below 1 cm. In real-world scenes with three images from differ-ent viewpoints, we successfully identify the best resulting pose estimations using uncertainty quantifications in most examples.