GraspTrack: Object and Grasp Pose Tracking for Arbitrary Objects

Abstract

A necessary skill for robots for human-robot-collaboration in Industry 4.0 scenarios is the manipulation of objects. To manipulate objects, the robot must estimate grasp poses autonomously. Object-agnostic approaches are often realized through frame-based deep-learning methods that take a single frame of information (image or point cloud) as input while discarding previous estimates, even if they were verified by execution. In addition, object-agnostic grasp estimation methods lack the ability to target specific objects in cluttered scenes. To address this, we propose GraspTrack – a class-agnostic pipeline that uses instance segmentation and object pose estimation to track estimated grasps for each object simultaneously. To evaluate our pipeline and its modules, we perform extensive experiments on the GraspNet-1Billion dataset and extend its evaluation to measure grasp quality for each object, capturing the ability to grasp targeted objects more effectively. Our experiments prove our pipeline to be robust against difficult viewing angles and occlusions, outperforming frame-based grasp pose estimation.