3D pose estimation of transparent objects for robot manipulation

Ovaj rad istražuje problem određivanja trodimenzionalnog oblika i položaja transparentnih objekata za potrebe robotske manipulacije. Predstavljena je metoda ClearGrasp temeljena na neuronskim mrežama koja se koristi za detekciju transparentnih objekata. Metoda uključuje estimaciju površinskih normi, rubova i maski transparentnih objekata. U radu se korištenjem rekonstruiranog oblaka točaka dobivenog primjenom ClearGrasp metode uz pomoć metoda Open3D biblioteke dolazi do centra i oblika objekta koji daju informaciju o položaju i veličini objekta. Izračunati položaj i veličina objekta su korišteni pri manipulaciji objektom kako bi robotski manipulator znao gdje se objekt nalazi i na koji način ga uhvatiti. Eksperimentalni postav je opisan, uključujući korištene hardverske komponente i programski okvir. Nakon toga su prikazani rezultati eksperimenta, koji uključuju evaluaciju uspješnosti detekcije i manipulacije transparentnih objekata.This paper explores the problem of determining the three-dimensional shape and pose of transparent objects for robotic manipulation purposes. A neural network-based method ClearGrasp is presented for the detection of transparent objects, involving the estimation of surface normals, edges, and masks. In this paper, the object's center and shape were determined by employing the reconstructed point cloud, which was generated through the application of the ClearGrasp method with the assistance of Open3D library methods. The calculated position and size of the object were then utilized in the manipulation of the object, allowing the robotic manipulator to be aware of the object's location and how to grasp it. The experimental setup is described, including the hardware components and the software framework utilized. Subsequently, the experimental results are presented, which include an evaluation of the success in detecting and manipulating transparent objects