Task-Oriented Visual Object Pose Estimation for Robot Manipulation: A Modular Approach
Abdelrahman, Ahmed So, Peter Le, Hoan Quang Swikir, Abdalla Haddadin, Sami
Abdelrahman, Ahmed
So, Peter
Le, Hoan Quang
Swikir, Abdalla
Haddadin, Sami
Supervisor
Department
Robotics
Embargo End Date
Type
Conference proceeding
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This paper presents a general method for object pose estimation from RGB-D camera data for robot manipulation tasks. We fine-tune off-the-shelf image detection models to recognize certain objects in color images then combine the result with point cloud information to estimate 3D object positions in a task-agnostic approach. By utilizing prior information about our manipulation task, we further estimate object orientations using additional heuristics. We demonstrate our approach and evaluate its performance on an electronic task board and release our adaptable and easy-to-integrate implementation as a re-usable software module under https://github.com/eurobin-wp1/tum-tb-perception.
Citation
A. Abdelrahman, P. So, H. Q. Le, A. Swikir, and S. Haddadin, “Task-Oriented Visual Object Pose Estimation for Robot Manipulation: A Modular Approach,” Springer Proceedings in Advanced Robotics, vol. 36 SPAR, pp. 242–248, 2025, doi: 10.1007/978-3-031-89471-8_37
Source
Springer Proceedings in Advanced Robotics
Conference
16th European Robotics Forum, ERF 2025
Keywords
computer vision, object detection, pose estimation, RGB-D, transfer learning
Subjects
Source
16th European Robotics Forum, ERF 2025
Publisher
Springer Nature