Proceedings of the
9th International Conference of Asian Society for Precision Engineering and Nanotechnology (ASPEN2022)
15 – 18 November 2022, Singapore

Digital Twin-based Object Detection of Camera Mounted Robot Arm using Deep Reinforcement Learning

Su-Young Park1, Hyungjung Kim2, Junghyuk Lee1, Suhwan Jeong1 and Sung-Hoon Ahn1,2,a

1Department of Mechanical Engineering, Seoul National University, Seoul 08826, Republic of Korea

2Institute of Advanced Machines and Design, Seoul National University, Seoul 08826, Republic of Korea


With the advancement of Industry 4.0, Digital Twin has been applied to robots used in various industries and is being used in research such as navigation, task automation, and collision prediction. The robot arm used in the traditional manufacturing industry is optimized to perform repetitive tasks in a pre-determined and fixed environment. In most cases, the robot arm moves to a fixed path or approaches and handles the target object using a vision system that estimates the location based on augmented reality marker. To detect a target object outside a specific range, recently, vision-based object detection and various related algorithms are being developed. However, there is no proper high-fidelity Digital Twin platform which can simulate the integrated robot automation system including a perception function. In addition, various evaluation of different locations of a target and obstacle is limited, and thus it is not possible to flexibly respond to changes in the environment. These limitations result in the decrease of reliability and simulation-to-reality transferability. This study proposes a robot arm object detection system using photo-realistic Digital Twin and deep reinforcement learning algorithm. A high-fidelity Digital Twin was reproduced on NVIDIA Omniverse platform, a state-of-the-art physics engine-based simulator. 3D models of the actual robot arm, camera, target object, and obstacle were reconstructed in Digital Twin. Virtual target images were created with NVIDIA Scene Imaging Interface, and it is learned with Deep Object Pose Estimation algorithm. Camera-mounted robot arm was also controlled by using reliable robot control package, ROS Moveit!. The target and obstacle were randomly generated on the working region by using domain randomization for each episode. The learned policy using deep reinforcement learning in Digital Twin were seamlessly deployed and evaluated in the actual robot system through ROS-based framework. This integrated system could robustly detect the target object which has various positions and obstacle. In the future, we plan to extend the system developed in this study to a mobile robot arm to explore and handle objects in more diverse environments. We hope that this study can increase the reliability of robot automation system and decrease the time of programming and developing robots.

Keywords: Digital twin, Robot arm, Object detection, Deep reinforcement learning, Simulation-to-reality.

PDF Download