Generating appropriate object orientations for robot-to-human handovers using synthetic object affordances
Term
4. semester
Education
Publication year
2022
Submitted on
2022-06-01
Pages
90
Abstract
This project is an investigation into applying object affordances to robot-to-human handovers. Our research makes two contributions. A state-of-the-art deep neural network for segmentation of object affordances named AffNet-DR, trained solely on synthetic data. Secondly, an object affordance enabled method for orienting objects appropriately for robot-to-human handovers. A user study with 6 participants showed that our method for computing handover orientations outperforms a method that uses random orientations. Finally, a robotic handover system was programmed in ROS Melodic and implemented on a KUKA LBR iiwa 7 R800 with an Intel RealSense D435i RGB-D sensor and a Robotiq 3-finger gripper. The system performs robot-to-human handovers with a success rate of 91.67 %.
This project is an investigation into applying object affordances to robot-to-human handovers. Our research makes two contributions. A state-of-the-art deep neural network for segmentation of object affordances named AffNet-DR, trained solely on synthetic data. Secondly, an object affordance enabled method for orienting objects appropriately for robot-to-human handovers. A user study with 6 participants showed that our method for computing handover orientations outperforms a method that uses random orientations. Finally, a robotic handover system was programmed in ROS Melodic and implemented on a KUKA LBR iiwa 7 R800 with an Intel RealSense D435i RGB-D sensor and a Robotiq 3-finger gripper. The system performs robot-to-human handovers with a success rate of 91.67 %.
Keywords
Documents