Generating appropriate object orientations for robot-to-human handovers using synthetic object affordances
Authors
Lehotský, Daniel ; Christensen, Albert Daugbjerg
Term
4. semester
Education
Publication year
2022
Submitted on
2022-06-01
Pages
90
Abstract
Passing an object from a robot to a person should feel straightforward. This project explores how object affordances—the parts of an object that suggest how to act on it, like a mug handle—can enable better handovers. We contribute two components: AffNet-DR, a deep neural network that locates and labels affordance regions on objects (affordance segmentation), trained solely on synthetic (computer-generated) data; and a method that uses these affordances to choose an object orientation so a person can easily take it from a robot. A user study with six participants showed that our orientation method outperforms choosing a random orientation. We also built a complete system in ROS Melodic on a KUKA LBR iiwa 7 R800 with an Intel RealSense D435i RGB-D camera and a Robotiq 3-finger gripper. The system achieved a 91.67% success rate in robot-to-human handovers.
At give en genstand fra en robot til et menneske bør føles enkelt. Dette projekt undersøger, hvordan objekt-affordanser – dele af en genstand, der indbyder til bestemte handlinger, fx et krushåndtag – kan bruges til bedre overleveringer. Vi bidrager med to elementer: AffNet-DR, et dybt neuralt netværk, der finder og markerer affordance-områder på objekter (affordance-segmentering), udelukkende trænet på syntetiske (computergenererede) data. Derudover en metode, der bruger disse affordancer til at vælge, hvordan et objekt skal orienteres, så en person nemt kan tage imod det fra en robot. Et brugerstudie med seks deltagere viste, at vores metode til at beregne overleveringsorienteringer er bedre end at vælge tilfældige orienteringer. Vi byggede også et komplet system i ROS Melodic på en KUKA LBR iiwa 7 R800 med et Intel RealSense D435i RGB-D kamera og en Robotiq 3-finger griber. Systemet gennemførte robot-til-menneske-overleveringer med en succesrate på 91,67 %.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Keywords