For decades, robots have been able to reliably follow precise trajectories making them ideal tools for assembly lines and other structured environments. However, pre-programmed motions fail under uncertainty and are unsafe around humans, making them inadequate for unstructured environments. This paper presents a framework to generate safe, robust, and generalizable robot behaviors for contact tasks where compliance plays a key role. First, we collect task data from haptic demonstrations. Then, we segment the data into a sequence of compliant primitives. Finally, we extract the key parameters required for a robot to perform each of the primitive actions using interpretable, model-based controllers.
This method was experimentally validated on a steel bolting task using a 7-DOF Franka Panda robot. By recombining the primitives, we were also able to screw a cap onto containers of different sizes, placed in arbitrary configurations, using two different 7-DOF manipulators. The results show that our method generates position and orientation invariant, robot-agnostic controllers.
Parametrization of Compliant, Object-Level Controllers from Human Demonstrations
Springer Proceedings in Advanced Robotics
International Symposium on Advances in Robot Kinematics ; 2022 ; Bilbao, Spain June 26, 2022 - June 30, 2022
2022-06-18
13 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
British Library Online Contents | 2017
|ROBOT LEARNING OF OBJECT MANIPULATION TASK ACTIONS FROM HUMAN DEMONSTRATIONS
BASE | 2017
|Learning continuous human-robot interactions from human-human demonstrations
TIBKAT | 2018
|