Self-Specialization of General Robot Plans Based on Experience

Autor: Michael Beetz, Gayane Kazhoyan, Sebastian Koralewski
Rok vydání: 2019
Předmět:
Zdroj: IEEE Robotics and Automation Letters. 4:3766-3773
ISSN: 2377-3774
DOI: 10.1109/lra.2019.2928771
Popis: For robots to work outside of laboratory settings, their plans should be applicable to a variety of environments, objects, task contexts, and hardware platforms. This requires general-purpose methods that are, at this moment, not sufficiently performant for real-world applications. We propose an approach to specialize such general plans through running them for specific tasks and thereby learning appropriate specializations from experience. We present a system architecture, which collects data during plan execution for making up supervised learning problems and utilizes learned models for specializing the plans in a closed loop. We demonstrate our approach by letting a PR2 robot specialize its general fetch and place plan, whereby learned results are automatically installed into the plan. We show that the specialized plan performs better than the original plan in a statistically significant sense.
Databáze: OpenAIRE