Explainable Hierarchical Reinforcement Learning For A Cooperative Sorting Task
Authors
Madsen, Asger Printz ; Adamik, Mark
Term
4. semester
Education
Publication year
2021
Abstract
This thesis investigates how to make hierarchical reinforcement learning (HRL) for a collaborative robot understandable to non-expert users in a cooperative waste-sorting task. We conducted a user experiment in a simulated environment where participants sorted items while receiving different types of explanations from the robot. The results indicate that graphical explanations can lead to significantly better task performance compared to no explanations. Guided by these findings, we designed a three-layer HRL system for a UR-3 collaborative robot. We present and test the low-level policies, while the mid- and high-level layers are specified but not tested and left for future work.
Denne afhandling undersøger, hvordan hierarkisk forstærkningslæring (HRL) til en kollaborativ robot kan gøres forståelig for ikke-eksperter i en kooperativ affaldssorteringsopgave. Vi gennemførte et brugereksperiment i et simuleret miljø, hvor deltagerne sorterede genstande, mens de modtog forskellige typer forklaringer fra robotten. Resultaterne viser, at grafiske forklaringer kan forbedre opgavepræstationen signifikant sammenlignet med ingen forklaringer. Med udgangspunkt i disse fund designede vi et trelags HRL-system til en UR-3 kollaborativ robot. Vi præsenterer og tester de lavniveaustrategier, mens mellem- og høj-niveau lagene er specificeret, men ikke testet og henlagt til fremtidigt arbejde.
[This apstract has been generated with the help of AI directly from the project full text]