Publication Details

SELECT * FROM publications WHERE Record_Number=11338
Reference TypeThesis
Author(s)Stadtmueller, J.
Year2020
TitleDimensionality Reduction of Movement Primitives in Parameter Space
Journal/Conference/Book TitleBachelor Thesis
AbstractAlthough there have been promising advancements in recent years, Reinforcement Learn- ing is not yet directly applicable to robotics, since the dimensionality of robot movements is high, so the computation time of the RL algorithms becomes prohibitive. Having a learning system on real robotics will enable industrial robotics to be easily reprogrammable to new tasks, and at the same time avoid the need for training in simulated environments. Engineering expenses to design the simulation can be saved and the reality gap can be overcome. We propose a new framework that works in parameter space to resolve the dimensionality of the representation. Our approach takes into consideration the similarity between movements instead of the redundancy in the kinematic system. Additionally, the proposed method is independent of the number of chosen features. We empirically show that the framework we introduce works efficiently for both complex human movement as well as in a simpler robotic scenario.
Link to PDFhttps://www.ias.informatik.tu-darmstadt.de/uploads/Team/SamueleTosatto/stadtmueller2020.pdf

  

zum Seitenanfang