RT Dissertation/Thesis
T1 Action Generalization in Humanoid Robots Through Artificial Intelligence With Learning From Demonstration
A1 Fernández Fernández, Raúl
AB Action Generalization is the ability to adapt an action to different contextsand environments. In humans, this ability is taken for granted. Robotsare yet far from achieving the human level of Action Generalization. Currentrobotic frameworks are limited frameworks that are only able to workin the small range of contexts and environments for which they were programmed.One of the reasons why we do not have a robot in our house yetis because every house is different.In this thesis, two different approaches to improve the Action Generalizationcapabilities of robots are proposed. First, a study of differentmethods to improve the performance of the Continuous Goal-Directed Actionsframework within highly dynamic real world environments is presented.Continuous Goal-Directed Actions is a Learning from Demonstrationframework based on the idea of encoding actions as the effects theseactions produce on the environment. No robot kinematic information isrequired for the encoding of actions. This improves the generalization capabilitiesof robots by solving the correspondence problem. This problemis related to the execution of the same action with different kinematics.The second approach is the proposition of the Neural Policy Style Transferframework. The goal of this framework is to achieve Action Generalizationby providing the robot the ability to introduce Styles within roboticactions. This allows the robot to adapt one action to different contexts withthe introduction of different Styles. Neural Style Transfer was originally proposed as a way to perform Style Transfer between images. Neural PolicyStyle Transfer proposes the introduction of Neural Style Transfer withinrobotic actions.The structure of this document was designed with the goal of depictingthe continuous research work that this thesis has been. Every time a newapproach is proposed, the reasons why this was considered the best newstep based on the experimental results obtained are provided. Each approachcan be studied separately and, at the same time, they are presentedas part of the larger research project from which they are part. Solvingthe problem of Action Generalization is currently a too ambitious goal forany single research project. The goal of this thesis is to make finding thissolution one step closer.
YR 2021
FD 2021-06
LK https://hdl.handle.net/10016/33536
UL https://hdl.handle.net/10016/33536
LA eng
NO Mención Internacional en el título de doctor
DS e-Archivo
RD 27 jul. 2024