Referência Completa


Título: A deep reinforcement learning method for humanoid kick motion
Autor : Luckeciano Carvalho Melo
Curso : Engenharia de Computação
Orientador : Adilson Marques da Cunha
Coorientador : Marcos Ricardo Omena de Albuquerque Maximo
Ano de Publicação : 2018
Assuntos : Aprendizagem (inteligência artificial)
t Dinâmica de robôs
t Redes neurais
t Robótica
t Computação
Resumo : Controlling high degrees of freedom for humanoid robot is acknowledged as one of the hardest problems in Robotics. Due to the lack of mathematical models, an approach frequently employed is to rely on human intuition to design keyframe movements by hand, usually aided by graphical tools. In this work, we firstly propose some methods based upon neural networks in order to imitate keyframe motions. Then, we propose a learning framework that not just imitates but also optimizes humanoid robot movement towards a task using Deep Reinforcement Learning. The developed technique does not make any assumption about the underlying implementation of the movement. The framework was applied in the RoboCup 3D Soccer Simulation domain and was able to improve the accuracy of given kick motion from 69% to 92% and also improved kick general distance.
Texto na íntegra : [Visualizar]