Juan Cristobal Zagal's Publications

Classified by Publication TypeClassified by Research Category

Embodied Robot Simulation

Juan Cristobal Zagal. Embodied Robot Simulation. Ph.D. Thesis, Universidad de Chile, 2007.

Download

(unavailable)

Abstract

Arising from an embodied cognitive perspective the approach of evolutionary robotics (ER) allows generating autonomous robot behaviors (or morphologies) as a consequence of the interaction between robot and environment. However, in order to generate a complex robot an extremely large amount of time is required, making the methodology generally unpractical. With great deception ER researchers have tried to use simulation in order to speed up robot adaptation. According to the "reality gap problem" robots generated under simulation fail to be transfered to reality, producing unexpected behaviors. Interestingly, recent findings from biological sciences show evidence for the cognitive and functional role that different types of simulation play in living systems. In contrast to a propositive representational simulation, this thesis proposes the use of an emergent embodied robot simulation; which is presented as a term for the use of environmentally coupled realistic robotics simulators that can be automatically generated during the robot behavioral adaptation process. In particular this thesis presents (i) an algorithm for generating embodied robot simulations, and (ii) a dynamics robotic simulator based on state of the art technology. A key advantage is that designer bias is reduced by coupling simulation with behavior. The presented approach also serves as an alternative for crossing the reality gap in evolutionary robotics. The proposed algorithm is experimentally validated by showing the successful development and continuous transference to reality of complex low-level behaviors with Sony AIBO robots. In the case of the quite standardized problem of gait optimization the algorithm shows the highest learning performance as compared to the other machine learning methods which have been applied to the same problem (Genetic Algorithms, Policy Gradient Reinforcement Learning, Evolutionary Hill Climbing With Line Search, Powell Direction Set). The methodology allows to successfully solve a problem that due to its complexity has not been addressed before: the automatic generation of unconstrained ball kick behaviors. Cognitive robotics, evolutionary robotics, embodied cognition, embodied simulation, co-evolution.

BibTeX
@PhdThesis{zagal2007b,
  author = 	 {Juan Cristobal Zagal},
  title = 	 {Embodied Robot Simulation},
  school = 	 {Universidad de Chile},
  year = 	 {2007},
  OPTkey = 	 {},
  OPTtype = 	 {},
  OPTaddress = 	 {},
  OPTmonth = 	 {October},
  OPTnote = 	 {},
  OPTannote = 	 {},
  abstract = { Arising from an embodied cognitive perspective the
                  approach of evolutionary robotics (ER) allows
                  generating autonomous robot behaviors (or
                  morphologies) as a consequence of the interaction
                  between robot and environment.  However, in order to
                  generate a complex robot an extremely large amount
                  of time is required, making the methodology
                  generally unpractical.  With great deception ER
                  researchers have tried to use simulation in order to
                  speed up robot adaptation. According to the
                  "reality gap problem" robots generated under
                  simulation fail to be transfered to reality,
                  producing unexpected behaviors. Interestingly,
                  recent findings from biological sciences show
                  evidence for the cognitive and functional role that
                  different types of simulation play in living
                  systems.  In contrast to a propositive
                  representational simulation, this thesis proposes
                  the use of an emergent embodied robot simulation;
                  which is presented as a term for the use of
                  environmentally coupled realistic robotics
                  simulators that can be automatically generated
                  during the robot behavioral adaptation process. In
                  particular this thesis presents (i) an algorithm for
                  generating embodied robot simulations, and (ii) a
                  dynamics robotic simulator based on state of the art
                  technology. A key advantage is that designer bias is
                  reduced by coupling simulation with behavior. The
                  presented approach also serves as an alternative for
                  crossing the reality gap in evolutionary robotics.
                  The proposed algorithm is experimentally validated
                  by showing the successful development and continuous
                  transference to reality of complex low-level
                  behaviors with Sony AIBO robots. In the case of the
                  quite standardized problem of gait optimization the
                  algorithm shows the highest learning performance as
                  compared to the other machine learning methods which
                  have been applied to the same problem (Genetic
                  Algorithms, Policy Gradient Reinforcement Learning,
                  Evolutionary Hill Climbing With Line Search, Powell
                  Direction Set). The methodology allows to
                  successfully solve a problem that due to its
                  complexity has not been addressed before: the
                  automatic generation of unconstrained ball kick
                  behaviors.  Cognitive robotics,  evolutionary robotics, 
                  embodied cognition, embodied simulation, co-evolution.},
}

Generated by bib2html.pl (written by Patrick Riley ) on Wed Sep 09, 2009 19:24:22