UWE Bristol Engineering showcase 2015
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
Jonathan Barnett<br />
BEng Robotics<br />
Project Supervisor<br />
Prof. Alan Winfield<br />
Analysing RoboGen for Evolutionary Design<br />
Project summary<br />
The aim of this project is the investigation of<br />
the RoboGen software for the<br />
implementation of evolutionary robotics.<br />
The primary appeal of the RoboGen software is its<br />
ability to co-evolve the robot body and brain<br />
simultaneously, creating a more harmonious<br />
design between the brain and body.<br />
One of the key aspects of the RoboGen<br />
software, is the ability to fully configure the<br />
setup of the robot build. This data can then be<br />
outputted to Matlab or other suitable software<br />
for rendering the information allowing for a<br />
generation by generation recount of the fitness<br />
development.<br />
The robot brain aspect evolves weights and biases<br />
of the robot within the body structure derived<br />
from the need to achieve the goal state, the design<br />
of the robot brain system is described as a “fullyconnected,<br />
recurrent artificial neural<br />
network”(Auerbach et al., 2014). The limitations of<br />
this robot brain are primarily based on the body<br />
structure that it is contained within.<br />
From Concept to Reality<br />
The RoboGen software allows for the 3D<br />
design of a robotic system using modular<br />
parts pre-configured to be wired up. The<br />
system then develops the artificial neural<br />
network within this body in order to solve<br />
the training scenario it has been placed in<br />
Project Objectives<br />
• Investigating the potential of the RoboGen<br />
software for evolutionary neural network<br />
development and co-evolutionary design<br />
of both brain and body, in a variety of<br />
situations examining the results that the<br />
system outputs<br />
• Investigation into the principle idea of real<br />
world testing of fitness, being reintroduced<br />
into the software population,<br />
adjusting its fitness.<br />
Project Conclusion<br />
In conclusion the RoboGen software is a remarkable<br />
application allowing for ease of use and accessible<br />
introduction to the principles behind both<br />
evolutionary neural network design and co-evolution<br />
development between the brain and body of a robot.<br />
The ability to configure a multitude of parameters as<br />
well as different scenarios for testing means that the<br />
software can be used for potentially multiple<br />
problems depending on how complex the user makes<br />
the terrain that the robot must navigate in order to<br />
achieve its goals. The co-evolutionary nature<br />
between both the brain and body of the robot, when<br />
used, opens up a potential gateway towards true<br />
independent design.