Neuroevolution can be used to evolve neural networks that can control
robot manipulators to perform tasks like target tracking and obstacle
avoidance in complex environments. Neurocontrollers have been
successful in the robot control domain because they are robust to
noise, they can be adapted to different environments and manipulator
configurations, and they can be used to implement controllers that can
perform online learning.
The focus of this report was to evolve neurocontrollers for two
environments. First, neurocontrollers are evolved for environments
without obstacles and their performance is compared to an inverse
kinematic controller and a potential field controller. Second,
neurocontrollers are evolved for environments with obstacles and
compared with a controller that uses potential fields to implement a
path planning algorithm. The neurocontrollers evolved in this report
come close to matching the performance of the analytical
controllers. The advantage of using neurocontrollers is their
robustness to noise and ability to adapt to different environments.
Masters Thesis, Department of Electrical and Computer Engineering, The University of Texas at Austin.