Applications to Robot Control

Genetic algorithms to robot control are briefly discussed in the following sections. These topics are dealt with in more depth in the following chapters on behavior-based systems and gait evolution.

 Genetic algorithms have been applied to the evolution of neural controllers for robot locomotion by numerous researchers. This approach uses the genetic algorithm to evolve the weight- ings between interconnected neurons to construct a controller that achieves the desired gait. Neuron inputs are taken from various sensors on the robot, and the outputs of certain neurons are directly connected to the robot’s actuators. successfully generated gaits for a hexapod robot using a simple traditional genetic algorithm with one-point crossover and mutate. A simple neural network controller was used to control the robot, and the fitness of the individuals generated was evaluated by human designers. evolved a controller for a simulated salamander using an enhanced genetic algorithm. The neural model employed was biologically based and very complex. However, the system developed was capable of operating without human fitness evaluators.

Genetic algorithms have been used in a variety of different ways to newl produce or optimize existing behavioral controllers. used a genetic algorithm to control the weightings and internal parameters of a simple reactive schema controller. In schema-based control, primitive motor and perceptual schemas do simple distributed processing of inputs (taken from sensors or other schemas) to produce outputs. Motor schemas asynchronously receive input from perceptual schemas to produce response outputs intended to drive an actuator. A schema arbitration controller produces output by summing contributions from independent schema units, each contributing to the final output signal sent to the actuators according to a weighting. These weightings are usually manually tuned to produce desired system behavior from the robot.

The approach taken by Ram et al. was to use a genetic algorithm to determine an optimal set of schema weightings for a given fitness function. By tuning the parameters of the fitness function, robots optimized for the qualities of safety, speed, and path efficiency were produced. The behavior of each of these robots was different from any of the others. This graphically demonstrates how behavioral outcomes may be easily altered by simple changes in a fitness function.

Example Evolution

 

Harvey used a genetic algorithm to evolve a robot neural net controller to perform the tasks of wandering and maximizing the enclosed polygonal area of a path within a closed space. The controller used sensors as its inputs and was directly coupled to the driving mechanism of the robot. A similar approach was taken in Venkitachalam 2002 but the outputs of the neural network were used to control schema weightings. The neural network produces dynamic schema weightings in response to input from percep- tual schemas.