Skip to contentSkip to site navigation
Cognitive Science
Completed Project

Evolutionary Robotics: Tools and Methods to Test Hypotheses about the Evolution of Evolvability

Evan Altiero, Vassar College ’16, Jessica Ng, Vassar College ’16, Joshua Ridley, Vassar College ’17, Nicholas Burka, Vassar College ’17, David Wallach, Vassar College ’17, Meghan Willcoxon, Vassar College ’18 and Profs. Ken Livingston, Nick Livingston, John Long, Jodi Schwarz and Marc Smith Collaborators: Joshua Bongard and Anton Bernatskiy, University of Vermont

We are using evolutionary robotics to test hypotheses about the evolution of evolvability in a population of behaviorally autonomous robots. We developed new tools and methods to measure various features of the physical environment and the robots’ behaviors and neural networks.

1. Understanding the world: Our robots (Tadros) swim in a 3-meter tank with a single overhead light source that creates an energy gradient that the Tadro detects via two photocells that function as left and right eyes. Output of these sensors can then be used to alter activity of the robot’s tail and change the direction of swimming. The amount of light collected from a third, center photocell provides the fitness function that determines the reproductive rate for a given robot. Robots that swim toward the light will gather more light and be more fit. We developed light intensity maps to characterize the gradient as measured by the three sensors.

2. Understanding the link between perception and behavior: Video recordings of each robot swimming were mapped to a grid frame, and computer imaging software was used to remove camera distortion. Image processing software was modified to allow mapping of position to the light intensity map of the environment. Additional software allowed digital overlay of sensor data onto the original videos.

3. Understanding each robot’s artificial neural network (ANN): A simulator of the code running on the Tadro’s microcontroller was used to test the effect of actual perceived inputs from navigational sensors on output commands to motors. The simulator allows selective removal of network connections to explore effects of such changes on tail behavior. To assess the degree of structural modularity in a given ANN, the Q value of the network was calculated. Q ranges from zero (low modularity) to one (high modularity). Calculated Q values for each agent in each generation were used to create a genealogical map of Q scores.