Supporting material for the paper:

Evolution of Signalling in a Group of Robots Controlled by Dynamic Neural Networks

by Christos Ampatzis, Elio Tuci, Vito Trianni, and Marco Dorigo
July 2006

Real Robot Experiments:

Below are sample movies from the experiments performed in the arena above:

Experiments with 2 robots.
Movies are encoded in wmv format.

• 2 robots, Env. A Video (11 MB)
• 2 robots, Env. B Video (5.5 MB)

Experiments with 4 robots.
Movies are encoded in mpeg format.

• 4 robots, Env. A Video (11 MB)
• 4 robots, Env. B Video (14 MB)

All videos recordings from all the trials can be accessed at:

http://www.swarm-bots.org/integration-over-time.html

Experiments in simulation are performed with two robots and in reality with two and four robots. Movies are taken from an overhead camera and a hand camera and are encoded in .mpg or .wmv formats. Notice that when the robots emit a sound signal we light their colour turret red for visualisation purposes.

The behaviour of the group for the two robots is analysed in the paper. Concerning the four robots experiment, the results are almost perfect. In one trial though, Robot E while performing antiphototaxis as a reaction to the sound emitted by Robot C, made a turn of 180 degrees and started moving wrongly towards the light. In all other trials though we did not observe this error and it looks to be a hardware crash. Another error which was not expected and revealed some property of our controller about which we would not have found out had we not performed the four robot test, is the fact that the robot-robot avoidance behaviour does not work while the robots perform antiphototaxis. In fact, as they leave the band after they perceive a sound signal, their sensorial input is ignored, with the consequence that in case they encounter another agent on their way, they collide against each other. A possible explanation for this is that this condition was never encountered during evolution, and therefore the mechanism shaped was confined to just leaving the band without paying attention to obstacles (other robots). Finally, by allowing more agents to interact in the target area, we discovered that the robot-robot avoidance mechanism is different once the robots are interacting there---a case not often encountered during the two robot experiments--and can be described as follows: if an agent detects others in its vicinity, it stops and spins until the other agents have moved away.