Supporting material for the paper:

Evolution of Autonomous Self-Assembly in Homogeneous Robots

by Christos Ampatzis, Elio Tuci, Vito Trianni, Anders Lyhne Christensen and Marco Dorigo
March 2008

Submitted to Artificial Life

Table of Contents
  1. Real Robot Experiments
  2. Comparison between simulated and real robots

Real Robot Experiments:



State diagram of events observed during the experimentation on real robots
State Diagram

This state diagram is the summary of all the possible scenarios observed in real s-bots during post-evaluation tests. States refer to particular s-bots' spatial configurations that are either representative of phases of a trial (e.g., RA-phase, Lost-phase), specific time in a trial (Start, End), or denote inaccurate manoeuvres that do not hinder the robots from eventually connecting to each other (i.e., I1, I2, and I3).The arrows indicate transitions between different states of a trial.

We remind the reader that: I1 refers to a series of maladroit actions by both robots which makes impossible for the s-bot-gripper to successfully grasp the s-bot-grippee's cylindrical turret. I2 refers to those circumstances in which both robots assume the role of s-bot-gripper and collide at the level of their grippers. I3 refers to those circumstances in which, after grasping, the connected structure gets slightly elevated at the connection point. Failures correspond to trials in which the robots do not manage to return to a distance from each other smaller than their visual field.

Below are sample movies from the experiments performed:

We also performed some sample experiments with 3 robots. Here is a movie sample of a successful trial: .wmv


Here you can access the page containing all the video recordings of the experiments performed and a brief description of their outcome.



Comparison between simulated and real robots

Time boxplot for simulated and real robots.
Time Boxplot

Boxplot showing the length of successful trials at the 1st, 2nd, and 3rd gripping attempt per post-evaluation tests G25, G30, B30, and R30. The boxplot labelled SimG25, SimG30, refer to the length of successful simulated trials in which the s-bots have been initialised as in the trials corresponding to boxplot labelled G25-1st, and G30-1st, B30-1st, R30-1st, respectively. Boxes represent the inter-quartile range of the data, while the horizontal bars inside the boxes mark the median values. The whiskers extends to the most extreme data points within 1.5 of the inter-quartile range from the box.

The boxplot shows the distribution of the length of trials with and without inaccuracies per post-evaluation test of real and simulated s-bots. Data concerning simulated s-bots (i.e., SimG25, SimG30) has been collected by placing the simulated agents in the same initial orientations as the real s-bots in the corresponding tests. That is, in SimG25 the simulated agents are placed as the real s-bots in G25; in SimG30 the simulated agents are placed as the real s-bots in G30, B30 and R30. By comparing the median of the distributions of simulated and real s-bots, we noticed that, for each pair of sets of trials, real s-bots tend to take longer time to accomplish their task than simulated agents. It seems that the real world noise, and other phenomena not modelled in our simulated world (e.g., friction), by affecting the dynamics of the evolved controllers, tends to increase the time required by the s-bots to physically assemble. The reader should bear in mind that, in spite of the difference between the length of trials of real and simulated s-bots, the robots' controllers proved to be robust enough to accomplish the task with high success rates in both environment. By looking at the figure, we also notice that the coloured light emitted by the LED does not seem to have any clear effect on the length of the successful trials with a single gripping attempt. As expected, trials that required more than one gripping attempt lasted longer than those in which the s-bots managed to assemble at the first attempt.