Modeling and the Reality Gap

The properties of bricks are variable. Differences in construction, age, dirt, temperature, humidity and other unpredictable factors produce seemingly random variations when their behavior is measured. These factors had to be considered in order to have buildable results. We have accounted for this problem using a safety margin: our model assigns 20% less resistance to all the joints involved.

The simplistic simulator described is far from modeling physics to its full detail, yet any model for modular structures, no matter how accurate, has to compensate for the random variability in the generic properties of the average brick, using similar methods. The value of 20% was set intuitively and may require further study, especially as our structures scale up in size and complexity.

Engineers have been using safety margins for a long time, to deal with all the factors that cannot be calculated ahead of time. In ALife, evolutionary roboticists have found unpredictabilities when attempting to simulate the environment for a real robot [72,98]. One simple technique used in ER is to add random noise to the simulated sensors in order to generate robust behaviors suitable to be transferred to the real world.

Noise and safety margins are variations on a fundamental principle of conservativeness. The so-called ``reality gap'', which is the difference between the model and the final artifact, does not mean that modeling is useless. But it must be taken into account in order to achieve successful transfer from simulation to reality. ALife research has addressed the reality gap problem in different ways, such as:

• Using safety margins, as exemplified by the present work.
• Evolving directly in reality, as in evolution in FPGA's [133] or robot behaviors [37].
• Evolving adaptable entities, as proposed by Di Paolo [32].
• Use a hybrid model, first evolving in simulation, then transfer to reality to do a final round of adaptation, has been proposed by our group [110] and implemented (for a game) in the two-layer evolutionary architecture of Tron (chapter 3).

We agree with Jakobi [70,71] in going for simulation at the right level, aiming at the aspects of reality that insure robust, transferable behaviors. We also think that the method of simulated evolution has inherent advantages in lieu of the reality gap problem: the dynamics of reproduction and selection subject to mutation lead evolving populations to explore fuzzy regions -- clouds within the solution space.

Those organisms that, when mutated, are likely to be successful, have a reproductive advantage: their offspring is more likely to survive. This means that inasmuch as genotypical and phenotypical variation are linked, evolved solutions will have a tendency to being robust, in the sense that small variations in their constitution will not incapacitate them (more likely, they give rise to successful mutants). If this phenomenon is true, then solutions obtained by evolution implement ``by default'' a conservative principle based on kinship. This is an open research question.