They use only one robot. Basically the drawback when implementing GA algorithms in real life (i.e. with real robots) is that the learning is not parallelizable but serial. In the specific work the single robot is loaded with the neural networks of each GA population's individual one after another. The results are kept and by software the next generation of neural networks is created. Then these are tested in the single robot etc.

Hmm that's what I thought. So for 1 generation to pass, we first try, say 80 generations in 1 robot, one after another. During this, we record each individual's fitness and then select the say, best 50%. Then recreate the neural networks for 80 new individuals and try it one by one on a single robot,. This goes like this right?

Thank you.

Yes that's right! Of course I would suggest to read a little about GAs (if you haven't done it already) and use things like elitism, roulette based selection, crossovers, mutations etc, instead of picking just the best 50%.

Cheers!

Yeah, I am in the process right now so I wanted to clarify these points.

Thank you so much.

Cheers!

The meaning of "population size=80" is, 80 number of individualas are considered in every generation. Crossover and mutation are performed on them. Then the best 50% can be selected as you mentioned.

The meaning of "population size=80" is, 80 number of individualas are considered in every generation. Crossover and mutation are performed on them. Then the best 50% can be selected as you mentioned.