Julio Jerez wrote:I finally saw the videos.
For thoss tah want to see you can download this codex
http://www.free-codecs.com/FFDShow_download.htmMan the run cycle is absolutely awesome; you are telling me that the genetic algorithm came out with that solution all by itself.
The run cycle really looks organic and that it came out from an evolution process. I guess that having a clear feeback really improve the evolution of the genne and reduces the mutations that lead to such moronics solution like I had seen.
Absolutly, using the good inputs (angle insteads of force) is less noisy, so it allows a better exploration of the solutions space. The other important point is the design of the controler : made from CPGs (Central Pattern Generator, see Auke Jan Ijspeert papers). Basicly, it's a network of non-linear oscillators. Those oscillators are a simplified model of a neuron group found in the spinal cord of the lamprey.
Julio Jerez wrote:I have seen many creatures walking with thist algorithm with other engines, but I never saw one with such symmetry and coordination.
Congratulation, I am looking for to see more.
Impossible without a good physic engine...
Julio Jerez wrote:I most say that the walk cycle is not as impresive as the run.
one question after how many iteration did you get that result?
Here it is the parameters of the evolution :
* Genetic algoritm : CMAES
* Population : 15 individuals
* Mu+Lambda ratio : 2
* Fitness measure : distance performed by the robot
Walk cycle :
* 30 seconds of evaluation per individual
* 5 generations
* 0h15 to evolve
Well, 5 generations, it's mostly random search
Running cycle :
* 90 seconds of evaluation per individual
* 30 generations
* 1h30 to evolve
Very soon : a snake and an articulated 24 legs bug...