Natural Motion has been working in a similar area for a while. I think they rely less on genetic algorithms and more on explicit behavior and physical reactions though.
Brings to mind Greg Egan's short story "Crystal Nights", in which a tech billionaire works out that the best way to create a singularity-type AI is to evolve one rather than try to design one. He sets up a virtual world inside a (conveniently powerful) supercomputer, starts with very basic organisms, and then manipulates the environment to encourage evolution in the directions he wants. Things don't quite go as expected. Well worth reading.
My first thought was "this is really familiar"... then I realized it's because I used to work in Hod's lab :)
I think it's interesting that most of the robots that evolved don't have rigid components. I wonder if reducing the structural support of the "muscles" themselves would favor robots with a defined skeleton.
It's surprising how much I've learned about evolution just as a natural phenomenon since I started learning about computational evolution and genetic algorithms. It's a fascinating field with applications I find very exciting, and lots of research to be done.
The idea to use just four types of blocks is brilliant. I tried using only bone blocks, with invisible muscle fibers connecting them to each other and pulsating at random pace. Wasn't so easy to visualize.
>> And in the greater scheme of things, 1,000 generations is not a whole lot: in human terms, that's only 25,000 years, while modern Homo Sapiens have been around for ten times as long.
That's a bit misleading. 1,000 generations to evolve extremely basic movement is nowhere near comparable to the complexity of the human body.
Most of the complexity of the human body (including its locomotive facilities) evolved well before humans existed. I think the analogy is just to put the time scales in perspective, not to equate their products.
I don't want to act flamey here, but you should read [The Blind Watchmaker][1], R. Dawkins. The whole book is pretty interesting and analyzes exactly that complexity from an evolutionary perspective. There's also a chapter in which he describes his own "genetic algorithm" which produces life-like creatures. Pretty related to the topic in question.
I remember being dumbstruck after watching that video for the first time and realizing the power of programming. Karl Sims website still looks a bit like it did back then http://www.karlsims.com
The effect of the moving building blocks (expansion, contraction) is apparently not triggered synchronously. What determines what order these actions trigger in? It looks like this plays an important role in actually getting these to move...
It would be fascinating to give these creatures a very simple, evolvable nervous system: for example, a table where rows are timesteps, columns are cubes, and the cell contents are expand/contract/nothing: sort of like the old MOD trackers.
Viktor Zykov (one of Hod's grad students) did something like this in 2006 [1][2]. He built a "starfish" then had it learn its own shape through experimentation, then evolve a way to move.
I don't think it really shows that. By its nature, software programmed evolution is intelligence guided evolution. Even though it's left alone, the rules built into the software already contain the seeds for the possible outcome. PLUS, a big plus, the selection rules. This is intelligent evolutive design. It's not mere evolution through random mutations and natural selection.
Darwin's reason for using the phrase "natural selection" was to draw an important analogy with "artificial selection" as is done by pigeon-breeders.
It is clear that the intelligence of the breeder does not somehow directly reshape the pigeon. Rather, the way the breeder expresses his preferences is to SELECT the traits he is looking for. The only understanding required to see how this works in the case of natural selection is to see that nature analogously 'selects' some things over others, without intelligence on nature's part. In other words, the breeder is a "fitness landscape" for his captive pigeons. The natural fitness landscape may not choose the same things as a pigeon fancier would, but it does 'choose' some things like traits involved in surviving to breeding age, attracting mates, having sufficiently many healthy children, etc. Although it is natural, it is still selection.
This is not "intelligent evolutive design" just because a human being is involved in the experiment. Evolution is evolution and other members of the same species, or even members of other species, are part of the natural environment imposing selection pressure.
This brings up a metaphysical point of whether humans really play a natural part in the evolutionary process anymore. For example, when mankind wipes out another species, in some sense you can say this is "natural", since evolution made us in the first place, and we're just doing what we do. In another sense, you can say we're assholes.
Personally, I do see a line being drawn, starting a few thousands year ago and coming to a head now with bio-engineering and robotics, etc., where human-level intelligence represents the next major stage in how matter is organized over time. First was the stellar life-cycle, then biological evolution, and now we're joining the party. If you view it this way, I don't believe we can say that we're part of the natural environment anymore.
Sorry, didn't mean to get all abstract...it's all physics anyway.
It's not bad as a model of evolution. In real world evolution the rules are defined (physics, chemistry) that contain the seeds for the possible outcome. The selection is merely: who survives to procreate. These sorts of models have to establish some arbitrary criteria that permits culling, but that's no different than someone in CFD modeling thousands or tens of thousands of atoms and molecules as a single entity and then continuing the model with thousands or millions of those entities. It's still a valid, demonstrative model.
"the rules built into the software already contain the seeds for the possible outcome" -- you could say the same for natural evolution, if you consider the laws of physics as 'the rules built into the software'.
Regarding selection, yes, this is artificial selection pretty much by definition, since it's simulated. Though it's an abstracted, simplified version of whatever natural selection is, it's not really very different in the end.
The problem of course it's so laid out to obtain the desired results. Artificial selection. But it's indeed shown that mutations, completely random, can breed evolution. The end results are unforeseeable solutions that work great.
Right, bad choice of words. This doesn't show the specifics of natural selection.
I believe the fact they didn't designed the final solution, but instead the rules and mutations, successfully shows mechanisms analogous to the ones involved in natural selection though, even if this is a human-made simulation.
