Evolving Swimming Soft-Bodied Creatures

Подписаться 856

Просмотров 236 тыс.

50% 1

Accompanying video for:
F. Corucci, N. Cheney, F. Giorgio-Serchi, J. Bongard, C. Laschi "Evolving soft locomotion in aquatic and terrestrial environments: effects of material properties and environmental transitions" arXiv:1711.06605 [cs.AI] arxiv.org/abs/1711.06605
Currently in press for the Soft Robotics Journal!
F. Corucci, N. Cheney, H. Lipson, C. Laschi and J. Bongard, "Evolving swimming soft-bodied creatures"
The Fifteenth International Conference on the Synthesis and Simulation of Living Systems (ALIFE XV) - Late Breaking Abstract
Paper: sssa.bioroboticsinstitute.it/s...
See also our growing soft robots exploiting morphological computation: • Artificial Evolution o...
More on:
sssa.bioroboticsinstitute.it/u...
www.meclab.org/

Наука

Опубликовано:

29 июн 2016

Ссылка:

Скачать:

Готовим ссылку...

Добавить в:

Мой плейлист

Посмотреть позже

Комментарии : 319

@deathpony698 7 лет назад

I want to show someone this but with out any subtitles or context.

@willwhite1987 7 лет назад

This must happen. :D

@Skwerll 7 лет назад

Once I read this comment, every new creature made me laugh.

@emeraldpopcorn 7 лет назад

Especially 3:57 - 4:15

@mrpenguin2083 6 лет назад

MAKE THIS

@ilyano1511 6 лет назад

also please with the same music

@lukehill6395 7 лет назад

congrats, you made a jellyfish simulator

@ensjo 7 лет назад

And yet NOT programming it specifically to behave like a jellyfish. That's the point. Evolved body form and behavior. :-)

@isaiahphillip4112 7 лет назад

You say it like it's not impressive!

@EighteenYearAccount 7 лет назад

Luke Hill interestingly enough jellyfish are some of the oldest and most basic organisms on earth

@rodrigoalvarado9524 6 лет назад

he kind of proved jellyfish are efficient organisms

@avengineering-yt 7 лет назад

I haven't got a single clue what's going on but the swimming cubes look pretty cute.

@areallylongnamethatyourest6509 7 лет назад

Basically, it's a computer program which creates creatures that can swim good.

@avengineering-yt 7 лет назад

A really long name that you're still reading for some reason. Why are you still reading my name? Well, I got that, at least. Heh.

@ensjo 7 лет назад

Hey, Inquiry. The interesting part is that those walking/swimming behaviors are not programmed. What is programmed is just how each little cube functions and interacts with surrounding cubes and with the environment (ground, surrounding "water" or "air"). The walking/swimming behaviors themselves emerge by an evolutionionary process that imitates what happens in nature. Initially the computer generates a bunch of random body structures that just wobble around crazily. But those which eventually move faster in a single direction are given the chance to "give birth" to a new generation. Those which perform poorer are eliiminated. The individuals of the new generation bear some mutations (random slight modifications in body form and behavior) that in some cases make them yet better than their parents. And through this process of reproduction, mutation and selection (survival of the fittest), efficient walking/swimming behaviors appear.

@manospondylus4896 7 лет назад

1:45 it's a Swiss jellyfish!!!

@henryambrose8607 7 лет назад

Disappointed Turtle Here we see a rare creature: the Swiss navy cube, in its natural habitat. Evolution has refined for many generations this most extraordinary creature, creating a box built purely for efficiency.

@peacedustinc.7108 7 лет назад

Disappointed Turtle a jellyswiss

@elanathevampireprincess1160 6 лет назад

Peacedust inc. Hahaha wow did you just go there?😂

@gimli3370 7 лет назад

It's quite interesting to see just how much more similar these creatures are to real aquatic creatures than their terrestrial counterparts, showing how it must be easier to evolve in the oceans than on land as the aquatic simulations seemed to come to a more realistic model in what I assume is about the same time. Also it is quite odd how the symmetry became much more prevalent in the aquatic ones which I wonder if that might have something to do with how in an aquatic environment the creature ends up exerting force all around it rather than on one surface

@Maurcusj777 7 лет назад

i think it's more that without rigid cells that can act as bones, fotming functioning limbs is really difficult.

@theholybuttfungus5911 5 лет назад

Yeah it is pretty fascinating. However that’s probably because the simulations only need to accomplish locomotion and not be interferes with natural selection etc.

@directrix1 8 лет назад

I feel like I'm watching a late 90s porn, but there is a little less flopping around in this.

@kai_unix 7 лет назад

you made this video alot uncomfortable for me

@miahlfranklin3794 7 лет назад

I knew if I looked, someone would've said this.

@user-ft8hh3pe6r 5 лет назад

The music certainly sounds like 90s porn.

@electrotoxins 5 лет назад

4:09

@clochard4074 7 лет назад

Fascinating. Have you ever considered the chance to add solid units and see if the robots will mimic bones? Would it be pertinent with your project?

@fcorucci 7 лет назад

Hi Francesco. Yes, it is already possible to allow stiffer voxels, as well as to evolve the stiffness distribution across the whole body as well. We would like to observe the emergence of skeletal structures, but apparently such structures are not easy to find. I guess it depends from the optimization algorithm at hand, and from the task as well. Those solutions are definitely within the search space generated by our generative encoding.

@bacicinvatteneaca 7 лет назад

Francesco Corucci You should do something about attrition in water, it seems it likes a lot to push flat square surfaces against the water. Maybe if it needed to be a bit more pointy it would result in fish-like robots. Also, is there a way to allow the robots to create new block types? With new rules? Because mutation also invents things :)

@bacicinvatteneaca 7 лет назад

@lxjuani 7 лет назад

I thought the same thing while watching this. Maybe make "offspring" of a certain creature with minor random alterations, re-train them and make them compete, then make the fastest have mutated offspring and so on. That would be very interesting.

@faeluvzelda5091 5 лет назад

Francesco Corucci i think a solution could be to make there be less options for stiffness similar to bone flesh and cartilage

@codiserville593 7 лет назад

That first water to land creature made itself into a rabbit

@Doubled952 7 лет назад

Codi Serville a lot of the creatures look pretty similar to real ones

@henryambrose8607 7 лет назад

Doubled952 Unsurprisingly. It makes sense that creatures in a similar environment would evolve similarly.

@elliotlea5457 7 лет назад

I think that this isn't actaully evolution

@henryambrose8607 7 лет назад

Elliot Lea It is a simulation of evolution.

@ensjo 7 лет назад

It's the evolutionary mechanisms of reproduction, mutation and selection (survival of the fittest), applied not to chemical cell-based organisms like we see in the real world (which would be terrible to simulate), but in virtual creatures made up of little cubes with simple functioning. And yet the evolurionary processes can make the random, pathetically wobbling bodies created for the fist generation mutate through generations into elegant swimmers. It is evolution, the same mechanisms, just in "another world".

@EmergencyTemporalShift 7 лет назад

I would say the reason for symmetry is that otherwise the lot of them would travel in circles.

@fcorucci 7 лет назад

The pressure to go straight is certainly one important factor. It is however true that this degree of morphological regularity did not emerge in similar works, where the fitness was still based on traveled distance.

@EmergencyTemporalShift 7 лет назад

Francesco Corucci So, whats the difference with this experiment?

@fcorucci 7 лет назад

We still need to look into this

@enchilordo8562 5 лет назад

“Here are some more evolved creatures.” *squishy cube*

@Oberon123 7 лет назад

came for the science, stayed for the music great video :)

@JacobEllinger 7 лет назад

add a water environment with a shore but put the reward at the shore but a denser reward farther inland so you can better see the transition from water to land where the bots that becomes fully land capable gets more rewards.

@fcorucci 7 лет назад

This is an excellent suggestion. We considered doing that, it just takes slightly more time to set up this kind of environment, so for the time being we tried something simpler, with the transition occurring over evolutionary time instead.

@fcorucci 7 лет назад

We will definitely try the setup you are suggesting in the near future, though!

@MarteonLP 7 лет назад

Jacob Ellinger

@JamUsagi 7 лет назад

The land creatures are called shuffledogs.

@daggius 8 лет назад

this jam is dope af

@noyz-anything 7 лет назад

Can I have one?

@MaximilianonMars 7 лет назад

I'm glad there's music to this. I've seen many that are in total silence, makes for boring viewing sometimes.

@Tlactl 7 лет назад

neat how the evolution of the robots comes out with the same results (jellyfish) as evolution in real life

@megalunalexi5601 7 лет назад

n e a t there were some with fins too!!

@lukabb7931 6 лет назад

But jellyfish don't actually move like this, right?

@arc3028 6 лет назад

They ALMOST move exactly like that, give them more time and they will figure it out

@pestianK 7 лет назад

0:42 I love how the thing moves with the music :D

@firstlast-wg2on 7 лет назад

You can realistically create art that you think is most beautiful by analyzing what one person finds the most attractive, traits they like (it would have to be a very extensive list of things) and simply generate random pieces of art, the art that the person likes the most or matches the criteria of what they like will "reproduce" until you have a very refined piece of art that somebody will likely love.

@fcorucci 7 лет назад

Yes, this is something very interesting, and a few people have indeed worked in this direction in the past. The intriguing thing is that if you have people in the loop, you actually don't need to encode the characteristics that will result in higher appreciacion, nor users need to be aware of those characteristics. The global degree of appreciacion a user expresses is enough to guide the search. Check this out: www.karlsims.com/genetic-images.html www.picbreeder.org/ endlessforms.com/ The last two use the very same encoding (CPPN) we are using here to evolve soft robots (picbreeder was actually developed by the inventor of such encoding, Ken Stanley, and his group). Different story if you try to produce interesting content with no humans in the loop. There has been work in this direction too: something that has been done is for example to maximize metrics based on information theory to produce interesting patterns (2D, 3D, or even animations). Intuitively, these metrics quantify the information content of a picture, that is somehow related to the notion of compressibility. A plain background is compressible (even a big image can be achieved by just knowing the color of a pixel and how many times it must be replicated). A complicated pattern, on the other hand, is less compressible (you can't achieve the whole image knowing just a single "pixel" of information. There are information theoretic metrics that capture this notion of information content, that can be used as fitness functions :-)

@firstlast-wg2on 7 лет назад

Francesco Corucci Oh shit, very nice and insightful, thanks.

@IJustLoveStories 7 лет назад

Interesting to see, really firing up my sensationalist excitement. What if you simulated multiple creatures together, give them a goal to reach to make them compete, and allow voxels to be destroyed by means such as excessive force (which could produce different results in different fluid densities as well), and allow them to evolve harder, yet less flexible voxels. Would they sacrifice flexibility and efficiency for durability? If so, would a sweetspot between flexibility and durability emerge? how would that sweetspot relate to the fluid density? Would they evolve to damage each other or to out-speed each other in order to get to the goal? if so, which is more likely to evolve (ratio of creatures that went either route) and which is more likely to succeed (ratio of success and defeat for each strategy). I don't know if these things can, need or should to be tested, it'd just be really interesting and cool.

@Schenkel101 7 лет назад

Although that would probably be harder to simulate, it does sound really interesting.

@Firebolt68 5 лет назад

Way to overuse the thesauruses.

@icecreamsamwich 7 лет назад

so many of these robots evolved like real-world animals cool!

@arantisdavis1623 7 лет назад

it's awesome to see the water/land evolutions all formed what could be considered, vertebrae.

@RB-fo5id 7 лет назад

Strangely beautiful

@KohuGaly 7 лет назад

Interesting. How about digging behaviour? The creature would have to push away material in front of it of it and also move forward. It might be a bit more CPU intensive to calculate all the collisions though...

@fcorucci 7 лет назад

KohuGaly thanks for your comment. Increasing behavioral complexity and diversity is certainly on the agenda, although at the moment we are focusing on different aspects. The setup you've described reminds me Josh Bongard's block pushers, you may want to check them out. And yes, you're right about collisions. If you're interested in the behavioral complexity/diversity, you might also enjoy Dan Lessin's work on this topic, you should be able to find a bunch of very cool videos related to his work here on RU-vid.

@KohuGaly 7 лет назад

Francesco Corucci I've seen Lessin's videos before. They are really interesting. There is another think I'd like to ask. In another video you seem to be able to affect the "squishiness" of the creatures. Have you tried including the squishiness in the genetics (aka the voxels can have different rigidity). It might be interesting to see skeletal structures emerge.

@fcorucci 7 лет назад

KohuGaly yes, the stiffness distribution can be evolved as well. It's not very common to observe the emergence of clear skeletal structures in this setup, but from time to time something interesting pops out in this regard. We are definitely trying to achieve this kind of result, hopefully we'll have some new results soon.

@blykgod 8 лет назад

Amazing results :)

@fcorucci 8 лет назад

Thank you! :-)

@Kram1032 7 лет назад

These are awesome! I know it's probably insanely time-intensive to do but I have now seen, I think, six experiments with these: - walking -- with varied gravity -- getting out of a box -- using different rhythm structures - swimming -- changing between walking and swimming - growing I think all of those could benefit in interesting ways if they weren't necessarily topologically restricted to fit into a cube and/or the resolution would be quite a bit higher. I know the arbitrary topology thing has partially been done by Michał Joachimczak at least in 2D and then there are the bone-and-muscle creatures by Dan Lessin, restricted to a very low muscle count. But what's possible on a much larger scale? And meanwhile, more achievably, how do growing creatures fare in the land/water switchup? (If you allow more materials.)

@fcorucci 7 лет назад

Hello Kram, thanks for your comment and for your interest in our research. Increasing the resolution is certainly possible (the encoding we're using -- CPPN) allows that. We could even produce far bigger morphologies already, using the genetic material at the end of evolution without having to rerun experiments (I think Nick presented a qualitative example of that in his 2013 paper). The only limitation when it comes to creatures' resolution is the computational time required to simulate the behavior of the robots, which increases considerably when adding more voxels. This is the reason why we tend to work with the smallest possible robot (which is, in fact, a pity). Consider that in order to run a meaningful experiment we need to evaluate a lot of individuals: say, 30 robots at each generation * number of generations (recently we tried from 1500 to 9000) * number of repetitions of the experiments to account for the stochasticity involved in the evolutionary process (15 - 20, ideally many more) * the number of different configurations we want to test. That's a lot of soft robots, and weeks of simulation time on a big super computer that sooner or later we'll end up melting down :-) Anyhow, I don't see any particular technological limitation in getting rid of the cubic workspace, and maybe we'll do that in the future. Allowing growth in the transition experiments is also very interesting, we have that in mind, but we did not have time to test that, yet. Cheers

@Kram1032 7 лет назад

Yeah, the computation time is what I was referring to. There is no guarantee that an upscaled version (by more finely sampling the CPPN) actually would behave as wanted. Any given creature's success might be an artifact of the particular points that are being sampled. I believe there even was some research roughly (though not quite) along those terms with a technique called HyperNEAT LEO, where the sampled points wouldn't be fixed in a static array but rather automatically determined by an extra output also generated over CPPN. What it effectively did was allowing to shift the sampled points along ridges generated by CPPN and slightly shifted by a mutation, thus making the space of viable geometries a lot more smooth, since small changes wouldn't as easily cause catastrophic jumps in behavior. But anyway, the bottom line is, all of these would actually have to be fully evaluated at higher scales to ensure those still work or work even better. Even just increasing the edge length by one (from 10 to 11) would increase the potential computational cost but also the potential fidelity by 33.1%. (1331 cubes instead of 1000). I know that this is very unfavourable for actually investing in larger-scale experiments, but it also seems like there are only so many different behaviors that even fit into this cube. There are some really cute variable builds in this video, so in terms of form there is quite a lot of diversity, but in terms of function the changes aren't nearly as different. Basically, from the top of my head, there are four-legged creatures, five-walled ones and two-sheeted/three-walled ones (for the swimming) and for each there is a version that manages to swim "forward" and one that swims "backwards" with any of those given shapes.

@Kram1032 7 лет назад

(assuming the computational cost is actually linear in the number of simulated cubes which I really don't know)

@Kram1032 7 лет назад

So how about this, is this computationally still feasible? Instead of limiting your shape to a box, limit it to _just_ use 1000 boxes but they can be arranged in any way, as long as they are all connected. Have your standard CPPN but with an extra layer, initialize your creature with a single cube at the 0-coordinate. The extra layer works like this: - take the current build of your creature - sample each point at the surface of the creature which would correspond to a potential cube. (Only faces count, so initially there are 6 viable positions.) - which ever gets the greatest weight, IF it is greater than a preset minimum threshold (a new parameter to fiddle with; alternatively also include that parameter in the evolution), gets to have a cube. - the CPPN you currently have going on is responsible for deciding which material goes where. What's no longer necessary is to decide that a position gets no cube. That's decided by the new layer instead. After that first step, you get 10 possible cube locations and the number grows each time. The algorithm stops once a) the box limit is reached or b) no remaining potential box passes the required threshold. This setup would at least allow structures that are much more directional in their primary shape, if not use more boxes. The only remaining problem I see right now would be what happens in the numerically unlikely case of a tie between multiple such directions. You could conceivably have very different morphologies just because one spot was chosen over another. Random tie breaking would therefore effectively mean a species with possibly significant morphological differences with the same genome. Either you are going to live with that (it could be interesting in its own right, although it would likely matter almost never) or a more sophisticated method could be employed. (Perhaps something like "the one closest to the origin or the current center of mass wins" and if it's still ambiguous, the one that's further up, then further to the front, then further to the right or something)

@fcorucci 7 лет назад

Very good observations, thanks for the suggestions. I think a few years ago Josh Auerbach (former PhD Student here at the MECLab with Prof. Bongard) did employ CPPNs to grow 3D structures following a procedure that had a similar flavor with respect to the one you're describing. Take a look at this paper: "Evolving CPPNs to Grow Three-Dimensional Physical Structures"

@gramursowanfaborden5820 7 лет назад

is it possible to enact multiple, independent waves to act as movement for separate joints, so that the creatures can evolve to be capable of walking? maybe add a random mutation for another pulse which may give rise to muscular structures as well as this sensual rhythmic jiggling..

@fcorucci 7 лет назад

Hi Mikail, thanks for yor comment. Yes, it is definitely possible!

@gramursowanfaborden5820 7 лет назад

it would be very interesting to see, i'd do it myself but lack the technical knowledge to write any of the code for it. in any case, this video has provided a great insight into the early evolution of multi-celled organisms, you and your colleagues should be proud. also, the video in itself has a wonderful whimsy to it, the combination of production style and music makes it a very pleasant and amusing video to watch, so well done for that too.

@fcorucci 7 лет назад

Thanks a lot Mikail! I'm glad you've enjoyed our work! It is indeed an exciting research field, and hopefully the best is yet to come :-) If you are interested in learning about these techniques, you might want to have a look at Ludobots (www.reddit.com/r/ludobots/wiki/index#welcome), a reddit based MOOC ran by our lab here at UVM (www.meclab.org). There are several tracks in which you can try and build your own evolutionary experiment. We are planning to start a soft bots track too, sooner or later.

@gramursowanfaborden5820 7 лет назад

Francesco Corucci it's all a little over my head but i shall have a look for sure. thanks man.

@ColinsCookingChannel 7 лет назад

I see a lot of these simulations. I love them for some reason. What if instead of (I am assuming yours is like the rest) instead of measuring performance over time you gave them an amount of energy to spend in their muscles, so that most energy efficient creatures are bred. Maybe it amounts to the same thing, idk

@fcorucci 7 лет назад

Hi Colin, I'm glad you've been enjoying our work. Thanks for your suggestion, too. Well, it would be interesting to see what would evolve under that energy constraint. This is sometimes done in computational experiments, but, usually, not when studying locomotion (as far as I know). For example, you can try and evolve foraging behavior by simulating some energy decay/consumption while allowing robots to somehow acquire energy in their task environment (e.g. collect energy "tokens" laying around, or going back to a fixed spot where they can recharge). A creature dies off when its energy level reaches zero, with fitness being based on its life span. The evolution of some pretty clever strategies has been shown in the past in this kind of setting. In a locomotion scenario, attributing a fixed energy budget and simulating energy consumption would probably result in evolution trying to optimize the transient phase (i.e., the locomotion of the robot while loosing energy), while we usually care about the regime locomotion pattern. What we've been doing in order to explore the issue of energy efficiency is to attribute a cost to active voxels in our optimization. This favor robots that manage to travel far while using few muscles over those that cover the same distance but need more muscles.

@thehomieofepicton 6 лет назад

The Land creatures were either really adorable or really funny.

@prestokrevlar 7 лет назад

2:56 ok this made me crack up so hard

@brandoncampbell9794 7 лет назад

4:09 dear god, its evolved a headcrab

@Sparton646real 7 лет назад

somewhere, in some place of the world this is someone's kink.

@phxcppdvlazi 7 лет назад

Keep up the awesome work!

@gabrielmoore1270 7 лет назад

This of course fits with smooth jazz

@drelthian1114 7 лет назад

You, my good sir, have just proved that jellyfish are the most superior being on this planet.

@wrenbeck3370 6 лет назад

4:11 that's a headcrab!

@Ava_The_Avatar 7 лет назад

outstanding.

@skiesquiggles7319 5 лет назад

Okay but it is really nice to see these things flop about

@alienplatypus7712 6 лет назад

Judging how aquatic evolution like this seems to yield far more symmetrical life forms than terrestrial evolution, if life were to form on an oceanless planet it could be perfectly possible that life simply wouldn’t evolve symmetry. Interesting stuff.

@4.0.4 7 лет назад

So let me understand, the perfect symmetry was purely emergent and it wasn't part of the mutation algorithm?

@RchamTV 7 лет назад

Actually they're symmetric because of the way their butthole is formed

@ensjo 7 лет назад

I bet symmetry emerges from the fact that symmetric bodies result in straight forward movement, and consequently longer distance, which is the fitness function.

@AJarOfYams 7 лет назад

The second to last Land->Water creature was fun to see

@eringotkilled Год назад

interesting to note the prevalence of radial symmetry among the aquatic robots and bilateral symmetry among the terrestrial ones, to the extent that many change from one symmetry to another when the environment is swapped

@re-instrumentals7144 7 лет назад

Crazy how nature do that.

@ReusSK8 7 лет назад

this is mindblowing for me.

@hazard7732 7 лет назад

Here are some more evolved creatures. A funking box.

@James-op6ls 8 лет назад

Really nice results. Do you have any ideas on how you might achieve this in real-time? Like perhaps precompute a lot of it. The body will stay constant, so it's only the brain that would need the evolution

@bigfootperson5285 7 лет назад

This is so cool

@GetToThePointAlready 7 лет назад

There's two things certain in life: 1. Death. 2. Some nerd uploading his evolution project with no accessible download link.

@--__--___- 7 лет назад

they're all so cute

@biggun8801 5 лет назад

very cool

@tuloski 6 лет назад

4:10 when dogs fu*k the pillows...

@curosuvshado 7 лет назад

3:58 lol

@respekterprivatlivet1367 7 лет назад

It reminds me of a dog scooting.

@ET_AYY_LMAO 7 лет назад

Really nice video! I like these software experiements based on evolution and neural networks etc. Did you also use some kind of NN for the locomotion or is it purely based on evolution? Thx for the interresting video!

@Alexaflohr 7 лет назад

Really cool. It would seem that symetry is more important in water than in land.

@SamuelVella1995 7 лет назад

Hey, this is absolutely awesome! How did you start of the process? What did the creatures begin with and what were the rewarded conditions (distance travelled, max speed etc)? Were the first generation just randomly generated?

@fcorucci 7 лет назад

Thanks Samuel! Yes, we start with a population of randomly generated morphologies and controllers. All individuals in the population are evaluated, and those that manage to travel further survive and produce children, which are randomly mutated copies of the parents. If you repeat this for long enough you end up achieving very effective and interesting robots. If you're interested, you can find more details about this and related works in the links in the description!

@pikminfan6778 6 лет назад

Here are some more evolved creatures. Swiss flag jellyfish!

@henryambrose8607 7 лет назад

I find it interesting how these all evolved like real-world jellyfish and other such organisms: exactly how boneless swimming organisms evolved in real life (albeit not as round)

@maxwell_edison 7 лет назад

I like how these ones seemed to be far more efficient than the previous land-only ones. Evidence we evolved from fish? Perhaps starting in water is much easier for life.

@WilliamHungVEVO 7 лет назад

I've seen enough hentai...

@CompOfHall 7 лет назад

Such cool stuff. What sort of baking times are you looking at to get results like this? How much longer would it take to do something with twice or even three times as large of a volume for the voxels? I wonder how jellyfish-like some of these water robots would become. I also wonder about radial symmetry. I would bet that some of the symmetry emerging in these trials is a direct result of the voxels and voxel space being cubes.

@patrick1532 5 лет назад

Why didn't you show the development of the creatures through each stage of evolution? That is what makes these kinds of videos interesting.

@officialpx3l80r6 7 лет назад

They are so cute! Well done! Is there a download?

@paweoriol4756 8 лет назад

Hello, this is just insanely amazing stuff... I finally see soft robot evolution treated seriously. One thing that borders me though, is that for some reason you guys may abandon what you are doing (in this case I mean switching focus, priorities), as it is so often the case with evolution projects. And you may do it before it yields the results and findings which could be just arround the corner. I humbly ask you to continue as it may lead to huge breakthroughs. And finally I see people addressing the problem of local optima and try to not just put some creatures on land and hope they will evolve ways of walking simillar to that of contemporary animals, but first try to evolve them in water thus changing the fitness landscape for the land phase, as it is dependent (the fitness landscape) on the current development of the creature in question. The problem with almost every evolution simulation that I've seen so far is that people hope that random mutation+selection will give some neat results on their own, when it all depends on the fitness landscape. It's no wonder that first there where fish which evolved undulating movements, and then these very same movements where used by amphibians/lizzards (just observe the spine) with the aid of limbs to cross rough terrain. Once you are that far the fitness landscape for moving toward fast movement becomes a ramp with fewer local optima traps. My point is that only different selective presures lead to something that otherwise only a design could yield. I will stop now, because I don't want to sound like someone who can talk the talk, but doesn't walk the walk. In half year time I will try to finish my work and if my findings are worth sharing (hard to tell at the moment) I will share my findings with you if you are interested. Before I go (and add you to my subs ;-) )I have one question to ask: can you put a figure on the % of simmetry occurrences in your creatures or would it be to troublesome? I would like to (no, in fact, I would love to) know if it's a prevalent characteristic or not. Wish you all the best, you have no idea how inspiring this video has been as far as I'm concerned! Best regards.

@fcorucci 8 лет назад

Hi Pawel, thanks for sharing your thoughts. These are very preliminary and qualitative results, but we are insisting in this direction, working hard trying to answer some of these questions. We are working on a new paper right now, which will enclose all our recent experiments with this setting (stay tuned!). We don't have any quantitative result regarding symmetry, yet, but that's definitely something we are measuring and monitoring in this setting. It indeed appears to be a common emergent trait, both in this and other experiments we've done with this system (see also: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-Cw2SwPNwcfM.html), and we are trying to investigate this aspect as well. I took a look at your videos too, cool stuff, congratulations :-)

@paweoriol4756 8 лет назад

Thanks for your reply. Can't wait for the second paper. I skimmed through this one (fascinating stuff, but TBN i'm only into the ordinary NEAT and haven't worked with CPPNs just yet, so can't fully appreciate this until I do some catching up, which I will) but once both papers are available I will study this more in-depth. Btw. if symmetry is over 50% then this would be huge. Also, the exaptation observed is mind blowing. Happy to hear that these are just preliminary results. Definitely staying tuned. Keep up the good work! And glad you enjoyed my videos :-)

@alexwelch1742 7 лет назад

Cool! just curious, do your simulations include water friction? I'm just wondering why the swimming creatures favored such flat and "un-aerodynamic" leading surfaces. Great work!

@aintnobody69 7 лет назад

This shit is quite trippy.

@pacificcoastbreeze 5 лет назад

2:00 Raise the roof yall!!

@fideliskarel6979 6 лет назад

4:08 when you're lazy but you're late to school

@17MetaRidley 7 лет назад

Beautiful! Where can I learn to do this?

@skyr8449 7 лет назад

hey, do you think you guys can also add a hard block they can use along side that?

@XeroAperture 8 лет назад

Just curious, all these critters evolved to use normal pulsating waves. Is it possible for spiral waves in the tissue to evolve?

@fcorucci 8 лет назад

Thanks for your question! Well, at the moment control is open loop (a sinusoidal signal for each voxel, with evolved phase offsets). The seemingly propagating waves that you see are in fact the result of Compositional Pattern Producing Networks "painting" a phase offset gradient (it's not actual information flow). In order to get spiral waves here we'd need CPPNs to paint such a pattern, which probably doesn't happen as it is possible to achieve effective locomotion with simpler patterns.

@a1b2c3z44 7 лет назад

also, nice song,

@heck_n_degenerate940 6 лет назад

The water ones are probably symmetrical because if they weren’t they’d be unbalanced and therefor slowly turn which would negatively affect their distance

@DT-yw4ob 5 лет назад

I read the title as "Evil Swimming Soft-bodied Creatures" and spent 10 whole seconds before noticing, just accepting that sea anemone are the spawn of the Devil.

@rangerCG 5 лет назад

One of the Land To Sea creatures moves just like a walrus/seal when it’s on land, which would make sense.

@AlexGonzalez-mr8mr 6 лет назад

2:03 Looks like 4 headless cultists performing a ritual

@teknonick3374 6 лет назад

I would love to see these evolve to want to go faster, or learn to compete with others. Perhaps more food than the other. That would be interesting :)

@thefallenprince9929 5 лет назад

I welcome our jellycube overlords

@carlostph 6 лет назад

Can energy consumption be taken into account in the objective function? Maybe the sum of accelerations of each voxel?

@Thats_Cool_Jack 7 лет назад

So are the shapes they evolved from truly random? It's really interesting how similar they are to real animals

@brendandor 7 лет назад

Awesome but that contrast of red and blue is nasty.

@LtPulsar 7 лет назад

But it does allow you to quite easily differentiate between passive cells and active ones.

@brendandor 7 лет назад

Hurts my eyes but yea it does need good contrast.

@jamesklark6562 5 лет назад

Are some voxels generated with a scale of ridgedy? Making some of them more stiff to movement than others?

@stijill 7 лет назад

what about adding light reactive cells. Even semi opaque passive cells. Pressure sensitive cells for "schooling". Even inert or ferrous minerals. Basically a different approach in folding.

@Poodleinacan 7 лет назад

I would have thought they'd try to make a robot that siphons "water" in and out through a hole it's front side (in addition to an inner system that would provide optimal motion with minimal reverse motion)..... Sort of like a mix between an airplane turbine and a squid bolting away (but I would image that in this case, rather than being large quick bursts with a setting time between them, it would hopefully become a continuous motion of "water in, water out").

@XDKvlogs 5 лет назад

I want one of these as a visualizer on my computer

@a-ragdoll 11 месяцев назад

what software is used to evolve these creatures?

@Flopsaurus 5 лет назад

Interesting how they all moved like jellyfish rather than fish

@TheFreaxTux 7 лет назад

Do the robot tissues conserve their volume?

@kalimdo2130 7 лет назад

But wat's going on if we custom gravity in the water? Y aura t-il des modifications physique?

@MyFavoriteDisease 3 года назад

1:50: the swimming first aid locker

@username4441 4 года назад

4:10 crowbar time

@faidheanta2611 7 лет назад

This stuff is fucking rad dude, I know I sound like a hippie but the stuff you're simulating here is fascinating. The one thing I would like to ask is how did you get them to change their structures? Did you implement some form of random mutation or did the robots design what they saw fit for future generations?

@Levi12O8 7 лет назад

Would this actually be possible with robots IRL? The voxels could probably be flexible, cell like (but obviously way bigger) structures with connections. The 'brain' could be in one or several of these cells if it wouldn't fit. The brain would send out commands to the active cells through the connections, each cell could have a label to make it possible. I mean, the evolution part wouldn't be but the movement could still be replicated.

@Calus767 7 лет назад

2:11 I feel like I'm watching some freaking porn again.

@THE_REDACTED 7 лет назад

Where and how do I go and create these myself?

@ItsRobinWhoTalks 7 лет назад

So the cube/flower thing literally turned itself into a jellyfish?! This is even further proof that evolution is absolute fact!

@EasyGraphicDesignca 7 лет назад

It seems that the front of all of the creatures is flat due to the voxel setup of your experiment, but is it possible that the flatness is preventing it from evolving further? Most organic aquatic life are non sharp and dont have to deal with the physics that comes with Angular shapes.

@DallyDragon 7 лет назад

When I saw the land ones I just cracked up laughing :D

@megalunalexi5601 7 лет назад

I wonder what would happen if you put some of the already-evolved ones into a tank and told them that the one that gets farthest wins? Would they start attacking each other in order to be the last one standing and win? If so, would they evolve defense mechanisms and weapons? Or would they just evolve the same way/become faster (since they're already evolved)

@s6th795 7 лет назад

So you took them for a swim then dumped them on land to evolve more. What if you had tested for fitness on both and weighted them equally, and evolved one that works well in both terrain?