AIs learn to WALK 

This is a reupload to fix sound issues

Solution 2 is so satisfying to watch, definitely my favorite

19:30 I love this part where Solution 2 seems to have a blip of consciousness like, "Hey! Wait a minute! These things just keep popping back up! What are we even doing here?!", before it bursts back into its programming. 🤣

Here's a simple reason why its more than 50% likely that two turns will be in the same direction: Consider a sequence of four points the agent travels between, A,B,C,D. Now draw the line through B and C. The turns will be in the same direction if A and D are on the same side of the line. (if you draw an example then this is clear) Say the probability a point is on one side of the line is p, so the other side is 1-p. To get the same side from A and D, you either get both left so p*p or both right so (1-p)*(1-p). So a total probability of p*p + (1-p)*(1-p) = 2p^2 - 2p + 1. This is always at least 1/2, and its only 1/2 when p=1/2 So its more than 50% likely that two turns will be in the same direction. (you can check with desmos) It's really just that if you flip a biased coin you are more likely to get the same direction twice. I have also tested it empirically and indeed if you choose 4 random points in a square then you have a 61% percent chance of the turns being the same. There might be some complications because its a sequence tho.

An idea for expanding this experiment: - limb count & length variability - segmentation possibility (not just adding another set of limbs, but making the whole creature longer/wider as they do it) - limb joint count variability (instead of each limb being a rigid stick, allowing for extra joints & muscles along the length) It'd be cool to see what types of body plans the AI comes up with.

3:45 "I also switched to a circle" Proceeds to use a square

Very cool! You have inspired me to start making my own experiments. Thank you!

Orange needed a quick nap. xD

Genius idea with the snow on the floor so they leave cute little traces.

Love your work dude - thanks for putting in so much effort!

Kind of interesting how the forward movement of each of the 4 solutions very closely mimics a different canine gait.

The amount of work, and the quality ... incredible.

Beautiful music, mesmerizing movement (love the sand addition and the particle 'explosion' effect on arrival, and the blinking of the targets.. you clearly have good taste in design and style, typically french eh ;)) and flawless presentation. (Good volume on the background information too, not too distracting).

The grey sand particles are just for show, right? With low friction of the foot, it just passes through, as if it was lifted above the sand. It would be quite interesting to see what walking tricks would evolve if the creatures had to push the sand out of the way and how different sand densities would affect this. There would probably be distinctive swimming, wading and burrowing. The creatures might learn to use an already cleared path again, building a system of tunnels over time... Your videos are really inspiring, making me think of a hundred more ideas to add. Never stop making those beautification features! Love your channel!

With every person doing these kinds of AIs I just want to watch 100 000 generations of it, with each getting a bit more complex through mutations or whatever. It's just so interesting

Very high quality stuff. Love it!

This channnel is so chill, i love it.

will an editor/source code for predator prey simulations ever be released? I really enjoyed the videos

that's so great can't wait for the more complex structures

I loved the idea of placing sand to watch where they go :) It looks amazing!

Love how the eyes are always tracking the targets 😍

this is the best youtube accounts ive found recently

Solution#2 is a legit walk that's so awesome. #3 was really interesting how it favored its right leg for power so heavily unlike the others which were a bit more symmetrical. Also I'm really surprised #1 plateaued where it did rather than refining things to something closer to #4. Really cool video thank you

Lets goo i love this channel sm

I would describe the types of movement for each solution like this: 1. Hop 2. crawl 3. Trot 4. Hop (But better able to turn)

Hmm... You've made rocket AI, Drone AI, and walking AI. I can't help but notice that all of these would be very entertaining if you paired them with the predator/prey AI... *drone phalanx attack*

Sorry for this being not related to this video, but later on whenever you feel like touching the Pred Vs Prey simulation again, I think I have some ideas that could produce interesting results: Prey remaining as-is in terms of directives, but with an added caveat that if a unit is surrounded too densely by other prey, they cannot gain nutrition. Tuning this would still allow for large groupings, but discourage simply sitting in the middle of a pile creating children that also do not move nor fight back. Other idea would be creating a middle tier creature (purple or something maybe?), that is able to eat prey, but can also be eaten by the old predator class. Possible food chain could be Large Predators (L) can eat Prey (P) in one bite, and can eat Small Predators (S) in two bites. From there, S can eat P in 2 bites, and can sometimes fight back and eat L in 3 or 4 bites. P can only fight back against S, losing to L in one bite, and do not gain food for a kill.

Imma say it again but I want to see this on a slippery surface on a race track to see them control at different speeds as they accelerate

Thought this was a sequel. I got exited😢

These are fascinating to me.

Solution 2 was easily my favourite little derp, then he just stopped dead. Now i think its my spirit animal.

So interesting that the networks had to develop a single pattern that worked well enough to control both distance traversal and turning. I'd love to see what would happen if neural networks were encouraged to develop two separate turning and distance traversal behaviors. Then to blend the two of them together using additional nodes. Sounds like another two videos!

In the race, Solution 4 (that jumps like a frog) outpaced Solution 2 (that crawls like a lizard). In real life, jumps require more energy - it's the price you pay for moving with the least contact with ground surface. On the other hand, lizard movements are really simple - you just move your limbs one by one. Thus, if you plan to enhance this learning algo and shot part 2, I'd vote for introducing "step price" aka "energy per movement" parameter - not to drastically improve the speed but instead to make it look closer to a real life! 🐸🦎

Superb! Thank you!

Liked and commented to boost algorithm as balance cos I don't intend to watch a second time so retention won't be as good. Great video tho.

Hey 👋 In the fitness function Is common to define a penalty for not solving the problem. The penalty is hard to set, but usually what I do is to define a per step small penalty that is slightly worse than doing nothing. You can also set a penalty for doing nothing for too long or for not solving the problem. The latter is useful to remove stalled agents.

the orange guy had to reconsider some life choices at the very end 😁

10 minutes ago!! Will leave it up in the background to build engagement

The main problem in wheir motions is "turning", turns take up to 50% of all time while moving. Despite the fact, that Solution 2 is quicker than Solution 1 on straight distanse, S1 is better on turns, the same goes to S3, it is moving like injured, but it is very good on turns. S4 is good in both turns and moving, thats why it is winning. In my opinion the movement pattern of S4 is better on longer distanses or higher speeds, as we can see in IRL like Felidae family do. And movement pattern of S2 is more mobile and "flexible", the example in the wild we can see as stated in Felidae family.

Its really cool to see how they resemble how real animals move even with such simple muscles. Orange is like a quadruped walking. Purple is like a quadruped running. And Green looks like a frog hopping along.

very good ❤

Salut l’ami ! J’adore ton travail

hoping for another pvp video soon!!! Very cool though

Very nice. A lot of work to implement as well as to test and document. Is this written to SFML or BGFX ?

I am more amazed by the way you did those smooth animated body thingy. the muscles and stuff like that. You are using SFML or SDL right... HOW do I do stuff like that?!

so the "real" point of efficiency is in their capacity/strategy to turn.

lets fucking go favorite channel out rn

whats the difference between 1 and 4 and between 2 and 3

Damn I feel pity to old generations who looks disabled and can't catch up young one. Reminds me my dog when he got back legs sick. Great videos man, very educational. Thank you.

19:22 Solution 2 starts taking nap

I think you should have a couple more parameters for the fitness function. Penalize energy use. This can be seen as a product of muscle contraction strength and duration. Secondly, penalize node count. This may also be framed as energy conservation as more synapses firing require more support from the host. Efficiency should be a factor considered.

hey i was wondering if you could make downloading of the ant simulation easier, Thanks!

Oh so it's not part 2... 😔

Bro I know how much work it takes to make videos like this. You are doing great quality stuff but you gotta make it shorter and more engaging. Keep up the good work though don't give up

This reminds me of the brandeis golem @ home project 🙂

Bro makes four-legged terminators

What tools / software do you use to create these simulations?

I have always asked myself this question, is it theoretically possible to have an overload of neurons and links by dint of making generations? having a network that is too large where certain “hidden neurons” come into conflict and therefore undermine the evolution of a network.

i like how solution 4 is green, as its very froglike in movement

Favoring one direction over the other. Could this be seen in humans preferring right handedness...

Late stage solution 1 seems really cool, but when you compare it to the others it just looks like its playing basically QWOP: madly flailing, which mostly works but its turning is just a mess XD

Why are they turning at all tho? Do they somehow get points for facing the target or something? Does it have something to do with the first simulation? I don't quite understand, I thought they were symmetrical.

How do you run these github projects? I don't know much at all about programming and so I was searching and searching for some hours now from my understanding I need to convert the main.cpp file into a .exe file? Is it possible that you could upload the exe file to github?

What do you make these things in? I see that you usually make stuff from scratch so I'm wondering how you run the code.

Reupload?

good vid

左右交互の方が早いと思わせて、結局4番が最速なの激熱

Solution 1 reminds me of a puppy running 2 reminds me of like a lizard 3 reminds me of a horse and 4 reminds me of a dog at full sprint And just like a lizard 2's brain just randomly shuts off haha

I don't understand how the fitness function is used to train the NN since none of the network outputs are part of the equation?

I clicked so fast

I was hoping this would be a new video But its just a reupload ):

Nice one gg

Oh 1 and 4 doff is one is pulling and other is pushing itself

Great video btw! Thank you!!

I'm curious, were they turning manually using the muscles and legs? or were the agents automatically rotating to face the goal?

connect the movement ai with the prey vs predator ai, i think that would look fun

so frogs are superior. i knew it. i will no longer walk and instead hop everywhere.

I'd be interested in seeing the final race but with a completely different target set. Something to show more "learned behavior" as opposed to "memorized movements to complete a set course". Not to say this isn't really cool! Just that I'd be interested in seeing how the various solution NNs handle a completely different "race" as opposed to what they were trained against. I feel that would give a better representation of which NN learned to walk better

I'm annoyed if music hasn't lower volume than the speach. Music should be in the background not dominate. That said, I can't remember anything special about the music in the first version of the video.

Can you make a update on 2d budget teardown? I or atleast upload it

Interesting how they never go backwards, instead they try to turn awkwardly every time. Aren't they symmetric?

Would be cool to see what would happen if you started a bunch of AIs off as omnivorous. Would they split into predators and prey? Would any omnivors remain? Or would they all stay omnivors? (Thinking about your previous vid on predators vs prey..!)

Scrungley litl guy.

Now we GAMBLE

do you remember all of the ant simulations? well i think they re missing one thing: the diging

this sand Kreygasm

Do you have discord?

Hello, I have a small question: Are you currently studying computer science or have you already graduated? please answer.

Un need more subs

I guess the details about solutions differences are kept in secret.

THEY WALK.!

Fais plus de vidéos !!!!!!

Le Frog clearly won.

Hey m8, I'm on a school project. Would want to know if we could talk about your and mine work. If you prefer we could talk about payments and those things. Let me know!!

Where part 3 of prey vs predator

à chaque fois que je recherche des gens qui font des IA, je tombe sur des français qui parlent anglais

Man I wish I knew C++

😊

Hi Pezza, please consider doing an economics based video, would be interesting to see how agents learn to trade and how it affects value. I'm sure that would be one of the most informative videos yet! Love your work

Instead of trying to make an improvement every single generation you should give it more time to make sure you're not getting stuck in a local optima