Thanks again to NordVPN for sponsoring this video. Get an Exclusive NordVPN deal here: nordvpn.com/CodeBullet It's risk-free with Nord's 30-day money-back guarantee!
When you started rewarding them for facing the right way, but they just turned their heads and kept walking backwards was like the ultimate "fuck you".
Idea: co-evolve two AIs in a game of chase. Reward one for catching the other, and one for staying away. See if they learn different walking strategies
Vitecplay - russian youtuber played the evolution style game, where he wanted to make a rolling circle, but it just kept collapsing on its own volution and then proceed to jump forward.
In fairness, I'm somewhere around 80% certain they're only doing that because CB accidentally made the back side heavier than the front, since they would, no matter what, always fall on their back when with humans that's more even between forwards and back, and walking is essentially falling forwards while continually catching yourself. Since they automatically fall backwards, they also learned to walk backwards.
@@theendersmirk5851 yeah I was kinda thinking about that. To be fair, I think it started off falling any both directions, but somehow early on it found it could get closer to the goal by falling backwards, so it became intentional by the AI. The funny thing is, I think even if he had the "look at the target" parameter in place from the start, it still probably would have fallen backwards, assuming the model could tilt its head up/back enough to see the target from the ground. Amusingly, the "point your d!ck at the target" aka have the pelvis face the objective I think would have been the best option for getting it to walk forward, since it has the least freedom of rotation separate from the rest of the body, it would have basically forced it into a forward walk. (Although if that was the case before he forced it to stay on it's feet, it probably would have started humping the air...)
@@Red-Tower Looking at the feet, the ankle joint might also be part of the cause for falling backwards all the time. Either it got reinforced to always go into plantar flexion early, or it doesn't have the joint mobility for dorsiflexion; either of those would encourage it to fall backwards, which would make the backwards walk the closer solution.
It definitely wouldn't be able to train in realtime, given how many hours it took to train, but you could absolutely train an AI separately and save the state of every generation so that you can have the AI to switch to later generations at certain points. I'm imagining a slenderman-like game, but with a more complex map, and maybe more of them spawn over time?
These aren't ever going to walk like humans. The human gait is caused by loads of factors that aren't being simulated here. Like specific joint structures
Started watching your videos a few years ago. Finished my degree in Game Programming last year. What you do is absolutely insane, and I don't fault you one bit for 2 - 4 videos a year. Having said all that... I need to see these boys walking forward, and you must absolutely bring back the death laser. EDIT: I made you a death laser and sent you a link to the repo.... soooo, you know, part three would be cool.
To make them walk straight, he would need to articulate their feet and toes, essentially adding 2 beans per leg (on for foot and one for toes) and the beauties will run for sure
Hi Code Bullet, I think the backwards issue is related to the starting position. The T pose is just a bit back heavy, so they naturally fall back and continue there movement in that direction. If you where to start them with their arms forward, they would likely learn to walk in that direction.
Honestly there's just a lot of issues to address. Like part of why we walk on two legs and forward is energy efficiency, our toes and foot flexibility, needing to see where we're going, etc. So you would just need a lot more inputs to try to make it happen in a realistic way that isn't just forced like he tried to do
I believe it has more to do with the AI wanting to use the hips as little as possible, so it prefers a backward gait due to the fact that knees can't bend forward. EDIT: more like the knees as little as possible, but exploiting a nonzero knee angle for stability
@@minamagdy4126 Well that's related to how the abdomen was simplified so much that the AI had much fewer option for maintaining it's balance than a person would. There were just a lot of problems relating this model to human movement, I doubt I could list them all. With that said, it was uncannily close in some ways despite those shortcomings.
They did pretty well considering they had no inner ear, spine or directional awareness related to their face, a working spine and the world. Something to give them priority for facing forward without twisting the neck, something to favour keeping the head relatively level with the horizon and something to favour holding a more upright position.
I think the proprioception idea is a good one, but I don't think its enough alone. In us humans, that sense is combined with the inner ear (which conveys a sense or roll, pitch and yaw) and vision (which provides directional targeting). So along with a sense of proprioception, you need to simulate an inner ear and a "cone of vision." Alas, you would have to train AIs to this task incrementally though. For example - you'd have to start training vision first (where it looks for a target in its cone of vision, and if it's not there, it turns its head/body until it finds it). The inner ear should already be working for this part of the training (so a sense of perceived direction and range emerges). And once it's learned to look at a target, then you can then try to train them to walk. As long as they learn to rotate themselves to look at the target first - they should walk forwards towards it. Moral of the story? They should look before they leap. :D
(To be clear by what I mean by a cone of vision... it's basically an invisible non-clipping cone that extends from where the eyes would be that has the same field of vision and range as human vision. For it to register anything as "seen" it must be inside the cone).
Another important detail the AI is ignoring is effort. Walking upright in the human style is extremely efficient. Walking bent over backwards with arms flailing in the air like the AI did is extremely difficult. If they were made to minimize energy use, I think the results would improve
I guess the models' feet are rigid, but the movement of toes plays an important role when walking and running, so adding another joint behind the toes might help a lot for AI to learn walking. (especially not backwards :) Additionally, I think instead of killing them when they fall, killing them only when their head touch the ground and giving serious penalty points for touching their body parts (may differ for front and back also) other than bottom of the feet, also giving less penalty points for hands may force them to learn getting up when they fall. This approach may make them learn to stand up properly and walk. Moreover, if you give them slight penalty points for touching the bottom of their feet than they may learn to jump towards the target which may lead them to learn to run even before walking :D All these might lead them learn to run on all fours at first, but they will learn to stand up eventually if the penalty and reward points are balanced well. If speed is rewarded more, then they will stay on all fours, but in this case if the height of the head is penaltied or rewarded, then standing on feet will be flourished. Maybe they can run like Tom Cruise some day :D Just speaking my mind, you are the expert :)
19:14 I love that they're so committed to the backwards thing that they hit the target, then _turn around so that they can still be backwards_ before heading to the next one.
They never learned how to walk forwards. That was probably the same issue with the turning necks thing: he added the "face your target" reward to a bot that already only knew how to walk backwards, so rather than re-learn an entirely new skill, they just turned their heads as part of the existing skill.
It is honestly amazing that CB managed to get them to walk without falling over. I expected the usual AI exploits physics engine to happen and the guys to vibrate across the floor at light speed instead
I think the main 2 reasons for the AI behaving the way it does is likely because: a) friction is not quite correct making it not actually use their legs to push forward but their lean over weight to pull them. b) I assume every bone is the same weight and every joint the same strength which causes strange priorities. Oh also there could be a added reward to keep the center of gravity beneath the head within a range that might also help
Agreed. Also, there are a lot more things that could be done to improve it. For example: calculating the energy expenditure and rewarding the most economical ones; rewarding those that the time of arrival is closest to a specific set goal (having the remaining time as an input together with the distance to target will help) will also make them learn different kinds of gaits for different desired speeds. Of course, changing how much reward each variable gives will drastically change the outcome.
@@somedude4487 Hmm how to describe this. Irl friction is a force that opposes movement. Pushing against it will let you use it to move i to the opposite direction of where friction goes. (Kind of think how high jumping with a stick works they use the elasticity of the stick and the opposing friction to propell themselves upward) However in this game friction presumably works closer to how magnets work. A magnet wont let you push against itself since it always pulls you back. A magnet will never push on you so you cannot use it to push yourself forward. So the Ai can only pull its feet off the groumd not push its body off the ground and thid movement is the result.
At 0 hours of training, the AI behaves remarkably like a newborn infant. The movements are random because it has yet to develop "awareness" of its limbs and how to use them. I just thought it was a really interesting comparison.
Imagine if he just left the crawling AI to learn for generations, I wonder if any of them would've evolved to walking like how a human learn to crawl first then learns to walk.
@@keysersoju9823 I doubt it. Newborns move on to walking eventually because crawling is more tiring, and hurts the hands and knees after a while, and also probably because they see everybody else walking. Without the punishment for parts other than feet touching the ground to simulate the discomfort of the first two reasons, it would not have any reason to stop crawling
it’s like watching infants learn, just like babies the AI only seems to know how to *push* at first lol. probably why it’s more comfortable moving backwards
I think the balance is off on the ai models. I think there's more "weight" at the back of the model so they end up stumbling backwards because they can't fall forward
@@Diathan Don't reply to bots (even the "ironic" ones). It just makes them look more legit to the TH-cam algorithm, and leaves a link to their channel in your post. Just report and move on.
@@Dark-Stryder Not really. If you move half way to the target in the same time another version moves 75% of the way. The further one would have been rewarded more points. So the one that got further is less time is already being reqarded better.
@@Krell356 I'm not an AI specialist but, if theres multiple targets and no cap on the score said AI could get in a single generation, theoretically a timer would still make it go faster. It's probably not necessary for this and would probably lead to the AI just discovering something unintentional like leaping towards the target as the timer is about to run out. Without a timer though, the AI that moves 75% of the way to the target in the same time as the 50% of way would still be rewarded more points but, a target that moves 50% of the way in 3 minutes vs 50% of the way in 3 seconds is also a huge gap where a timer would improve it, and in turn teach it to go faster.
@@axeldasilva8060 -- Australia is uninhabitable due to the temperatures and the spiders and a million other nasties. France is uninhabitable due to the French.
So, we actually have neurons in our spines that generate cyclic rhythm for walking. Look up Central Pattern Generator. It's pretty cool to generate cyclic impulses out of just a bunch of neurons connected in a loop (albeit, that obviously does not fit the layers of neurons models you can find in most NN packages...). Second important thing in walking, is that proprioception is used for reflex acquisition. Eg. certain angles of certain joints trigger the next movement in the sequence. Which for "AI simplicity" can possibly be done as a feedback/reset input on some CPGs.
^^^^Genius. I think rewarding them for having an average floor touches/second over a given amount of time (or whole run) would be a good way to implement this slowly, avoiding incessant steps. Another way you could diversify the AI’s ability would be to set the cyclic rhythm goal at a random value within a reasonable range.
I saw a NN video the other day in which they simply included an oscillator as one of the inputs; perhaps giving the AI output control over the frequency of those oscillations might be a simple way for it to feed itself information... Just a handful of sine waves, etc? I get that with enough information it might be able to synthesize cyclical motion based on its current state, but wondering how many shortcuts/better architecture there is for locomotion activities.
While I buy a dakimakura for my crippling loneliness, Code Bullet makes his own AI so he has friends which will follow him. Now that's some determination!
Pretty sure the reason they walked backwards is that they fell backwards in the default position, causing them to take a step backwards to prevent falling over. If you would just have tilted them forwards a little at the start, the first step would have been forwards and they would have continued walking forward. The backwards thing was funnier tho soI can see why you didn't try too long and hard to fix it.
Interesting... I thought that the problem was that the feet seem to be modeled as one inflexible object... and obviously, feet don't work that way. I thought that at least making the feet hinge at the toes would avoid the backwards, tippy-toed approach the AI is using. Or is could be a little of both of our ideas together might finally get hem to walk forward.
Give them a reward for moving in the direction of their front, a reward for keeping their shoulder height up (but penalized for shrugging XD), and a penalty that increases the longer their toes are on the ground without their heels also touching the ground, and make them have a vision cone for finding the target they have to move to :D
perhaps the toe stance discouragement could begin if there is 1 to 3 seconds of toe touch without a heel touch (possibly also the reverse to discourage heel walking too) making the foot a little rounded in the back and with a flexible front section (toes) might also help to better match real feet. as for rewarding standing, reward head above shoulders, shoulders above torso, torso above legs, upper arm above lower arm, upper leg above lower leg, lower leg above foot (probably by failure hurting score, rather than success helping it much)
Your comment made me realize this is pretty much exactly what pain does for humans. Maybe penalize the AI slightly any time it overextends or over taxes its joints? Cause that shit would hurt if a human tried it.
Indeed bro, also, I think that what the AI learned is actually more applicable to 4 legged animals walking with the way the have learned how the bent knees are the actually going forward.
By Golly, not only the themes you chose for your videos and your skill to actually do them and show them are more than on point, but also your editing and silly jokes are hilarious, I cried from laughter more than once.
Actually if he could get this working where they walk forward then change the avatar to a zombie then give yourself a shotgun I think he might have the start of a epic zombie game
I hope one day that the person who made the model with all the hopes in the world that its a self input model of himself, and he finds the weird stuff codebullet does, and enjoys what is being done to his virtual self.
@@krishanSharma.69.69f didn't google successfully create an AI that created a new method for training AI? Edit: Because no one believes me I did some more research. The AI I'm thinking of is called AutoML, it uses an AI model built by Google to produce AI models faster and more accurately than those produced by human training.
We have been blessed by two code bullet videos in like a week. They are for the same thing though and we probably won't see him again until in half a year but we are grateful
I liked the video up until the point you made their necks twist 180 degrees to circumvent the back-facing walking, I would've loved to see them actually walking, I think the video was really funny and wish to watch part 3, with more accurate posture and speed, maybe less disjoined necks
I find it interesting how similar this AI learns to walk as a baby does. Start off just randomly flailing with no real sense of what they’re doing, then slowly work out how to crawl towards an objective. It’s shockingly similar to real living things learning to walk, even considering all the jank
@@jtcoding6422 Yes. Based off, but does not work as our brains do. There are many differences. And, neuroscience isn't far enough to even determine how the brain actually works. But here's a simple example to show the differences. Neurons in neural networks are not like brain neurons at all. A brain's neuron is either on or off. A neuron in a neural network, however, can be on from any range between 0 an 1. You can have a 0.5 activated neuron in a neural network. Yes, it is true, neural networks are based on how the brain works. But to say they work as the brain does is a misconception and false. And our scientific knowledge is not far enough to even determine something like that yet.
This two-part video contains an impressive amount of getting sidetracked. Hilarious, I might add. Watching AI fail at walking is an amazing and comforting sight. Laughing at real people falling feels bad because they get hurt, but watching an AI-controlled puppet fall in virtual 3D seems harmless
I think encouraging the chest to point at the target would help immensely. Also, wouldn't mind seeing running, and/or jumping if a part 3 is made. Maybe hands that have a stick to things and not stick to things state could be added for parkour? But, you did make walking AI, so if you are done, then so be it.
Nah, that's too artificial. Humans don't have a reward for keeping our chest pointed at things. What we have is a reward for looking at things (seeing where we're going) and a penalty (pain) for over-rotating our neck.
Great video, I think the backwards walk stems from you spawning the guys in random orientations to the target while training. Most of the time they will be facing away from it and because they dont get rewarded for moving away from the target briefly they never turn, therefore developing the backwards walk. When you introduced the reward for looking at the target you continued with already trained models so they just turned their heads instead of developing something new. Might be interesting to see what happens when you retrain with them always starting facing the target, as knees only bend in one way and can be untilized differently when walking forward in comparison to backwards.
As funny as VR shenanigans are, now I'd really like to see the chest-and-dick-facing-target thing with like 100h of training to really see what's possible.
Dude! Two videos in less than a week! Oh my GOD. That never happens. Looks like the AI that writes the script for the videos has been fine-tuned quite a bit...
Honestly bullet, I really loved the way you did this video. Loved the editing, loved the way you through in the extra biat besides telling us what you're doing. Overall it was a new experience in the way you did your video. Would love to see more of this style. But I do alao enjoy your normal stuff too
Great video! Instead of positive reinforcement for facing the goal, you could also try making the AI feel “tired” from using their core or leading back, so they look for efficiency which may avoid the backwards walk
Now I want to see forward only learning for a week straight. Could we ever get relaxed gentleman walking or athlete running? What if you just train them to stand still for awhile, and then give them a target and train them to turn to face it with as little neck rotation as possible, more points for the closest to neck straight forward.
the reason humans walk efficently is because we actually expend energy so maybe assign an energy consumption to each action give them a determined amount of energy see who goes the farthest
@@HurricaneLantern Also pain. Crawling around on your elbows and back, with your head hitting the ground, would be ridiculously painful regardless of how energy efficient it is.
Thanks again to NordVPN for sponsoring this video. Get an Exclusive NordVPN deal here: nordvpn.com/CodeBullet
It's risk-free with Nord's 30-day money-back guarantee!
BOO!
I will never use nord vpn
code bullet loves money!
wait a minute this says 7 days ago
First😊
He really did have the second video ready. So proud Mr. Bullet.😊 With two videos in a week, I look forward to the next video in 2024.
@SiZzyVFX that’s being optimistic
I was so confused when the 2nd part came out earlier than 2 years later
Think you misspelled it's 2034
Being a bit hopeful there aren’t we?
@@Inveist 3034*
When you started rewarding them for facing the right way, but they just turned their heads and kept walking backwards was like the ultimate "fuck you".
So true
They like to take a shortcut 😂
Time stamp
@@anshik.k.t 16:50
They followed the instructions exactly as given
Seeing the time lapses of the AI in their individual cubes feels disturbingly dystopian
Reminds me of the Truman Show
What's with the santa account that I assume is a bot given the obvious copy paste message all of them have
@@Dolat1984 AI learned to scam
@@stupit467 dear god AI is taking scammers jobs too hahaha
@@stupit467 the next logical step
Idea: co-evolve two AIs in a game of chase. Reward one for catching the other, and one for staying away. See if they learn different walking strategies
There already is one, search for prey and predator ai simulation on yt
@@drm.himself yeah but do those just squirm of the floor or prefer to walk backwards like our boys here?
@@lui5gif It's not 3D, but some of them do walk backwards
Isn't that literally what the Google AI did to beat DeepBlue?
Well this should play out just fine in my dreams tonight. Hopefully they all stay the same size
This guy is the most sane Aussie programmer I’ve seen
I like how all the most popular aussie youtubers are also the “most sane”
low bar tbh
E
@@enlightenedbanana it's like being the most lively corpse in the morgue
@@enlightenedbanana I think "How to Basic" is Australian. I know Max and Chad from "Cold Ones" are so maybe that levels the balance a little.
AI is finally being added to this AI project, and it only took half of a half hour video of homoerotic character rigging to get this far
;)
Code bullet is an ironic name, because the videos take ages to get here
fr
Your usual code bullet videos right there
And We Also Need To Wait Ages For This Series
AI finding a loophole and looking backwards instead of learning how to walk forward is oddly relatable
Vitecplay - russian youtuber played the evolution style game, where he wanted to make a rolling circle, but it just kept collapsing on its own volution and then proceed to jump forward.
In fairness, I'm somewhere around 80% certain they're only doing that because CB accidentally made the back side heavier than the front, since they would, no matter what, always fall on their back when with humans that's more even between forwards and back, and walking is essentially falling forwards while continually catching yourself. Since they automatically fall backwards, they also learned to walk backwards.
@@theendersmirk5851 yeah I was kinda thinking about that. To be fair, I think it started off falling any both directions, but somehow early on it found it could get closer to the goal by falling backwards, so it became intentional by the AI.
The funny thing is, I think even if he had the "look at the target" parameter in place from the start, it still probably would have fallen backwards, assuming the model could tilt its head up/back enough to see the target from the ground. Amusingly, the "point your d!ck at the target" aka have the pelvis face the objective I think would have been the best option for getting it to walk forward, since it has the least freedom of rotation separate from the rest of the body, it would have basically forced it into a forward walk. (Although if that was the case before he forced it to stay on it's feet, it probably would have started humping the air...)
@@Red-Tower Looking at the feet, the ankle joint might also be part of the cause for falling backwards all the time. Either it got reinforced to always go into plantar flexion early, or it doesn't have the joint mobility for dorsiflexion; either of those would encourage it to fall backwards, which would make the backwards walk the closer solution.
E
Now you need to make a horror game where throughout the game, the enemies AI progressively gets better and better at walking as they chase you.
It definitely wouldn't be able to train in realtime, given how many hours it took to train, but you could absolutely train an AI separately and save the state of every generation so that you can have the AI to switch to later generations at certain points.
I'm imagining a slenderman-like game, but with a more complex map, and maybe more of them spawn over time?
Rainworld
Especially if they enable the full 180 head turn over time too
underrated
Part three? We still need to see them walk like humans.
Wait, you don't walk like this?
wtf
What if they could actually run towards you. That woundn't be scary at all.
@@Known_as_The_Ghost nah bruh we fly 💀
@@juicetinthemakar Bro you don't conjure a bubble of blood and other fluids to swim along the ground in?
These aren't ever going to walk like humans. The human gait is caused by loads of factors that aren't being simulated here. Like specific joint structures
I honestly thought he would just never do the rest. Finishing a project is very out of style, you go Evan!
But he didn't finish. He gave up before getting them to walk forward or stand up after falling!
@@ZoltarDeathNnja bullet found a way of turning into a homoerotic fantasy AGAIN
@@brunogonzalez7453 Did he find a way to do it, or did they AI force it upon us?
@@ZoltarDeathNnja Lol
@@ZoltarDeathNnja If so, I'm a big fan of this AI
Started watching your videos a few years ago. Finished my degree in Game Programming last year. What you do is absolutely insane, and I don't fault you one bit for 2 - 4 videos a year. Having said all that... I need to see these boys walking forward, and you must absolutely bring back the death laser.
EDIT: I made you a death laser and sent you a link to the repo.... soooo, you know, part three would be cool.
I admire you being the change you want to see in the world
@@strykerwaller3784me too. Don’t stop that
23:20
Laser Wall: th-cam.com/users/shortscYxzTAuTYWM?si=M2atFnIMKVcFrUXi
could milk a third video from this making them walk straight
Oof. I mean, true, but oof.
Or make it animals walking on 4 legs
Girl they were sucking each other's dick within 2 hours of being created they would never be able to do something straight
That’s a big ask, does anything in this video seem straight to you?
To make them walk straight, he would need to articulate their feet and toes, essentially adding 2 beans per leg (on for foot and one for toes) and the beauties will run for sure
"we're gonna have to force them to stand up" yeah because they were gonna evolve into crabs if left unchecked
Because in this universe everything evolves into crabs eventually
@@AkoSolius the natural state of being
Hi Code Bullet,
I think the backwards issue is related to the starting position. The T pose is just a bit back heavy, so they naturally fall back and continue there movement in that direction. If you where to start them with their arms forward, they would likely learn to walk in that direction.
There’s a lot of smart people on the Internet and this is one of them.
So the problem is that their booty is too _thicc_
Honestly there's just a lot of issues to address. Like part of why we walk on two legs and forward is energy efficiency, our toes and foot flexibility, needing to see where we're going, etc. So you would just need a lot more inputs to try to make it happen in a realistic way that isn't just forced like he tried to do
I believe it has more to do with the AI wanting to use the hips as little as possible, so it prefers a backward gait due to the fact that knees can't bend forward.
EDIT: more like the knees as little as possible, but exploiting a nonzero knee angle for stability
@@minamagdy4126 Well that's related to how the abdomen was simplified so much that the AI had much fewer option for maintaining it's balance than a person would. There were just a lot of problems relating this model to human movement, I doubt I could list them all. With that said, it was uncannily close in some ways despite those shortcomings.
They did pretty well considering they had no inner ear, spine or directional awareness related to their face, a working spine and the world. Something to give them priority for facing forward without twisting the neck, something to favour keeping the head relatively level with the horizon and something to favour holding a more upright position.
I think the proprioception idea is a good one, but I don't think its enough alone. In us humans, that sense is combined with the inner ear (which conveys a sense or roll, pitch and yaw) and vision (which provides directional targeting). So along with a sense of proprioception, you need to simulate an inner ear and a "cone of vision." Alas, you would have to train AIs to this task incrementally though. For example - you'd have to start training vision first (where it looks for a target in its cone of vision, and if it's not there, it turns its head/body until it finds it). The inner ear should already be working for this part of the training (so a sense of perceived direction and range emerges). And once it's learned to look at a target, then you can then try to train them to walk. As long as they learn to rotate themselves to look at the target first - they should walk forwards towards it.
Moral of the story? They should look before they leap. :D
(To be clear by what I mean by a cone of vision... it's basically an invisible non-clipping cone that extends from where the eyes would be that has the same field of vision and range as human vision. For it to register anything as "seen" it must be inside the cone).
Another important detail the AI is ignoring is effort. Walking upright in the human style is extremely efficient. Walking bent over backwards with arms flailing in the air like the AI did is extremely difficult. If they were made to minimize energy use, I think the results would improve
E
or fall to the ground like a ####### loser
this is the guy code bullet is scared of
Fellow fans of Code Bullet: it doesn’t matter if you watched the previous video before this one.
Yeah, but I think this one hits different if you did. I personally think it was worth it.
It was totally worth watching but like.... Yeah you really don't need it for context
Honestly, I didn’t like the first video very much, this video is much better
Well, if you're a fan of Code Bullet, it helps him out a lot if you watch both videos!
If you didn't watch the first one, how would you know humans are just beans?
Honestly love the line graph bit
👍
Best
Nice
Should have been left to right, but still nice 👍
Best
I guess the models' feet are rigid, but the movement of toes plays an important role when walking and running, so adding another joint behind the toes might help a lot for AI to learn walking. (especially not backwards :)
Additionally, I think instead of killing them when they fall, killing them only when their head touch the ground and giving serious penalty points for touching their body parts (may differ for front and back also) other than bottom of the feet, also giving less penalty points for hands may force them to learn getting up when they fall. This approach may make them learn to stand up properly and walk. Moreover, if you give them slight penalty points for touching the bottom of their feet than they may learn to jump towards the target which may lead them to learn to run even before walking :D
All these might lead them learn to run on all fours at first, but they will learn to stand up eventually if the penalty and reward points are balanced well. If speed is rewarded more, then they will stay on all fours, but in this case if the height of the head is penaltied or rewarded, then standing on feet will be flourished. Maybe they can run like Tom Cruise some day :D
Just speaking my mind, you are the expert :)
You are absolutely right, be he will avoid doing that kind of work? And instead waste his time with insufficient methodologies :)
I’m really sorry but you lost me at *TOES*
19:14 I love that they're so committed to the backwards thing that they hit the target, then _turn around so that they can still be backwards_ before heading to the next one.
They never learned how to walk forwards. That was probably the same issue with the turning necks thing: he added the "face your target" reward to a bot that already only knew how to walk backwards, so rather than re-learn an entirely new skill, they just turned their heads as part of the existing skill.
Part 3: Ai learns to run
Part 4: Ai learns to hug
Part 5: Ai learns love
Part 6: I married my AI
Part 7: WE GOT KIDS!
Part 8: Help
@@urmum8540 part8:the divorce
Part 9: depression
@@audionewpaper1 Part 10: What was the AI cooking?
E
I love how they immediately find a way to exploit the system to avoid walking forwards
this is genuinely one of the funniest videos I have seen in a loooong time, my throat hurts from laughing so hard. You're legendary CodeBullet.
It is honestly amazing that CB managed to get them to walk without falling over. I expected the usual AI exploits physics engine to happen and the guys to vibrate across the floor at light speed instead
well the several constrain prevent that from happening.
@@kivylius Plus he's using a much more popular physics engine (PhysX, Unity's physics engine), so most of the easy to trigger exploits are patched.
I think the main 2 reasons for the AI behaving the way it does is likely because:
a) friction is not quite correct making it not actually use their legs to push forward but their lean over weight to pull them.
b) I assume every bone is the same weight and every joint the same strength which causes strange priorities.
Oh also there could be a added reward to keep the center of gravity beneath the head within a range that might also help
Agreed. Also, there are a lot more things that could be done to improve it. For example: calculating the energy expenditure and rewarding the most economical ones; rewarding those that the time of arrival is closest to a specific set goal (having the remaining time as an input together with the distance to target will help) will also make them learn different kinds of gaits for different desired speeds.
Of course, changing how much reward each variable gives will drastically change the outcome.
Technically we also use our weight to walk, so the friction is fine, the AI just found a local maximum to attach itself to
@@somedude4487 Hmm how to describe this. Irl friction is a force that opposes movement. Pushing against it will let you use it to move i to the opposite direction of where friction goes. (Kind of think how high jumping with a stick works they use the elasticity of the stick and the opposing friction to propell themselves upward)
However in this game friction presumably works closer to how magnets work. A magnet wont let you push against itself since it always pulls you back. A magnet will never push on you so you cannot use it to push yourself forward. So the Ai can only pull its feet off the groumd not push its body off the ground and thid movement is the result.
ok
@Yara Yo The Chinese/Japanese (I don’t know either give me a break) bots have infected the English comment threads!
he finally mastered the art of making the audience believe he is dead and then boom he is back
Are we sure he isn’t a zombie
@@arfb01 Now that you mentioned it, his face does look awfully green...
E
@@EEEEEEEEe
@@EEEEEEEEe 🫡
Accidentally created a JoJo pose generator in the process of trying to get a hot man to walk.
At 0 hours of training, the AI behaves remarkably like a newborn infant. The movements are random because it has yet to develop "awareness" of its limbs and how to use them. I just thought it was a really interesting comparison.
If my baby started doing this im taking it to the hospital
Yeah, computer science can be surprisingly childish.
Which is strangely not a derogatory term, here. **g**
Imagine if he just left the crawling AI to learn for generations, I wonder if any of them would've evolved to walking like how a human learn to crawl first then learns to walk.
@@keysersoju9823 I doubt it. Newborns move on to walking eventually because crawling is more tiring, and hurts the hands and knees after a while, and also probably because they see everybody else walking. Without the punishment for parts other than feet touching the ground to simulate the discomfort of the first two reasons, it would not have any reason to stop crawling
So from the whole beginning I was just an AI
it’s like watching infants learn, just like babies the AI only seems to know how to *push* at first lol. probably why it’s more comfortable moving backwards
I think the balance is off on the ai models. I think there's more "weight" at the back of the model so they end up stumbling backwards because they can't fall forward
@@SilverWolf_-cj4qn cuz of all that cake
@Rahaf Ahmed THE BOTS ARE LEARNING NEW LANGUAGES
GOD HELP US ALL
It feels strange to have 2 codebullet videos to watch in the span of 3 years
10*
@@Diathan Don't reply to bots (even the "ironic" ones). It just makes them look more legit to the TH-cam algorithm, and leaves a link to their channel in your post. Just report and move on.
E
@@HiddenWindshield not a bot hahahah
It’s amazing seeing the AI use the arms to maintain balance
E
Makes ai walk and the first thing he does with it is make them do the fitnessgram pacer test
Ok who the mofo deleting all the comments telling telegram Evan bot to *shut*
Honestly love the line graph bit. It also totally would work. Hours Trained x Walkability.
This made me realize why crabs keep independently evolving
the fact that codebullet got a bunch of buff men to moonwalk at him chode-first is truly the peak of AI
Omg, a vr horror rougelike where the ai monsters progressively get better at chasing you
Create a time limit on how long it takes them to reach the target so they have to learn to sprint
In theory they should already be getting rewarded more frequently based on the distance closing so a timer shouldn't be necessary.
@@Krell356 it would speed up the process tho wouldn't it?
@@Dark-Stryder Not really. If you move half way to the target in the same time another version moves 75% of the way. The further one would have been rewarded more points. So the one that got further is less time is already being reqarded better.
@@Krell356 I'm not an AI specialist but, if theres multiple targets and no cap on the score said AI could get in a single generation, theoretically a timer would still make it go faster. It's probably not necessary for this and would probably lead to the AI just discovering something unintentional like leaping towards the target as the timer is about to run out.
Without a timer though, the AI that moves 75% of the way to the target in the same time as the 50% of way would still be rewarded more points but, a target that moves 50% of the way in 3 minutes vs 50% of the way in 3 seconds is also a huge gap where a timer would improve it, and in turn teach it to go faster.
I think that yes, this could be a weight that can be conditionally added into the model once the AI model manages to reach goal 1
10:18 He says he's connecting to habitable country while showing france truly the funniest person on the internet
I don't get it
I mean, compared to Australia? As funny as "huhuhuh french people silly dumb cowards oui oui" is, Australia do be fucking savage.
@@axeldasilva8060 me neither
@@axeldasilva8060 -- Australia is uninhabitable due to the temperatures and the spiders and a million other nasties.
France is uninhabitable due to the French.
fr*nce 🤮🤢🤢
So, we actually have neurons in our spines that generate cyclic rhythm for walking. Look up Central Pattern Generator. It's pretty cool to generate cyclic impulses out of just a bunch of neurons connected in a loop (albeit, that obviously does not fit the layers of neurons models you can find in most NN packages...). Second important thing in walking, is that proprioception is used for reflex acquisition. Eg. certain angles of certain joints trigger the next movement in the sequence. Which for "AI simplicity" can possibly be done as a feedback/reset input on some CPGs.
^^^^Genius. I think rewarding them for having an average floor touches/second over a given amount of time (or whole run) would be a good way to implement this slowly, avoiding incessant steps.
Another way you could diversify the AI’s ability would be to set the cyclic rhythm goal at a random value within a reasonable range.
It turns out walking is hard to replicate by wiggling joints at random until you get closer to a target point.
Joint angle proprioception determining the next timestep's angle outputs reminds me of ohmganesha's kinesthetic worms
I saw a NN video the other day in which they simply included an oscillator as one of the inputs; perhaps giving the AI output control over the frequency of those oscillations might be a simple way for it to feed itself information... Just a handful of sine waves, etc?
I get that with enough information it might be able to synthesize cyclical motion based on its current state, but wondering how many shortcuts/better architecture there is for locomotion activities.
biologist entered the chat
9:12 “AI Learns to Eat Ass” video when?
While I buy a dakimakura for my crippling loneliness, Code Bullet makes his own AI so he has friends which will follow him. Now that's some determination!
Pretty sure the reason they walked backwards is that they fell backwards in the default position, causing them to take a step backwards to prevent falling over. If you would just have tilted them forwards a little at the start, the first step would have been forwards and they would have continued walking forward. The backwards thing was funnier tho soI can see why you didn't try too long and hard to fix it.
Interesting... I thought that the problem was that the feet seem to be modeled as one inflexible object... and obviously, feet don't work that way. I thought that at least making the feet hinge at the toes would avoid the backwards, tippy-toed approach the AI is using.
Or is could be a little of both of our ideas together might finally get hem to walk forward.
@@buckykattnj nah, the real fix is to put a clone with a machete behind them that already knows how to walk.
Also the first part (without touching penalty) could be interesting with forward falling -- sort of animal 4-leg walk
Maybe randomizing the falling direction as well as adding a vestibular sense would allow for a more robust locomotive system.
Give them a reward for moving in the direction of their front, a reward for keeping their shoulder height up (but penalized for shrugging XD), and a penalty that increases the longer their toes are on the ground without their heels also touching the ground, and make them have a vision cone for finding the target they have to move to :D
perhaps the toe stance discouragement could begin if there is 1 to 3 seconds of toe touch without a heel touch (possibly also the reverse to discourage heel walking too) making the foot a little rounded in the back and with a flexible front section (toes) might also help to better match real feet. as for rewarding standing, reward head above shoulders, shoulders above torso, torso above legs, upper arm above lower arm, upper leg above lower leg, lower leg above foot (probably by failure hurting score, rather than success helping it much)
No need for a reward for moving in the direction of their front. Instead, penalize exponentially for head pivoting away from neutral.
@@goldfndr Yes, that worked for me. I tried to walk like them and my neck started hurting after 5 seconds, so I turned around.
Your comment made me realize this is pretty much exactly what pain does for humans.
Maybe penalize the AI slightly any time it overextends or over taxes its joints? Cause that shit would hurt if a human tried it.
Indeed bro, also, I think that what the AI learned is actually more applicable to 4 legged animals walking with the way the have learned how the bent knees are the actually going forward.
By Golly, not only the themes you chose for your videos and your skill to actually do them and show them are more than on point, but also your editing and silly jokes are hilarious, I cried from laughter more than once.
i wanna see a part 3 where they walk like you know humans
Humans... CAN walk like that... technically
what do you mean? how do YOU walk?
Same
You don’t walk with your shoulder blades?
@@OKayD3N when he sped it up it looked like a cockroach race
2 videos in 1 year? we must have been nice this year to deserve this much content
I think that the sight of several men walking towards Bullet in the most unhuman way possible was worth all the effort he put into this.
10:50 I love this guy’s humor and comedic timing. I can’t believe I’m saying this for the ad segment.
4:55 that cursed Section went on for far too long. It was perfect. 10/10 Content :)
23:24 "My God that is hard to look at". No pun intended.
Then proceeds to turn the camera back and look again.
Actually if he could get this working where they walk forward then change the avatar to a zombie then give yourself a shotgun I think he might have the start of a epic zombie game
That’s exactly what I was thinking. It could be an insane vr zombie game
could be an interesting concept. A wave based shooter where the enemies gradually get smarter as the game progresses.
“I don’t know why they’re walking backwards” *spawns them facing away from the target
Well, we will never see him again, he learned how to walk away from us thanks to AI
Made himself AI friends so he doesn't need us anymore 😭
@@Red-Tower no, he'll be back, he still has to eat and without youtube, he is doomed
E
@@Ferrari255GTO He can eat the AI
@@EEEEEEEE E
I wonder what would have happened if the “touching the floor” punishment were a reduction in score instead of immediate failure…
Or if instead of a penalty, it was a reward? I want to see how fast they could worm 🪱.
Rolling around at the speed of sound
the long awaited sequel finally arrives
It been 4 days but yep
I hope one day that the person who made the model with all the hopes in the world that its a self input model of himself, and he finds the weird stuff codebullet does, and enjoys what is being done to his virtual self.
This man, whenever he releases a new video I rewatch him for days
Like the same video over and over or going back to his back log?
You should teach ai to code and make an endless loop.
teaching an A.I. to code would be very interesting but may end up creating skynet lol
@@MultipleC9 that is exactly what most humans are trying to make ... a recursive self improving AI.
@@krishanSharma.69.69f didn't google successfully create an AI that created a new method for training AI?
Edit: Because no one believes me I did some more research. The AI I'm thinking of is called AutoML, it uses an AI model built by Google to produce AI models faster and more accurately than those produced by human training.
@@samwilde8311 nah, that's media bullshit
There already is an ai that can code btw.
We have been blessed by two code bullet videos in like a week. They are for the same thing though and we probably won't see him again until in half a year but we are grateful
These videos are beautiful. Please never stop being chaotic Code.
I liked the video up until the point you made their necks twist 180 degrees to circumvent the back-facing walking, I would've loved to see them actually walking, I think the video was really funny and wish to watch part 3, with more accurate posture and speed, maybe less disjoined necks
agreed, it's a shame he didn't added that chest and d**k reward.
Yea same
I agree, the neck thing was pretty weak
I agree as well
This teaches us that feet without muscles or functional toes are really just in the way
I find it interesting how similar this AI learns to walk as a baby does. Start off just randomly flailing with no real sense of what they’re doing, then slowly work out how to crawl towards an objective. It’s shockingly similar to real living things learning to walk, even considering all the jank
That's because ai literally works as our brains do
@@__u__9464 It actually doesn't. That's a common misconception though.
@@-TheBugLord it’s like the same concept, neural networks r based off brains
@@-TheBugLord yes it does. It aims to recreate the function of a brain
@@jtcoding6422 Yes. Based off, but does not work as our brains do. There are many differences. And, neuroscience isn't far enough to even determine how the brain actually works. But here's a simple example to show the differences. Neurons in neural networks are not like brain neurons at all. A brain's neuron is either on or off. A neuron in a neural network, however, can be on from any range between 0 an 1. You can have a 0.5 activated neuron in a neural network. Yes, it is true, neural networks are based on how the brain works. But to say they work as the brain does is a misconception and false. And our scientific knowledge is not far enough to even determine something like that yet.
This two-part video contains an impressive amount of getting sidetracked. Hilarious, I might add.
Watching AI fail at walking is an amazing and comforting sight. Laughing at real people falling feels bad because they get hurt, but watching an AI-controlled puppet fall in virtual 3D seems harmless
They call him code bullet because he's slow (like a bullet)
yes classic bullet behaviour
I think encouraging the chest to point at the target would help immensely. Also, wouldn't mind seeing running, and/or jumping if a part 3 is made. Maybe hands that have a stick to things and not stick to things state could be added for parkour? But, you did make walking AI, so if you are done, then so be it.
Nah, that's too artificial. Humans don't have a reward for keeping our chest pointed at things. What we have is a reward for looking at things (seeing where we're going) and a penalty (pain) for over-rotating our neck.
@@heliomance760 but, it would look good enough while introducing the fewest new rules.
Or maybe.. spawn them facing towards the Target?
@@cooperburtonburger1827 He is doing that..
16:12 And in the end, it didn't even matter...
HE TRIED SO HARD AND GOT SO FAR
Am I the only one that the audio is way ahead of the video?
Great video from start to finish, I especially liked the breakthrough and 18:16
That ending too, 10/10
Bro how you already watched it, it hasn’t even been out for 10 minutes
The VR part had me dying. The little quips were perfect. Great video.
Great video, I think the backwards walk stems from you spawning the guys in random orientations to the target while training. Most of the time they will be facing away from it and because they dont get rewarded for moving away from the target briefly they never turn, therefore developing the backwards walk. When you introduced the reward for looking at the target you continued with already trained models so they just turned their heads instead of developing something new. Might be interesting to see what happens when you retrain with them always starting facing the target, as knees only bend in one way and can be untilized differently when walking forward in comparison to backwards.
These two videos might be the funniest videos you have ever made. Cannot stop laughing every time you get submerged in all these player models in VR
As funny as VR shenanigans are, now I'd really like to see the chest-and-dick-facing-target thing with like 100h of training to really see what's possible.
With different weights on the reward-and-punishment values, too.
Maybe the average of the body parts facing the target could circumvent the issue entirely. But yeah it would be nice to se what actually can be done
Then they just bend over more, limbo style
Dude! Two videos in less than a week! Oh my GOD. That never happens. Looks like the AI that writes the script for the videos has been fine-tuned quite a bit...
I have a feeling you watch spiffing Brit
Beautiful. Incentives don't always work the way we want them to.
This is by far the best and funniest codebullet video to date!
Holy shit, codebullet actually released part 2 in a few days and not a few months. Impressive
17:22 this is how nature intended humans to walk
3:21 That flying dab tho!
23:40
How tf did it take him so long to realize
Honestly bullet, I really loved the way you did this video. Loved the editing, loved the way you through in the extra biat besides telling us what you're doing. Overall it was a new experience in the way you did your video. Would love to see more of this style. But I do alao enjoy your normal stuff too
Never again did I think I'd see the day where 2 code bullet videos get uploaded in a week
Never thought we'd be getting Code Bullet cake for Christmas
Nice touch with the @kickingmustang airsoft hits for the heel strikes at 22:25
Great video! Instead of positive reinforcement for facing the goal, you could also try making the AI feel “tired” from using their core or leading back, so they look for efficiency which may avoid the backwards walk
This man is making the line "AI is taking over the world" come into real life even faster.
next video: "AI learns to code"
can you stop commenting on everything its not fun
"AI taking over the world one step at a time"
3:55
According to the bots just below you... you're not wrong. Not by a long shot.
I can't wait for part 3 where he trains them to walk forwards in his new horror game.
4:00 is the hardest I've laughed in weeks
For the first time in Earth's history, codebullet actually released multiple videos in a single week. Good job man. Good job.
Now I want to see forward only learning for a week straight. Could we ever get relaxed gentleman walking or athlete running? What if you just train them to stand still for awhile, and then give them a target and train them to turn to face it with as little neck rotation as possible, more points for the closest to neck straight forward.
Also, give them points based on speed. Either they find a physics bug or we get the Usain bolt of AI
the reason humans walk efficently is because we actually expend energy
so maybe assign an energy consumption to each action
give them a determined amount of energy
see who goes the farthest
@@HurricaneLantern Wow, I wasn't even thinking of that. Excellent point!
@@HurricaneLantern Also pain. Crawling around on your elbows and back, with your head hitting the ground, would be ridiculously painful regardless of how energy efficient it is.
@@equidistanthoneyjoy7600 This sentence has me laughing like a madman 🤣
2 Code Bullet vids in a year this is insane
The same month even!
24:29 Caught me all the way off guard while I was drinking. Almost ruined my keyboard lmao
25:03 "The scariest part is the framerate." got me floored. That's one horror game enjoyer right there! Numb to the terrors surrounding him.
I'm just surprised we managed to get a part 2. If only there's another part 2 for a certain enigma
15:01 British soldiers bravely charging through no man's land into machinegun fire (1917)
AI is gonna kill me in the future for my laughing-my-ass-off during the *CRAKED* JoJo posing montage 😂
2 videos in 7 days? the AI must have taken over his channel
We want a part 3! Walking forwards and standing up after falling where it's reset if they spend more than 10-15 seconds on their back.
The clusterfuck that’s at 4:00 onwards looks like the fairly odd parents into when they’re turning Vickey into different things