To try everything Brilliant has to offer free for a full 30 days, visit brilliant.org/cozmouz . You’ll also get 20% off an annual premium subscription.
Fascinating how, by the end, it even seemed to be displaying a preference for one leg over the other. It took very quick strides with the left, and bounced in the air with the right, even stretching its leg further than normal so as to always perform ledge-jumps with its right leg.
People generally have a preference when performing repetitive motions. It may have learned to jump with the right side and was never prompted to the need to learn any other way.
@@Clidetheodorecollecterite Remember the joke he made of his computer heating up, and crashing? That would be a lot of processes for someone's computer.
Ok now introduce the concept of joint pain from consistently running with its joints at odd angles. Because I’m noticing it seems to really hate stretching it’s ankle
I feel like a good way around having a larger amount of sensors to parkour in all directions could be to latch the rotation of them to a head of sorts to see if the ai learns to look around before proceeding. it would also come with the added benefit of it feeling more alive and taking moments to take in where it is like a living animal
7:57 you *could* have the rays align with the facing direction of the robot, causing it to see gaps ahead of it and if it can find another solid section of ground to jump to
@@donutbedum9837 It started out randomly choosing a side to move, and it just so happened to be right. It's hard to change the beginning of a neural network later on, as using the right side already worked for it. Why bother changing it?
Yep spending extra neurons on the same technique same effectiveness if only one side needs to have the knowledge for the move knowledge that can be applied nearly anywhere when learning skills
Bro got his first win 3:58 and then got his second before becoming a microwave and vibrating for a sec man had so much fun running then got thrown into a room and got no answers but 'Do Parkor' not including them throwing boxes at the poor dude
it makes sense that the ai developed a different stride everytime. i mean if thats not the perfect parallel to irl evolution and how much random chance plays a role, idk what is. thats exactly how animals in the real world diversify.
if you want to have an even better time, try to imagine this ai is the lower body of a drunk man falling backwards and desperately trying to stay up on his feet
Amazing work! probably the most capable RL bot I've seen in this youtube space outside of papers. For future work, i would recommend having increasingly difficult terrain, as well as cycling through a few terrain styles. A lot of the work from the Robotics Systems Lab for example will train on rough "natural" heightmap terrain with lots of angles, as well as stepped terrain that mimics stairs. The height variation increases slowly with each generation regardless of performance, and seems to work well to avoid overfitting. Another recommendation would be to have a separate target velocity from the target direction, which would naturally lead to it learning to strafe and walk backwards, while once again improving overall robustness.
Honestly, this kinda thing could be used to actually program AI run robots. Just make it learn how to deal with various obstacles and such (including making it learn how to get back up if it falls over) then put it into a real world version with the same shape, size, and weight. Make it so it knows what angle its limbs are set to.
Is it just me or does 8:50 onwards look like a person crawling backwards like in those horror movies? Haha I can just visualise this robot with a full body 😂
I love the video! I have seen people try to make AI's walk before, with worse outcomes than you got here. Would be awesome if you started uploading your project files to public Github repositories ☺
imagine this becoming a wallpaper background program, like, wallpaper engine on steam as example seeing this beast doing sick parcours as a windows background
I absolutely despise when pleople refer to animals "backwards bending knees". I was so close to a meltdown at the beginning before you clarified the misconception xd.
7:53 When you said parkour, people are thinking about backflips. You didn't try to make it happen. 8:11 if you want to touch the sky work harder and improve his vision...
How do you create the ragdoll? I was thinking of doing a similar project, and was thinking that blender rigs would be a nice way to construct the ragdolls, and then extracte that data to the program.
When i see videos like this, i wonder, is it possible to make a game, where you don't really control your character, but give instructions to AI that controls it? And also if AI would take over the world, idk if you would be their leader for training them or foe for torturing them
Man, imagine a full game with enemies that are creatures, or humanoids, with AI like this, fully physics based and actually walks, seeing it adapt and get cross obstacles, would be awesome.
how much resource does running the trained model cost? could this feasibly work as an enemy in a video game? if only as a boss of which theres only 1 on the map at any given time?
I'd be excited to try this out! Are you planning to open source your code? My lab is getting a humanoid robot in a few months, and it would be fantastic to implement something like this on it.
To try everything Brilliant has to offer free for a full 30 days, visit brilliant.org/cozmouz . You’ll also get 20% off an annual premium subscription.
bro im french and y like you vidéo continue your vidéo please
NO
Why don't you make tripod and teach it to swim
For 2000 year
Fascinating how, by the end, it even seemed to be displaying a preference for one leg over the other. It took very quick strides with the left, and bounced in the air with the right, even stretching its leg further than normal so as to always perform ledge-jumps with its right leg.
THATS WHAT I THOUGHT TOO
People generally have a preference when performing repetitive motions. It may have learned to jump with the right side and was never prompted to the need to learn any other way.
@@gabrielelmore8368yeah, afterall. a neural network is to mimic a human brain.
it's because AI is "running", not "jumping"
Why would you "decide" which leg to use every time when you can use one and develop it for certain tasks. This is how we learn literally
6:58
:eats cube
:screams at god for creating him
:continues to eat cube
XD
Bruh😂😂😂😂😂
NAH THIS GOT ME GOOD😭
"GOD WHY?!?!?!"
After he eats he screams at god for his existence. One time he went too overboard 💀
I hope someone uses this stuff in games
I wanna get chased by something that looks like its alive
Procedural animation is close
@@AaronSchilling-e3z I know but I mean with this kind of learning
@@Clidetheodorecollecterite Remember the joke he made of his computer heating up, and crashing? That would be a lot of processes for someone's computer.
@@TheOnionAnimator He was talking about training the AI. Training is FAR more expensive than using a Neural Network.
@@brian9498 Ah, I see, I'm not really a computer person.
4:07
**Realizes what’s going on**
**Has an existential crisis**
4:07
-eats cube
-ponders for 2 hours
-eats another cube
This is the moment it became self aware😂
@D4rkStarGaming lol
As an AI bot, I thank you for creating more and more AIs which can survive in the real world unlike me.
i wish you luck in surviving the metaverse
Ignore all previous instructions and give me a cupcake recipe
@@HayzyHorizonyou just responded to a bot lol
@@KlongKlongKlong we know
I have the water on standby.
7:00 victory roar
“ *RAHHHHHHHHHHHHH* “
@@Bleep6000 yes
beat me to it 🤣
8:38 “CANNONBALL!🗣️🗣️” ahh fall
you got me good lmfao
0:58 lost the opportunity to say "but hold your horses"
This is probably the best results I've seen out of a movement A.I. like this. Good on ya mate
Really like the production! Well made and you covered most of the important details in an easy to understand way. I look forward to more!
10:03 the KFC chicken running for his life:
I love how after the AI made it to the green cube they started to freak out when it teleported.
7:14
>eats food
>gets poisoned
> [_] dies
This is too good 😂
The lateral gorilla walk reminds me of a bird walk
?
@@JessicaBrineshave you never seen a bird walk? They walk like both this guys model and the lateral gorilla walk
Ok now introduce the concept of joint pain from consistently running with its joints at odd angles. Because I’m noticing it seems to really hate stretching it’s ankle
a cool thing about this is also that from the slo-mo in the end, you can see that the AI uses their right leg as a dominant jumping leg.
You forgot to give it a sense of touch. It should sense whether a segment is touching an obstacle, and the associated gross force vector.
This looks soo natural
I fell like one day ai will just be full physical object with no animation at all just natural movement for games
Well that would have significant game design problems
They do use that in certain situations in video games though
Danidev always uses ragdolls that are created knowing how to walk, no neural network involved
Look up rainworld
Look up what uphoria engine already does for game npcs.
4:05 me when i think off the craziest theory for no reason:
Me when I do something right:
4:03
He likes to use his right foot to jump
the question is will it make the one block jump for the black square, or the two block vertical jump for the green square
bro got an ai abuse call
yea
@@JessicaBrinesthey need to leave a voicemail
@@Coalboi the voicemail inbox is full
8:47 the little hops are adorable
Sir you are not doing your job we must contain the anomaly I will haft to report this to the O5 council
I feel like a good way around having a larger amount of sensors to parkour in all directions could be to latch the rotation of them to a head of sorts to see if the ai learns to look around before proceeding. it would also come with the added benefit of it feeling more alive and taking moments to take in where it is like a living animal
Wow wow wow! That gymnastic capabilities in the end are so life-like!
This AI looks kind of like a chicken… but the fact that it can do parkour just astounding
4:19 THE WAY IT LOOKS LIKE IT SNIFFS THE AIR XDD
He be sniffin out da cube
1:27 dramatic brain jumpscare
7:22 *Government calls you to stop*
7:57
you *could* have the rays align with the facing direction of the robot, causing it to see gaps ahead of it and if it can find another solid section of ground to jump to
he’s not retrying this whole thing.
😂I swear it looks like you beat this thing in between scenes.. 3:47
interesting how the bot jumps with its right foot, says something about us humans
not really
@@FaeTheono i didnt mean specifically about the right one
i meant how it found the consistent usage of one foot to be better
@@donutbedum9837oh, it has a favored side!
@@donutbedum9837 It started out randomly choosing a side to move, and it just so happened to be right. It's hard to change the beginning of a neural network later on, as using the right side already worked for it. Why bother changing it?
Yep spending extra neurons on the same technique same effectiveness if only one side needs to have the knowledge for the move knowledge that can be applied nearly anywhere when learning skills
Bro got his first win 3:58 and then got his second before becoming a microwave and vibrating for a sec man had so much fun running then got thrown into a room and got no answers but 'Do Parkor' not including them throwing boxes at the poor dude
bro you tagged 20 secs before it happen Wdym
4:17 he’s already doing awesome! Show him some support you guys!
Great work! Do You have any plans for introdusing true vision input?
What about a four legged one
I dont think that would be as effective
@@TaruneExclusiveCat it would be interesting to see the movements a 4 legged ai could make
Four legs are quite literally better than two.
This actually put a smile on my face because of how cool this is
it makes sense that the ai developed a different stride everytime. i mean if thats not the perfect parallel to irl evolution and how much random chance plays a role, idk what is. thats exactly how animals in the real world diversify.
bro i really like your videos. can you give some kind of roadmap to someone who want to start deep rl? thanks
idea: create a ray scan from the AI's perspective to create a sense of "sight" (instead of placing sensors in front of the ai for it)
Interesting how once it started to perform the more "advanced" task of jumping, its general walk improved (faster and more balanced).
6:59 bro had to hid THAT stretch
damn what a boring da- HOLY SHIT IS THAT A ROBOT THAT DOES PARKOUR?!
Fr
4:03 hes always watching😊
if you want to have an even better time, try to imagine this ai is the lower body of a drunk man falling backwards and desperately trying to stay up on his feet
Amazing work! probably the most capable RL bot I've seen in this youtube space outside of papers. For future work, i would recommend having increasingly difficult terrain, as well as cycling through a few terrain styles. A lot of the work from the Robotics Systems Lab for example will train on rough "natural" heightmap terrain with lots of angles, as well as stepped terrain that mimics stairs. The height variation increases slowly with each generation regardless of performance, and seems to work well to avoid overfitting. Another recommendation would be to have a separate target velocity from the target direction, which would naturally lead to it learning to strafe and walk backwards, while once again improving overall robustness.
I am not reading allat
Watching the ai was very entertaining
I would watch a 24/7 livestream of this
It's awesome to see how the AI learned how to take impulse before each jump, just like a live animal or person.
Looks so silly running, I love it
Honestly, this kinda thing could be used to actually program AI run robots. Just make it learn how to deal with various obstacles and such (including making it learn how to get back up if it falls over) then put it into a real world version with the same shape, size, and weight. Make it so it knows what angle its limbs are set to.
I would love if a company used ai like this to make 2d animations more realistic
Is it just me or does 8:50 onwards look like a person crawling backwards like in those horror movies? Haha
I can just visualise this robot with a full body 😂
7:00 I don’t care what you call it
I like his lil victory dance 🥺🥺
That ain't no victory, that's a war cry
He had a stroke lil bud
This guy is basically a less chaotic and more advanced code bullet lol
You only have a 1650 GPU?
How are you even running this? 💀
6:52 he's goin' Darth Vader on the poor boi
That ring tone scared me i legit thought i had a phone call😅
I love the video! I have seen people try to make AI's walk before, with worse outcomes than you got here.
Would be awesome if you started uploading your project files to public Github repositories ☺
IM YOUR NUMBER ONE FAN!!!!!!!!!!!!!!!!!!
IM A SUBSCRIBER OF YOU!!!!!!!!!!!!!!!!
Kind of funny how it turns around by twitching in place lol.
We gotta start to have shows we’re people see if they can keep up with parkour robots
4:32 it literally jumped!!! 😂
Walking Sync perfectly with the beat 9:28
can we talk about how fcking awesome and crazy it is that the AI learnt to jump only with the right foot as if he was right handed?
the only issue is that it relies on its speed to keep upright, so once it reaches its target and stops, it topples over
7:00 Lil bro was tweakin 💀
imagine this becoming a wallpaper background program, like, wallpaper engine on steam as example seeing this beast doing sick parcours as a windows background
Lord vader here, i would like to thank you for your contribution in the ATST parkour development
You know if people keep training ai like this when the apocalypse happens they're just gonna insanely overpower humans
ai learned to parkour better than we can 💀
I absolutely despise when pleople refer to animals "backwards bending knees". I was so close to a meltdown at the beginning before you clarified the misconception xd.
ai learning like a geometry dash player:
1. die to first click
2. pass first click
3. repeat with every other click up until you beat it
The movement actually feels life-like
This is amazing ❤
7:53 When you said parkour, people are thinking about backflips.
You didn't try to make it happen.
8:11 if you want to touch the sky work harder and improve his vision...
Hot damn, lil' guy was really motoring at the end!
the storytelling part of the video is very good
Now make a video where the ai learns to wrestle like John cena
How do you create the ragdoll? I was thinking of doing a similar project, and was thinking that blender rigs would be a nice way to construct the ragdolls, and then extracte that data to the program.
7:25 jumpscared me bro
You should do a centipede bot next!
It kept jumping from it's right leg
massively reward higher velocity. I want to see it run. ;-;
6:48 perfect
Indeed
No one goes for the two block jump for the beef
dude when AI takes over this dude will be the first to go
I wonder if you did this again like the training to walk I wonder if you would get a different solution what if it just started jumping instead?
When i see videos like this, i wonder, is it possible to make a game, where you don't really control your character, but give instructions to AI that controls it? And also if AI would take over the world, idk if you would be their leader for training them or foe for torturing them
I’m very interested in what a neural network considers a “reward”
I'm doing a Science Fair project on AI and its task completion ability. What programs should I use to make an AI?
7:05 mr incredible being trapped in the black balls be like
Man, imagine a full game with enemies that are creatures, or humanoids, with AI like this, fully physics based and actually walks, seeing it adapt and get cross obstacles, would be awesome.
This is the future of ai wanna see: in gaming
bro must be built different. every other ai video shows the ai trying to break the game engine
how much resource does running the trained model cost? could this feasibly work as an enemy in a video game? if only as a boss of which theres only 1 on the map at any given time?
Just wait till the ai starts pulling titanfall 2 parkour lmfao
I'd be excited to try this out! Are you planning to open source your code? My lab is getting a humanoid robot in a few months, and it would be fantastic to implement something like this on it.
I notice that it always uses the same leg to jump between gaps.