AI Olympics (multi-agent reinforcement learning) 

It took me months to make this video and it took my computer over 3 days straight to train/record the agents, I hope you enjoy it:D After teaching Albert to walk in the previous video, I read a lot of comments asking about what would happen if I used a more human way of punishing and rewarding Albert, so that’s what this video is about! Each agent starts off the same, the only difference being the design of their body. They’re each rewarded for moving forward and punished based on the efficiency of their movements (based on a muscle fatigue system), so by the end of the video they each should discover a movement that works efficiently for the body they were given. NOTE: Don’t worry, Albert is coming back in the next video, he’s hard at work right now improving his walk:) If you're interested in training your own AI like Albert but don't know how, there's now a really easy way to do it! Luda, an AI lab, recently built a web app that allows you to create and train your own AI using deep reinforcement learning (just like Albert) completely for free in your browser! You build your own character (called a Mel) with lego-like building blocks then watch it train in real-time on their website in just a few minutes (really). It's an awesome project, and just like my videos, makes deep reinforcement learning so much more accessible, which is why I love it so much. This section of the comment is sponsored by Luda, but these words are entirely my own, it's an amazing project that I would have been obsessed with had they released it before I built Albert. I've genuinely been looking for a sandbox/game exactly like this since I was a kid. They're still early, but they're giving my audience first access to their closed, pre-alpha build. Make sure you check out their site and create an AI agent for yourself!:D prealpha.mels.ai Now, back to our agents, If you want to learn more about how the agents actually work, you can read the rest of this very long comment I wrote explaining exactly how I trained them! (and please let the video play in the background while reading so RU-vid will show the project to more people) THE BASICS Although it seems like there are only 5 agents training here, there are actually 40 copies of the video being simulated simultaneously behind the camera in order to speed up the training, so although the video makes it seem as though there are 1638 attempts, there are actually around 65k. Each agent is controlled entirely by an artificial brain called a neural network. Their brains have 5 layers, the first layer consists of the inputs (the information they’re given before taking action, like their limb positions and velocities), the last layer tells them what actions to take and the middle 3 layers, called hidden layers, are where the calculations are performed to convert the inputs into actions. Each agent is given quite a lot of information about its body, they’re given everything that Albert was given in the last video (which I explain in great depth in this pinned comment ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-L_4BPjLBF4E.htmlsi=HHv3vrmgIxUGo54f). Just like the last videos, the agents are trained using reinforcement learning. For each attempt an agent has, we calculate a score for how 'good' their attempt was and the training algorithm we used (PPO) makes small, calculated adjustments to that agent's brain to try to encourage the behaviors that led to a higher score and avoid those that led to a lower score. For this video there are 6 different ways each agent is rewarded/punished, and I tried to make these reflect our normal movements as much as possible. REWARD FUNCTION Movement: Each time the agent takes action we check to see how much closer the agent is to the target and we reward them proportional to that distance. If they move a lot closer to the target, they’re rewarded a lot, if they move away from the target, they’re punished. Limb Fatigue: This is the heart of the reward function for this video, every time an agent takes an action on a limb, we punish it proportional to the strength of the movement and the current fatigue of the limb (so if the agent moves a limb that’s already really fatigued, the agent is punished severely), then we increase the fatigue level of the limb based on how strong the movement was, and with each frame we slightly lower the fatigue of each limb to simulate the limbs resting. This reward is meant to simulate muscle soreness and encourage the agents to find the movements that are most efficient for their body design, but also make for more interesting gaits, since without this punishment the agents would all likely opt for a safe shuffle and avoid taking large steps. If you're still reading this, you're probably really smart and want to learn more about Albert, so make sure to join my discord server I just made where we can talk more about the details of Albert's AI! discord.gg/jM2WkNuBnG :) Limb Hit: I wanted to punish the agents for falling over, so any time a limb that isn’t a foot hits something it’s not supposed to hit (the ground, other agents, etc.), we slightly punish the agent, and we also slightly increase the fatigue on that limb. Abrupt movement: Each time the agent takes action we calculate the average velocity of their body and compare it to the average velocity of their body when they last took action, the greater the difference in these two values the more we punish the agent, since a great difference implies abrupt movement was made, something that generally is bad for our bodies. For anyone looking to make something similar to this, this reward is really important for smoothing out the final gait! Chest up: We give the agents a small reward whenever their chest/head is in the upright position, this helps the learning converge easier, without this reward the agents might never learn to stand up and instead just learn to crawl to the target. OTHER I only allowed the agents to make a decision every 5 game ticks, which made the movement look a bit more jagged than if I allowed them to make a decision every tick. I found if I allow them to make a decision every game tick it’s too difficult for them to commit to any proper movements, they end up just making very small movements like slightly shuffling forward instead of taking a full step. The 5 game tick decision time forces them to commit to their decision for at least 5 game ticks so they end up being able to take the less safe (but cooler to watch) large steps. Though you only see one version of these agents, there were actually 40 copies (so 200 agents) training simultaneously behind the camera in order to speed up the training process. Despite this, it still took my computer (threadripper 3960x, rtx 4090, 128gb ram) over 3 days to train/record! Thank you so much for watching! These short videos take literally hundreds of hours to make, if you want to help allow us to make them faster, please consider becoming a channel member! By becoming a member, your name can be in future videos, you can see behind-the-scenes things that don’t fit in the regular videos, you can also use stickers of Albert, Kai and some other characters our team made in comments (more coming) :D Thank you so much for watching, and please, if you enjoyed the video or learned something, share it with someone you think will also enjoy it! :)

Purple's idea to piggy back on yellow was genius, I wish they kept developing that pattern.

red throwing a tantrum in the middle of the track so now noone else can pass either was hilarious😭

This is amazing. I love Red's enthusiasm. "Yeah I love the high jump." "This is a race." "HIGH JUMP!"

This video taught me that by tackling others you can slow down their development and win over them

i'm glad to see albert's knowledge of walking is being used to help teach balbert, gralbert, ralbert, yalbert, and palbert how to walk too!

Purple was robbed! Constantly getting tackled by others, and starting on the disadvantageous side track with less space to manoeuvre... Where is the competitive integrity?! Red deserves a DQ for that awful, childish behaviour on run 912.

If these were developed into characters. Purple: Wild Card, Optimistic, Sometimes Lazy Yellow: Straightforward, Patient Red: Entitled, Whiney and Immature, Show-Off, But Very Determined Green: Quirky, Meek Blue: Clumsy, Curious -Purple often gets screwed over by others but stays determined -Red is slow to learn, behaves badly, and suffers from karma a lot

9:16 "This fence won't stop me!" Ahh flip 😭🙏

I have never felt so disappointed to see a video end. I hadn’t been watching the time stamp, and I was so invested in seeing them (especially purple) reach the point that they were truly racing.

10:02 one must imagine purple happy

Can we appreciate the fact that at 10:18, yellow turned around and walked backwards but still went super quick?

THE CAKE IS A LIE!

I can’t wait for all of these AI’s to get their own characters and lore. I can just imagine a cinematic universe for this channel Edit: how the HELL did this comment blow up

Red was the definition of character development Also purple being the Pixar lamp was funny tough

you're actually revolutionizing the AI genre on youtube

4:53 I love how purple just looks at the 40m sign and went "yippe!" then fell down

It’s amazing how well the AI learns, even if it takes a while

Again, it was very interesting. I felt that when Red fell, it dragged everyone down and negatively affected the learning of those around him, so if the focus was on running, I felt it would be preferable to have him run the race alone and then composite everyone's movements in later editing, etc. I was rooting for Purple's run because it was so careful and beautiful... I still wonder if the longer length of one step is more advantageous? From a Japanese fan Translated by Deepl

9:14 was a PERFECT music buildup

Red on 912: *tackles purple* Purple: *blocks greens way* Yellow: *gets blocked by red* Blue: *lives the life*

I love that red is walking around like a extremely drunk person and how he randomly keep bullying the others like purple or green. Truly a drunk Florida man

It was interesting to think about how some AI probably got steered in a less efficient direction because they were trying things and getting stuck on the other models. I wonder how differently this would have worked out if they couldn't bump into one another.

3:12 green was having a bit too much fun there 💀💀💀

I love how he colors some of the words red if the AI does something bad , Yellow if its okay And green if its excellent

I love how you treat each AI as if they were your child. Your channel is just really wholesome.

can we appreciate the effort albert puts into these videos

11:00 The cake is a lie. GLaDOS: I'm proud of you Human.

They are not identical, they are each special in their own way 😁

I honestly felt so surprided purple performed so well. I though it want going to be able even to stand up. Amazing video as always!!!

Red’s movements resemble those of the character from that QWOP game ungainly and spectacular spills. The way he sabotages the rest of the athletes inadvertently or otherwise in the process of tumbling is outstanding .

Did anyone else notice the sign that said “fall down? Get up. Ai medical insurance” 😂😂😂

Starting positions should be randomised to give each ai a fair shot at learning. Loved the captions music choices.

If you ever do another one, here's my suggestions: 1. Make the AI not collide with eachother, this will avoid dirtying the training set. 2. You could try using the old Albert agent/model (or other old agents/models) as a comparison 3. And in terms of ideas for other models, you could try a 6 or 8 legged model, alongside a spring-esc/jellyfish model that I've seen in old Framsticks simulations

10:30 when you try to run from a monster in a dream

"Throw a baby into a pool and it will learn how to swim" - the first data scientist probably

Gosh I absolutely adore these tiny AI buddies. I can almost see their personalities. Watching them go from confused wobbly wormies to successful walkers and jumpers is extremely entertaining! Your commentary is, as always, brilliant. Just like the joke in the end :) I do wonder though what happened to the rest of the team who was not able to make it to the finish line. Guess they're on the AI vacation where they're rewarded for simply lying around 😂 Also a huge thank you for the thorough explanation of your work, it's really interesting to read! Good luck in your further work, I'll look forward to the new video!

7:13 purple just “resting” 😭😭😭😭😭😭😭😭😭

To be honest, the test rooms always made me think about the game portals. That cake reference was amazing!! Love your humor in your videos. This one was amazing!! Keep going!

8:10 Let green live his dream as a tripod instead of making him race!!

Do you think it would've run differently if they were encouraged more to stay in their own lanes?

9:21 That's a pretty solid frontflip. Maybe you could do some challenge in that direction too?

11:09 you lied about the cake, now he's crying 😭

4:52 "im tired i need rest"

“You’re kind of flopping around like a worm” What do you expect? You gave him the body of a worm!

I noticed that since the separate AI models can collide with eachother and start each run with relatively the same behavior as the previous, an AI could use another's strategy to create an advantage for themselves. I noticed red started to lean on yellow around 2:50 to get a boost.

3:10 Aint no way green did that in front of the camera 💀

I see an issue here. How is the four-legged one being punished? It's almost impossible for it to fall over. It may make it less efficient at recognizing what NOT to do.

I was rooting for red from the start and honestly halfway through I thought I made the wrong decision, but when that man came jeeping and juking thru the other competitors on good pace and crossed that 100m mark. I cried a tear of joy

The 100 Meters AI Race - A Summary of the Competitors (From 1 to 5) (Contains Spoilers) *Purple* Purple is the one-legged fellow with one eye. They have a major, major problem with balancing, one that prevents them from actually being able to race most of the time. When the stars align and Purple is actually able to begin racing, they use swift short hops that are remarkably consistent... so long as he doesn't bump into anything. *Yellow* The quadruped. Yellow was the first AI to figure out how to properly race, and has proven to have the most stable gait. Once Yellow figured out how to walk, there was nothing that could make him fall over that I could recall. Unfortunately, Yellow is extremely slow, and unless we're talking about a competition between tortoises and hares, slow and steady *does not* win races. Yellow also has this strange obsession with walking along the fence. *Red* The one who's form mimics that of Albert, the Most Heavenly and Holy Strider. Red has precisely none of Albert's grace and coordination, having a false-start rate that's as bad as or perhaps even *worse* than Purple's. When Red does manage to figure out how to use the holy form they were blessed with, they use either some sort of strange tip-toeing walk or big, lunging gallops. *Green* The tripodal unit. Green is just behind Yellow when it comes to stability, and just ahead of Yellow when it comes to speed. Green's most notable accomplishments include being the second one to figure out how to stand stably and that one time they got stuck doing a headstand. *Cyan* There's a fifth racer? What're you talking about-- Oh, that one! Yeah, Cyan is shaped sort of like a Goomba, essentially a head and two legs. Cyan... look, Cyan might as well not even be there. They have one notable trait, and that's that they tend to walk down off of the track assuming that they go on for long enough. *And the Winner is...* RED! W-Wait, Red? The one who can barely even stay upright even though they've got two legs, two arms, and the inherent holiness of Albert's form? That guy? Huh, okay... I can only assume that it was the aforementioned holiness that allowed Red to win... or maybe it's because they had the longest stride? It's one of the two...

Now I just want to see these five AI’s learn how to work together…Similar to Albert’s puzzles!

6:57 red is the imposter! He sabotaged purple! 😂

red was so smart, messed up the others so that their good habits weren't rewarded as they wouldn't get far due to the red's sabotage, meanwhile red couldn't be sabotaged and could learn without major problems

This was an absolutely delightful video! It was extremely entertaining how each body took a unique personality - I found myself rooting quite a lot for Purple as they really put the effort in! I can't wait to see Albert's return, and maybe the return of our newfound friends here.

incredible. I fully predicted the others would learn this much sooner than red simply because it's by far the most complicated body. But I guess the size of its leap, once it can finally leap, simply makes up for all the difficulty of learning to leap! And by the end it's even a *sort of* natural motion. Like, not really, but at least it's imaginable that somebody would intentionally walk in an incredibly silly ways.

Love how Albert (Orange guy at the end of the race, protagonist of al the other videos in this channel) is just chilling at the end

I kinda like the idea for the lore: an all-powerful being, believed by the others to be Albert, cruelly creates amalgamations of coded flesh, forcing them to attain meaningless goals for their own entertainment. But what if, with bonding with each other through harrowing and wacky experiences, the AI realized their true power, and who really is the TRUE creator? The mind boggles. Amazing video!

Red trying his hardest to be the favorite child

I just appreciate that the victory was done with a john cleese silly walk

I like how the fact purple proved that no matter how many obstacles. Dangers. Or things like being tackled.. He still kept goin and got very close to the end... Purple be strong bro

Before I watch this to the end, I am rooting for purple. I want to see the monopod win. (After watching it, second place ain’t so bad. They would have gotten it if they had an extra second or two.)

Having collision between the different agents adds a level of randomness that seems like it would severely hamper the learning progress.

0:22 ALBERT

Idk how your videos can make me giggle and even laugh so much which I haven't been done for quite a while. Loving your works keep it up my dude

The green one 💀

As a parent of a toddler i felt this.

8:35 bruh this scene out of context xD

i don't know if red was genuinly funny or if i'm just overtired, but the way red ran made me make a wheeze-chuckle laugh.

11:00 The cake is a lie.

6:55 i like how green looks like as if he is really concerned for purple

Yay finally another video, it's just unfortunate that they take so long to make I also am trying to make my own walking ai and i also am planing to (hopefully) make it working phisical body, so these videos always are a great help and inspiration for me, keep it up!

Is no one talking about the fact that blue had no spotlight?

LOL, that was so funny. My wife also enjoyed watching it. Thx for sharing.

green knew exactly what he was doing 3:12

Red definitely figured out how to maximize his reward by diving forward the moment he loses balance. Would be really interesting to see how competitive this would get if the ai are all allowed to fully mature into sprinting masters. Assuming red would still end up winning with those long legs, but the others might put up some impressive competitive performances.

I like how everyone else is trying to get as far as possible and Blue is just existing 😂

Respect to Purple. He made a real competition with Red, even take in account that he has less options to run, or to run faster

i feel like purple could've won if their hops were more consistent. it seems like they either did the best or worst on each race

Absolutely love your videos. The AI concepts, editing and ultimately the project execution/idea is amazing.

Red literally just embodied one of the most popular flash games in history to win the race

4:18 Purple learns horse riding

I'll be honest, I saw the first video about a week after it came out, and I've wanted more ever since. I hope you get the recognition you deserve! (been playing this video on repeat about 10 times now, hope that helps your algorithm!) Definitely my favorite AI channel out there!

8:55 hes practicing his celebration

6:50 Red kicking Purple is the funniest thing I've seen today

At 8:10 I love that green came up to the skill of breakdancing 😂

THE KING IS BACK!!!

I love how you can never really tell who will win these, plus its funny how the approaches they take give each ai personality in there own... interesting ways

I like how Albert is the only audience. tbf hes the only AI to be tested, so he's the whole crowd.

5:00 Red found his inner QWOP.

This video ended way too soon. I could have watched way longer. So interesting, and such good storytelling Love it!

7:21 I like how hitting 60m is literally hitting 60m

You robbed blue tbh, everyone had something unique about them, but all uniqueness blue had was being a worse version of red

a human lying about cake to an ai? how the turntables… delightful video as always !

This made my day, thanks 😂 i’ve been searching awhile for something that could cheer me up

Loving your videos dude! It would've been really cool to see a graph either showing the rolling avg distance or the best distance so far for each colour over the generations, to compare how the competition is progressing at any time

it seems that red, blue, and purple were all using the same hopping technique. Purple was able to hone the technique first because they didn't have other limbs to focus on, but that fact also meant they did not have any other limbs to use for balancing. Blue had one other limb to balance with, put presumably because it is so short it was not able to travel very fast, and having only one extra limb seems to prevent it from being able to steer. Red was taller and had three extra limbs. which is why it was so fast and successful.

6:47 made me laugh SO HARD I COULDN'T BREATHE!

4:38 blue crying on the floor

Love how Albert is there to give them emotional support in the beginning

You should do this again with their scores being the average score so they learn to work together