AI Learns to steal
ฝัง
- เผยแพร่เมื่อ 18 ธ.ค. 2023
- Based on a true story!
Key Hyperparameters for those interested:
type: PPO
Batch Size: 512
Buffer Size: 4096
Hidden Units: 512
Hidden Layers: 4 (yep, 4 actually got me the best learning at a decent rate, typically I'd try with 2 or three layers first)
total training time:
approx 8.4 hours
20,800,000 steps
music:
1 - 368 - Dyalla
2 - All I Am - Dyalla
3 - Beyond - Patrick Patrikios
4 - TORSION - Density & Time
5 - AETHER - Density & Time
6 - Organic Guitar House - Dyalla
textures from:
PBR Material Bundle Vol1 - Cole
100+ PBR Materials Pack - Integrity Software & Games
AllSky - 220+ Sky / Skybox Set - RPGWHITELOCK
and some bad ones from me!
models from:
Low Poly Ultimate Pack - POLYPERFECT
Low Poly Disco Bar Pack - GOLS Consulting - วิทยาศาสตร์และเทคโนโลยี
![[Full Episode] MasterChef Junior Thailand มาสเตอร์เชฟ จูเนียร์ ประเทศไทย Season 3 Episode 2](http://i.ytimg.com/vi/YpQnHoQvZt4/mqdefault.jpg)
Imagine if a dude broke into your house, did sick flips for 3 minutes then left without saying a word.
I can imagine that
I would respond by doing nothing.
@@soup9242 popcorn
@@soup9242 I would respond by joining him because BINGUS those are sick flips
Sad to see Dingus fall on hard times.
And on stairs
And off of the stai-- oh you got to it first
Floors*
Seal :3
Dingus is not a thief, he is a GREAT ASSET to the company
Hopefully he doesn’t find any flower men.
weeeee looooove the company
Asset-Great great asset
great great asset great great asset
@@adora_was_taken WEEEEEE LOOOOVE THE COMPANY
I imagine Dingus was given no memory which explains why Dingus immediately moves on when objects fall out of view. With that in mind, going up the staircase and jumping off is actually a brilliant strategy because that path has more visual coverage than anywhere else in the room.
Dingus hasn't developed object permeance
AI learns to get around short term memory loss
I think it's more likely that he recognises moving valuables is rewarding, but does not realise he gets *more* for completing delivery.
I am on board with this theory @bitblit, also in supposing that there's rewarding going on for exploration. It seems that Dingus does not know where all the valuable items are from the start and must first locate them, which would suggest that any maneuver which drastically gathers information about where the valuable items are (such as eliminating the most space without valuable items) is rewarding enough to do repeatedly.
I'd be interested in seeing Dingus try to rob from completely different buildings with this same training.
I'd love to see him try and steal in even more environments, with uneven or unnatural terrain, or more hazards.
ye like guards
bro wants to see the world crumble before his eyes 💀🗿
maybe fire exits
Lasers on the floor
@@elizathegamer413 a few thumpers for good measure
His addiction to boulder rolling has led Dingus to a life of crime.
Maybe him liking pearls at start is ptsd of some sort
if i ever see 3 stairs i will now purposefully avoid the middle one at all times even if it's an easier path
9:13 Reminds me of video game speedrun brute forcers that find crazy tricks like this by testing every possible permutation of inputs. Very cool discovery.
iirc, this also happened in the Open AI hide and seek demo. The seekers learned to abuse the physics engine to launch themselves over obstacles, even though they only had directional inputs to work with.
I remember that vid, such a classic.
Sounds cool, what's the video?
I wonder if Dingus would be able to easily generalize to new maps, or if he's just learned how to beat these particular maps. Future video idea?
It's a good question, I'm curious myself, my guess would be he'd likely need to learn a wider pool of maps if new ones were substantially different.
But I have a trained model now so I can stick it into some new maps and see how it behaves
@@Dingus_Labs That might make a decent follow up video to this one. Just following the adventures of Dingus the Thief as he robs a liquor store of all its speakers.
@@Dingus_Labsoh, that's something I'd be really interested in seeing
@@Dingus_LabsI feel like you shouldve added a guard that also trained like dingus but catching dingus instead and the guard traine dsame rate as dingus and was named "Bingus"
I have a feeling the stairway loop has something to do with the reward system in some way. Maybe pushing down a valuable generates a high reward value as opposed to just nudging it like in the beginning, since it travels much closer to the car, thus leading Dingus to associate the stairway with that rewards spike.
I think your right... Is the moving items closer to the door working right? It looked like he would get close enough to git the reward and then run off 😅
could be right there, I tried to keep the "move the object towards the entrance" reward low so that it wouldn't overly impact later behaviours in training that got much higher rewards, but it may still have had a long term impact on his behaviours.
@@Dingus_Labs sounds like it's going to be a reset, go back a save, or remove the reward and penalty for moving it until you can bend hem back some. but most of the time unless it's part of the mission to keep the IA's memories intact it well take 5 or 6 times longer to unlearn something then restart sadly
@@Dingus_Labs maybe if you make the hitbox of the van for putting away the loot really tall and have dingus track the closest path that could solve it? Then the items descending due to gravity wouldn't generate as many points
If he struggles with this imagine this with guards
all future dinguses should be granted small rewards for doing sick flips from staircases because those really were some sick flips
I love the part where Dingus starts chugging wine and shoves the speaker as far as he can under the table. I can picture an actual incompetent thief getting wine drunk and violently proceed to force the electronics under furniture.
I love how dingus was literally just any player of a game goofy around at the start
I want to see one with two competing AIs training concurrently, like a cop AI vs a robber AI
I actually wonder how that'd go, huh
Whether they are even capable of adapting to each other
seeing how high he got id love to see a video that rewards him for height but has no obvious ways to do it too see if he can wall run consistantly
A good representation of how I learn how to play a new game.
Go dingus...
Dingus may be a idiot but he's our idiot. Never change dingus. Never change.
Honestly, I’m surprised it went so smoothly. My guess is that you left out a lot of the trial and error that I know goes into this-I’ve heard of people running simulations for literal days and coming out with an ai that was only slightly better than before.
There were 67 prior runs performed before I had a model that was reliable enough I was happy to make a video with!
I had to redesign reward functions and the actual gameplay a few times as part of that and tweak a lot of hyperparameters, I also for the first time needed to increase the number of hidden layers to 4 this time before I saw really good progress.
A number of my earlier attempts did yield an AI that could solve all levels, but not reliably enough for my liking!
hes struggling to resist the voices’ demands to climb up the stairs and flip off the balcony
Now AI are putting thieves out of work. A thought that simultaneously sounds great and scary.
So sad seeing dingus have to resort to this 😔
5:50
That laugh kills me 😂
next episode, DINGUS LEARNS HOW TO ESCAPE THE POLICE
Would love to see how he preforms in a new map he never got to train in
you should do a video where dingus has anxiety and randomly gets negative values for absolutely no reason and see what happens
I thought you were gonna convince an ai to make the moral decision to steal
Ai becomes romanian
Dingus was just checking if there was fall damage. Smart guy
So glad I watched your old videos before you removed them. You've improved a lot!
thank you!
The old vids are still available on a playlist for the channel, I just wanted some level of seperation because the new content is pretty different in format to the old content and felt them being intermingled could cause some confusion.
I think the reason he kept going down the stairs was because he got rewarded for going fast, and falling was the fastest way to move in his reality.
One could argue they're already very good at that
How would one get started training 3d models like this? To be clear, I have 3d skills in Blender, but besides that, I have literally zero clue how I would start. Any insight would be greatly appreciated!
it'd be interesting to see a dingus that combines all the previous ai's
This is awesome! Didn't think the agent would learn this quickly.
Also, thank you so much for including the mlagents hyperparameters in the description. Much appreciated 👌
is this why dingus was sent to the underworld
AI is already really good at stealing
10:40 Him talking to dingus feels like the narrator talking to Stanley.
Thought this was dank pods at first. Love the video bro
This is one of my favorite channels rn, im suprised its not larger, I love Dingus
good to see Romania putting some effort to expand the limits of ai
DINGUS I NEEDED YOU IN THESE TRYING TIMES
but the real crime was dingus stealing our hearts...
This is the only acceptable kind of AI learning, I would be ok if Dingus stole my job. He stole my heart already
Great video as always. Would love to get something like this setup and play around with it in multiplayer.
Could be interesting to see the AI having some basic understanding from the beginning, and then see how quickly it can learn. Like a tutorial or intro sequence for us mortalt. So for example a pre-recorded playthrough but with the AI still getting the rewards prior to the AI taking control, just so it got a "taste" of what works and what doesn't. A kickstart one might call it.
I give my life for glorious dingus nation
Average Romanian Training Regimen
i think he kept doing kick flips because he associated the reward not with the wine but instead with the drop off the stairs
I really like that he found a valuable item and was rewarded for, but coincidentally flipped off the second floor, and was forever convinced that flipping off the second floor was good, so he did it whenever he was lost
This is really underrated!!
this is like playing any games for first time
OMG HE'S BACK
Wow your acc is growing lol!! So excited to see another video
should give Dingus a friend
You should've told Dingus he was an artist, he would've figured out how to steal 50% faster
AI learns to borrow things extremely unethically
Did you maybe give it an award evertime it sees the object? That would ecplain the flips and circles.
Dingus really said "W I N E... F L I P S..." Within the first 20 minutes of their life
This and epic ai wars are now my go to places for all things ai entertainment
I love to see another installment in the dingus franchise!
I’m so proud of your AI son.
This man just made a lethal company bot
Imagine just being hired to just say "too dang slow Dingus!" every 200 seconds
Big fan of the videos, I think it’d be neat if you elaborated on how the ai works and/or the inputs dingus gets
Hmmmmmmmmmm, he's new but he's got it.
i feel the dingus lore growing
This guy's TH-cam channel is gonna be the entirety of our universe's terminator lore
So, an AI image generator, basically? Thank you, thank you, I'll be here all week
They call me dingus
ia learns how to comit planetary crimes
9:13 dingus the speedrunner
5:30 he probably thought that stair jump + item = reward lol. Ai is fascinsting
(reads title) it already does
More powerful than the will to win is the courage to begin.
Or the courage to be born black
Now he's so good with the basics would be amazing to see him take on bigger maps with security cameras/patrolling agents 😅
great watch, thank you for this video! I hope your sickness has completely left you
Childhood simulator with strict parents
This is the most fun AI video I've ever seen.
1:28LIES! we need dingus guitar
Thanks for the video! This data will be very useful for my new ai powered mech 😊
it would be cool if you did more levels that had guards for him to sneak past
Now you gotta train police AI to capture this thief
that ai is so quick it learns it absolutely no time. damn
The switching between the different Dinguses testing gives me major Aperture Science vibes
I would pay good money to see this entire video remade with a real life actor, although I can't help but think this is just what an actual burgler is doing as they're looting.
Crazy to see how dingus is initially trying random things, like going up the stairs, to try to get rewards - just like how if you put a human in a room with a bunch of stuff and told them to solve it, they would try random stuff.
Even more crazy to see is that once dingus is successful, it still unnecessarily goes up the stairs before depositing each item, because when it first ever deposited an item it went up stairs beforehand. So dingus pretty much created a ritual, seeing the mission to be to go up stairs and then deposit, instead of just depositing.
This might explain why humans created rituals for things that are out of our control (rain dances) or for things that have complicated patterns that are hard to see at first.
Would be cool to have table wines only accessible via doing sick flips.
Yingus is here. Sisyphus didn’t teach him a lesson yet
This man IS boosted with the ytb algorith but this level of excellent content i love it
DINGUS IS ADDICTED TO SICK FLIPS
I would love to see guards added
I kinda wonna see two of them try to carry sofa out
Dingus:try not to do sick flips challenge (impossible)
14:33 DONT WORRY HES JUST DOING THE BEST SITUPS EVER
Dingus the sick flipper
4:37 most relatable line of the video.
Would you be able to reuse the ai learning from another project to give Dingus a headstart on a challenge in the future? Like maybe the box pushing ai would have given Dingus a good idea of how to push the loot around his environment.
You can definitely build on top of existing models (in multiple ways), the main difficulty is that in order to reuse a model it needs to have the exact same neural net properties and use the exact same inputs and outputs in order to be compatible with the new environment, often times a new environment may require changes to one of these 3 things in order to perform optimally which means starting from scratch.
I think dingus's flops and struggles are important but may have ended up a little overemphasized here, I felt like I didn't hear much about what was actually going on with the numbers. it's also possible I'm just out of it rn so I'm way off the mark, but who knows. anyways dingus is my idol I love him hehehe
i love your channel.
This man wants a parkour course. Sick flips down the stairs are the best nlu way he can communicate it