Building the AI of F.E.A.R. with Goal Oriented Action Planning | AI 101
HTML-код
- Опубликовано: 19 июн 2024
- In this episode of AI 101 we celebrate AI and Games reaching 100,000 subscribers as I return to the technique - and the game - where this channel began. A popular alternative to Behaviour Trees, Goal Oriented Action Planning (GOAP) was popularised by the AI of F.E.A.R. and has been used in numerous games since.
But how does it work and where did it come from? We explore its origins and go into even more detail on its application within F.E.A.R.
Chapters
[00:00] Introduction
[01:11] Automated Planning
[05:26] Goal Oriented Action Planning
[09:01] The AI of F.E.A.R.
[15:50] Closing
Some additional assets to check out if you're interested in GOAP and FEAR:
Holistic 3D's video on GOAP
• An Introduction to Goa...
Jeff Orkin's GOAP website when at MIT, including papers, presentations and other resources
alumni.media.mit.edu/~jorkin/g...
GOAP: 10 Years of AI Planning (GDC 2015)
• Goal-Oriented Action P...
Three States and a Plan: The AI of F.E.A.R., Jeff Orkin, 2006
alumni.media.mit.edu/~jorkin/g...
Game AI Planning Analytics: The Case of Three First Person Shooters. Eric Jacopin, 2014.
www.aaai.org/ocs/index.php/AI...
--
AI 101 is a series that looks at the tools and techniques used in video game development.
AI and Games is a RUclips series on research and applications of Artificial Intelligence in video games. It's supported through and wouldn't be possible without the wonderful people who support it on Patreon, plus RUclips members and Paypal donations.
/ ai_and_games
/ @aiandgames
www.paypal.me/AIandGames
--
Get yourself an AI and Games t-shirt over on Teespring!
teespring.com/stores/aiandgames
You can follow AI and Games (and me) on Facebook and Twitter:
/ aiandgames
/ aiandgames
/ get_tuda_choppa
#FEAR #GameDev #FEARgame
The AI of F.E.A.R. is still held in high regard to this day. It is cunning, responsive, reactive and vicious. It was the driving force that resulted in the very first video on this channel. Now 6 years and 100,000 subscribers later, I figured why not celebrate by returning to where it all started. I hope you all enjoy it and maybe learn a thing or two along the way.
This is my thank you to all who have watched, liked, subscribed and shared the blogs and videos from this show over the years: to revisit a game I covered before in even more detail than I did previously - and deliver it in the style that this channel has slowly developed. Your continued support has been a huge motivator in pushing me to make more content, but is also the driving force that helped this channel grow; becoming increasingly more recognised within different corners of the gaming community. It has been incredibly rewarding to see the channels impact as I receive emails, tweets and comments providing feedback and support from aspiring indie devs all the way to AAA developers from all over the world. My goal for each and every video is for viewers to come away having learned something - and hopefully be entertained along the way. It's a fresh challenge with every new video and success is not guaranteed as every new viewer takes a chance and clicks on the thumbnail. It might succeed hundreds if not thousands of times, it might only succeed once, but if it helps educate and inspire others, then that once was enough.
And of course, a special thank you to my Patreon supporters, who fund the development of these videos and enable me to dedicate the time and energy necessary to do this work as best I can! Each episode can take anywhere from days to weeks to research, record, edit and release, with many episodes - including this one - going through numerous edits. I appreciate for some talking money may be crass, but if I'm going to make a video, I'm going to make it the very best I can - and that comes at a cost. Without the financial support of my Patreon community, AI and Games would have shut down many years ago.
Once again, my sincere thank you. I hope you continue to stick around for future videos and wish you all the very best.
- Tommy
(the Scottish voice in the video)
Absolutely tremendous work on this video, Tommy. Explanation is clear and easily understood and makes me appreciate FEAR even that much more 15 years later.
Why exactly was this AI hardly ever used again? Or at least spoken about more. Then again, for all I know its lurking in the background.
The AI in FEAR were a gem and it feels really like nothing has come close
I've read Monolith's witepaper about this technic and they making jokes in it about how F.E.A.R AI is simillar to Half-Life specforces behaviour and how it's a shame that game which came out 5 years after Half-Life expose same behaviour, but people like it.
@@MegaSmouke where can you find this whitepaper? I'd be really interested in reading it
The A.I. in the Stalker games is really good, especially modded (Complete Mod or Anomaly)
trepang 2:
Better and more responsive AI is what I am looking forward to in more next gen games. Imagine a nemesis system from shadow of morder that remembers the tactics you use against them and learns to counter your behavior. Would be pretty sick
The knowledge that none of the NPCs are aware of each other's existence makes me sad for some reason. It's like a weird episode of the Twilight Zone.
Sad at first, but it's like the paper shredder not knowing about the fax machine (wow this is dated). Take solace in the fact that they are blissfully unaware of each others existence and in their ignorance, are happily carrying out any action provided to them. I like to think this gives them a feeling of purpose :)
Kinda ironic given that they're also supposed to be part of a hivemind in the game
Another sad part is that they're not supposed to live more than, let's say, 5 seconds on screen. Hence the plans with only 1-2 goals and just a few more actions, like T.T. explained. Short life. Filled with purpose, indeed, but still a short life.
I'm binging AI and Games as well as New Frame Plus, hoping to become a one-man-army game developer lol
Chad Gamedev ?
Lol, best of luck to you soldier
Take a look a Hakita& his game UltraKill
It's worth a look
A new hope
U just wasting time
F.E.A.R. Is simply one of the best FPS of all time. I *love* it.
It certainly is my friend, it certainly is.
Considering that most video game protagonists can open doors without using their hands, I'd say a telekinetic AI that can open doors from afar isn't too far removed from what usually happens in a game.
I mean yeah if you wanna get technical pretty much every single AI is telekinetic. It just has a veeeeeeeeery limited range (if you've fixed all the bugs). :D
I really appreciate your channel delivering technical explanation on higher level concepts, rather than catering to the large population of gamers with no programming experience. Almost all gaming video-essay creators produce videos for fans of video-games, but you deliver videos for game developers and people trained in tech, who understand computer science concepts but want to learn how they are implemented in video games.
I can't understate how refreshing it is to have a game development channel focus on complex ideas. Imagine other game-dev youtubers casually throwing out 'finite state machines' and using them to motivate a game mechanic, it'd never happen.
Thanks for the kind words! Though I will say I am only now happy throwing around terms like Finite State Machines because I already made an entire episode on that. The channel has now ran for so long it's massively self-referential. Having episodes I can point to about 'that thing' so I don't have to explain it for the 10th time is pretty much what AI 101 is designed to achieve.
@@AIandGames Speaking of self-reference I'd love to see a bit more of structured guidance on your content. I mean, I just got to the channel and am already loving it, but also I don't know what to watch (or HOW to watch it). This video for instance is in the top of the 'AI 101' playlist but clearly is not the first of the series (is it the 5th, 6th? I can't tell). On that note, I appreciate your timestamping chapters and the addicional links on the video descriptions. Btw congrats on the milestone.
Something I noticed in your first part (you may have talked about this later in the video); The NPC needs to be able to read a state before a plan can be formulated. An example of why this needs to be so; if the door is open at runtime, the player can close it. This introduces false information, which can cause the AI to get stuck. Generally, the best practice for this is, instead of developing a state-based script, it would probably be easier or better for it to be a stateless script. This is a script where each state in the script is checked as the AI works its' way through it. For example; If the AI plans to jump over a table, go through a door and jump through a window on the other side, it would read the state only as far as the next objective. If "true" proceed directly to the next one after. If false, change the state to true, then proceed.
God, I really need to reinstall FEAR and play through it now that I'm seeing this episode... I've been meaning to do that for a while.
Found this video because of Civvie11's recent video on F.E.A.R. A fantastic explanation of what is still some of the finest AI in video games. Great job!
Pointman Insertion.
"there are just shy of 70 goals implemented in the code"
...
...
Nice
Took me a second, heh
I still await the STALKER A-Life video, but damn, good video mate. Congrats on the achievement!
for what it's worth, the stalker A-life systme is just goap mixed in with toggling between full-ai when the player is close, and a simplified sim when the player is far away. It's actually not very complicated on an abstract level, it just has to deal with alot of random shit happening in the world.
And it comes really handy with STALKER 2 on the next corner.
I'm really glad to see your channel come so far. Back in college your long form videos of your classrooms were like a hidden treat that provided information that no one else did. Now you've got proper editing, audio and gameplay combined with your professional commentary. There's no one who does what you do like you do.
Keep doing what you're doing Tommy. You're one of a kind.
On a side note, I must say its really interesting that you started with FEAR and now here we are, 5 years later, and still talking about it. Says alot about both FEAR as a game and the industry as a whole.
Thanks for your kind words. It's funny when I speak with some of my peers in the game AI research community as we talk a lot about how to get the message out and deliver more 'pop science' explanations of the work being done. I get a lot of people commenting on how much work it was to get the one or two videos done that they made and how hard it is. I recall making the first episode without having any idea how to make a video. I'd never done it before and had no desire to do so. It was so much work and I was so proud of it back then. Now it makes me despair! :D It's interesting for me to see the growth of the channel, as painful as it can be. I see and hear every mistake in the edit and try to stop that happening again. One step at a time!
Fun fact: Due to being made around the same time, Condemned: Criminal Origins (another First Person Horror game) uses the same GOAP AI as F.E.A.R. 1 & 2.
I also think that Wolfenstein: The New Order and The Old Blood used GOAP AI as well.
So the Wolfenstein games use hierarchical finite state machines, as do all games running on the id Tech engine. I talk about this at length over in the video on DOOM (2016), given many of the AI systems in that game are adapted from Rage.
@@AIandGames Oh. Still great AI, though!
No. I thank you for making such awesome content that deserves much more than just 100k subs.
It should be noted; the Lord of the Rings games do NOT use G.O.A.P. Instead, they use a very similiar version of it, often referred to as M.O.A.P. This is a system capable of long-term planning (upwards of 30 actions in a sequence, and over 500 actions overall). It plans a sequence of actions; if the enemy achieves their goal, the overall state reverts to what it was previously. If it fails in its' goal, the reasons for the failure are considered, and applied to creating the special enemies. The players actions in those sequences are taken into account, whether the AI succeeds or fails, and applied to future encounters; with the characters even referencing them, sometimes.
One of the most memorable video games in my life is FEAR, a first-person shooter horror game that came out in 2005. I was a freshmen in high school when I built my first custom built PC, and FEAR was the first game I installed on it. I was so excited to play it, because I had heard great things about its graphics, gameplay and story. I remember being blown away by the realistic lighting, shadows and physics, the intense combat and the creepy atmosphere. FEAR was not just a game, it was an experience. I felt like I was part of a movie, a thrilling and scary one. I loved every minute of it, even when I was scared out of my wits by the supernatural enemies and the mysterious girl Alma. FEAR is nostalgic for me because it reminds me of a time when I discovered the joy of PC gaming, and when I was amazed by the possibilities of technology. FEAR is more than just a game, it is a part of my history.
Thank you AI and Games for doing F.E.A.R. Is one of my favorite games that, in my opinion, have the best AI. Especially during their time. The AI still gets me sometimes to this day. Love your videos of gaming AI (since I mostly play single player games all my life) and loved playing against bots instead of players.
Glad you enjoy it! Stick around, I have some fun stuff planned for the next couple of months.
Not many games have stories that keep me coming back for another immediate playthrough. But FEAR is definitely one of them. I've played the campaign dozens of times and still never got bored with it. While I enjoyed the creepy story itself, the combat is definitely what made FEAR's replay value sky rocket. Granted, once you've mastered slow-mo. Some areas kinda made squads into cannon fodder. But even then, it was still fun. You still felt like a special forces badass that could take squads of soldiers single handed.
Boy am I glad Civvie 11 brought me here
Great video. Usually videos about FEAR AI tell us something about how it's not actually AI is so good, but developers trick us for it to appear smarter than it really is. It is the first video I've seen on this topic that lets you actully learn something about AI, not about how to trick player.
Is there another game that uses the AI like this?
It's always a blast to watch how the AI responds to your actions
As mentioned at the end of the video, a lot of games have used GOAP. But not many first person shooters.
I suspect it's a combination of building the tools but also that behaviour trees and Halo 2 really defined how FPS AI should be done going forward (go watch that AI 101 episode if you haven't yet).
I'd argue that a lot of FPS games aren't built for FEAR (or even Halo) style combat: sandbox combat in fixed open-area spaces. I think it's worth looking at it historically that this came along at the same time as Call of Duty rose to prominence, which isn't designed to be as sandbox-ey (for lack of a better term) and is more concerned with coordinated set pieces. The AAA industry has only in the last couple of years snapped back towards this more open idea of sandbox shooters again.
Fallout 3 and Deus Ex Human Revolution also used GOAP.
@@KingThrillgore wait, fallout 3? The AI is brain dead in that game lol
Congrats on passing the 100.000 subs mark! Once again a really interesting video!
Congrats on 100k, man! Another great video per usual, always love your content!
Unreal quality of your videos, been here since the early days, AI and Fear always stuck with me as a kid, I saw your first video on it and knew I’d found the channel for me, keep it up mate. Fantastic work.
Just discovered the channel through your Dark Souls video. I'm having a blast watching your videos, man. Very well done!
I found this channel through the Alien video like many others. I just wanted to thank the incredible production quality you have with these videos, you deserve more subscribers!
great videos. just the amount of details and technicalities that I like
Thanks for all the videos - they are fascinating, in depth, and really inspiring me to experiment with different types of AI. Much appreciated.
I love your videos. Well deserved success - congratulations on the benchmark!
AI and Games: Thank you for 100,000!
Me: You have that little of subscribers? You deserve more.
I'd like to live on a planet where a higher proportion of people are interested in AI at this level of detail!
Very well explained! Love your videos.
I've heard much of _FEAR_ 's Replica A.I. It's always a treat to learn even a crumb of the sheer complexity involved with making believable baddies.
I watched your coverage of _Alien: Isolation_ and had to tune in again!
i am dreaming of fear kind of game in this era
How have I not found this channel till now, this is amazing!
Welcome aboard! 🙂
Well this is a coincidence. I’m actually playing through F.E.A.R right now.
It's no coincidence, Google is always watching and listening, and the algorithm is always rolling
Damn ,.. you must be smart.
I love learning about AI thank you for the videos man.
I know I'm late in saying this, but congrats on the 100000 subscribers count! I'd love to see you make more videos dude 😉😁
Awesome video. Very clear and well explained!
I'm assuming the AI cycles through a list of commands until it comes to one it feels is the right successor to its original plan? Fascinating as always and congrats!
It runs the A* search algorithm on the current state. Only evaluating the states it knows it can execute (based on action preconditions) and searches for a combination of actions that satisfies the goal. The thing I omit in the video is that it requires a heuristic within the search to evaluate 'good' actions that will satisfy the goal. Each action can be scored to understand how 'good' it is to do at that point in time and prioritises searching good ideas first.
🔥🔥🔥🔥Thank you for material!
Someone at cdprojekt red should've watched this dudes videos before releasing cyperjunk 2077
Another useful video...
💯💯💯💯💯💯
Half-life, Halo CE and FEAR are easily my favourite shooters (also loved BF2 online)
Something about these games is still missing in most new games…
Passion for the project and innovation with NPCs
Fear AI is still some of my favourite ever
Can you do a video on Far Cry 2's AI and why it seems so alive and dynamic??
Or better, on why, despite using the same system, is Stalker's AI so bad?
fc2 ai is underrated
Needs a remake. But even if it never does, its a legendary game that did so much for the 7th gen console generation and the fps genre.
So I just implemented my own GOAP in C#, but wound up building the plans as trees instead of a list of actions because it was just easier. I guess I inadvertently implemented HTN?
Hey Tommy! Found you through your Alien Isolation Revisited video. But, I bet you could have guessed that with almost 100% certainty lol. Great work, have watched probably 15 or 20 by now in a marathon.
I was curious if you take on suggestions for games to cover. Either by, AI or design, whichever is more pertainent or interesting to you.
Your video on Ultima, got me wondering this. Since I still to this day adore these games.
Anyways, here are my requests/suggestions...
Ultima Online (1998 - First real MMO)
EverQuest (2000 - The first hugely popular and successful MMO)
Star Wars - The Old Republic (2014 - The most expensive and fully voiced MMO)
Just curious, genuinely, curious on your thoughts and what what you may know about design, AI and what actually made these games so great/popular (IMHO).
Like I said, just a request on my curiosity on what your specific opinion is of the game(s). I can find, I'm sure any long state video on RUclips about each game. But, I am genuinely interested in your specific opinion. You have a beautiful voice to listen to and your thoughtfulness in your scripts and points are wonderful.
Thanks in advance, if you see this.
very good video :D
Niiiice..
Keep up the awesome work Tommy!
Question though... what if the AI doesn't fully open the door?
Will it not recognise it now as it's no longer a door, it's A Jar?
*Badum-tsh*
Hnnnnnnnnnnnng.... :D
(In all seriousness, you're in-game systems would need to be able to properly model whether the door is fully open or not and then decide whether that means it's a) still closed, b) open or c) stuck in the middle. At which point you'd need for either the game to shut the door when the player isn't looking, or add an action to the planning handle that addresses how to open a slightly ajar door)
FEAR is a legendary benchmark for AI in games. Trepang2 has released and is said to be heavily influenced by FEAR.
A comparison between the AI systems of both would be very interesting.
I bought Trepang2 recently, and have been so busy I haven't had a chance to play it yet. But yeah, I'm very keen to see how it plays.
I know people dont rate the horror the original FEAR game, but honestly this game freaks me the fuck out, i play it like every 5 years or so, and christ its scary....
This might be a dumb question, but something I've been struggling with is platforming for AI, in both 3D and 2D games. Getting an AI to jump up onto a higher platform, or pathfind over multiple platforms to jump between. This seems to be a difficult topic to search. Do you have any suggestions on tutorials for this specific corner case?
Thank you. Is this related to the means-end analysis championed by Simon and Newell in the early 60's?
Is HTN kind of a nested GOAP?
You could have a GOAP designing a strategic plan, then nested, more specific GOAPs on how to carry out elements of the strategic plan… The strategic planner would decide the next step towards victory is to secure this sector, and then the detail planner takes the "secure sector" goal and comes up with observe, orient, decide, act except organically
Yeah sorta. HTN has two types of actions: primitive and compound actions. The former is exactly the same as GOAP. Compound are macros, which have several actions in a sequence (it can even have other compound actions in that sequence). So it gives designers a lot more control, given you can have 'secure sector' as a compound action and make sure it does exactly what the designer needs.
You can’t talk about video game AI without talking about FEAR 😉
12:22 This is so sad I cry everytime
Any chance you could update the two links for Jeff Orkins papers? Both links return error 404s.
Does anyone know if MT Framework game engine is available for free download? If so, where can I download this?
Does anyone why GOAP is now used less frequently? We are considering using it for our next game, but I would like to know why it was not used very much in the last decade before committing to it. I have not been able to find much information about the downsides other than that can be hard to debug and there can be performance issues. Is HTN better in those areas?
I alluded to this a little in another comment, but behaviour trees kinda beat GOAP to the punch and is a little more amenable for designers to get their head around. That said, while GOAP is great, I do think one big differentiator between GOAP and HTN (which I will focus on during the inevitable AI 101 episode on HTN) is that it's best to think of GOAP as bottom-up design versus HTN as top-down.
The paper I refer to from Jacopin acknowledges that a lot of GOAP plans are often very short, whereas his examination of HTN planners in Killzone 2 and Transformers shows HTN often generates much longer plans. That's in part because compound tasks in HTNs can glue multiple actions together, but GOAP never really looks at making larger, more contextually relevant decisions. Largely because the planning horizon is very short. If you consider how FEARs goals are all about avoiding the threat of the player, the plan to address that for the immediate future won't be very long. So you wind up with plans that make sense to execute in that 2 second epoch, but perhaps over a course of 10 seconds was actually a bad idea. HTN plans address this by grouping 'good' sequences of actions into compound tasks. In GOAP smart sequences are at the mercy of the planner picking them because it seemed like a good idea to satisfy the goal, whereas in HTN designers have a lot more control by designing those smart sequences into the task networks.
Could someone point me to the C++ code that he said was linked to in the description?
I clicked all the links but couldn't find the list of all goals used in FEAR.
Can you do some tutorials about doing this in ue4. There is not much documentation on this
Anyone know if there’s a remastered version of the games for the PS4? One, cause I really wanna play them and have them on my computer but it’s fucked so I can’t play them and Two, I’m to lazy to look it up myself.
So, a good fan-patch would be to remove the rats?
Civvie 11 sent me here!
Dame Pesos in da comments
Man, I always kill rats in video games, who knew I was saving system resources :)
hi civvie
Can anyone show me how to open F.E.A.R code? I have tried download the SDK but it doesn't seem to work :(
i like your voice :)
Anyone managed to get the Public Tools working? It says that I need to install FEAR before installing the public tools, I do have the Steam version of FEAR (patch 1.08) installed. The installer ignores it. The public tools are downloaded in the game folder along with the Steam instalation, so it's weird that they would include non-working tools.
E: Someone on Steam forums said that you can fix this by editing the patch value in regedit. I tried that (There was no Patch key, I had to create one) and it now says that I have the wrong version of the game. So that's at least something new, but still no bueno with the installing.
You can try this something like this
Windows Registry Editor Version 5.00
[HKEY_LOCAL_MACHINE\SOFTWARE\Wow6432Node\Monolith Productions]
[HKEY_LOCAL_MACHINE\SOFTWARE\Wow6432Node\Monolith Productions\FEAR]
[HKEY_LOCAL_MACHINE\SOFTWARE\Wow6432Node\Monolith Productions\FEAR\1.00.0000]
"Patch"=dword:00000008
"InstallDir"="d:\\FEAR_Platinum\\"
Or download standalone version from fear-combat.nl/FEAR-Public-Tools.html or www.moddb.com/games/fear/downloads/fear-public-tools-v2
I don't see the link to C++ code ://
the kinda....dumbing down o the AI in fear 2, mostly the replicas, could be maaaybe explained away as in fear 1, they had fettel leading them, they had a sort of, signal strength enhancer that allowed them to be smarter
in fear 2, their IS no psychic commander, alma activated them but she isent giving them direction
in fear 1, they had direction, leadership
in fear 2, they are awoken and not told anything; replica command has to try and....think of a plan and lead their troops
This games AI puts modern devs to shame
There is something in the water!..
Plz Do a Killzone 2 AI
If you want a great example for how easily the planning method can fall apart, watch the viral video of the programmer telling his children to write him instructions on how to make a peanut butter and jelly sandwich.
You keep refering to the first episode or the thrid episode. Your episodes arent labeled.
Thank God he's a Scot
The video was very informative, but lacks footage of what made that AI fun to interact with.
There was too much footage showing how underwhelming the AI can be when attacked in slow motion mode... and not enough of them actually looking for the player, flipping tables and all that.
all and all, good video regadless.
Versu design should had have more influence 😭
Rat lag got me killed
12:45
an entity using GOAP would not swallow a spider to catch the fly because the escalation to catch the spider, bird, etc lead to swallowing a horse and dieing, putting the entity into an invalid state.
the creepest game i have played as kid
Civvie11 brought me here
He who is afraid of AI is no different than he who is afraid about calculator.
No.
What the fuck are you on about.
That blur is killing my life... q_q
My friends and i became aware of the AI behaviour, when we first played Half Life 1...
Funny detail: had a Celeron 166MHz with 12 MB SDRAM and a Vodoo 2 Card, 2 MB (i still remember it, because the game ran veeeeery poorly - played it anyways.
... we were dreaming about how the AI will be developed in the future. Well, today i'm disappointed. All that counts are graphics as sellingpoint, and hidden gambling mechanics to maximize the income for the publishers. Maybe we get something groundbreaking out of chatGPT. Who knows...
Nakey Jakey baby, dog bless
It is a little disappointing that they didn't cheat the acronym to be Goal Oriented Action pLanning
lol GOAP
jakey, jakey, jakey, attorneys at law send me here...
From 5mins its nonsense. There is navmesh, npc walks from A to B. If door encauntered / door opened.
I dont get it, hes comparing the action of opening a door to a robots sent to mars
lol look at the AI in the game Rust, you'll get a laugh, it's like the worst AI of any game ive ever played, i know of games that are 20 years old with better AI/npc's it's unreal how bad it is
Man, why's your game so slow? Eh it's got a bunch of rats. Do you mean bugs? No, RATS.
lets face it the ''AI'' instilled into gaming thus far isn't ai in nature at all what so ever , this is the case until ''AI'' it's self does the programming , what we have to date is just flawed human conditioned rough random crappolla. ''suck it''