Go to this video's Sponsor: snhu.edu/llamacademy to learn more about an accredited degree program in Game Design and Development! Special thanks to SNHU for sponsoring this video! In case you skipped the description, here you can find some free and paid solutions for State Machines, Behavior Trees, and GOAP!: ⚫ UFSM - A free and open source Hierarchical Finite State Machine framework: github.com/Inspiaaa/UnityHFSM ⚫ Fluid Behavior Tree - a Free and Open Source Behavior Tree Framework: github.com/ashblue/fluid-behavior-tree ⚫ Opsive Behavior Designer: assetstore.unity.com/packages/tools/visual-scripting/behavior-designer-behavior-trees-for-everyone-15277?aid=1101l9QvC ⚫ AI Tree: assetstore.unity.com/packages/tools/behavior-ai/ai-tree-229578?aid=1101l9QvC ⚫ Node Canvas: assetstore.unity.com/packages/tools/visual-scripting/nodecanvas-14914?aid=1101l9QvC ⚫ Free GOAP Framework: assetstore.unity.com/packages/tools/behavior-ai/goap-252687?aid=1101l9QvC | github.com/crashkonijn/GOAP Some of these links may be affiliate links, which at no additional cost to you, gives me a small portion of the purchase.
Thanks for the nice video! As some other comments have mentioned, combining hierarchical state machines with behavior trees can be very powerful. I've done this with behavior trees as the base framework of the AI, and the state machines existing as "Selector" nodes in the tree. When the state machine node is ran, it runs the statemachine logic and selects the branch that corresponds to that state. If a single state's tree becomes too complex, it can be broken down into smaller substates each having their own behavior tree and a new statemachine node at the root of them. I suppose it all comes down to preference but I prefer to handle scripted decision making with heirarchical state machines as you can have short lists of all the conditions that cause state transitions in one spot for each state or substate. Whereas behavior trees can become quite unruly if you have a single large tree jumping around to different points in that tree. (You have to remember which parts of the tree jumps to which other parts and how those nodes decide to jump. Essentially, the transitions feel scattered across the entire tree which has driven me crazy in the past) So, I try to keep behavior trees smaller and more readable by keeping them as directed acyclic graphs. Basically, just a convenient way to form sequences of actions with a given flow, while HFSM handles the transitions. If the desired behavior needs to more complex or emergent behaviors are desired, I agree that other frameworks like Utility AI or some kind of planner like GOAP or HTN can be nice. Really just comes down to what kind of behaviors you want from the AI and weighing the pros and cons, but I often like using the most simple frameworks first, and only delving into the more advanced frameworks for the specific AI states that need to be more complex.
GOAP can implemented in a behavior tree. Just add the concept of a Triage Sequence. You can make a weighted sequence of goals, and update their importance over time. the most important goal is run, and the dependency chain of actions are just behavior tree sequences.
I thought so, but upon looking closer, no I don't think so. The decoupling of actions is something unique to GOAP. Essentially, if killing the player requires getting a weapon first, then in a BT you need to encode that into the "kill the player" sequence. In GOAP, if you have some abstraction, it doesn't have to. For example, in GOAP you can give items a "damage" variable and tell the AI that to kill a player it needs something that does damage. Now much later in development, you add a barrel and decide that if it falls on a player, it should do damage, so it gets its "damage" value set to something not zero. Suddenly, without any changes, the AI may decide to pick up a barrel and throw it. I don't see how to do that in a behaviour tree. Behavior trees can do Utility decisions (an AI framework ignored in this video). Maybe Utility AI is kind of a between?
@@LemuriaGames the behavior tree agent would have a list of items with WeaponData components, that labels its fire rate, damage, range, reload speed, etc... and the killPlayer Goal would triage those items with a pickBestWeapon function. Adding components allows decoupling, and anything that doesn't specifically fit into the concept of a behavior tree node can be a function called by the behavior tree. so triaging with components makes a behavior tree into a planner.
@@RPG_Guy-fx8ns Yes, like I said: A kind of hybrid. Since BTs are essentially just a bunch of boolean decision points, you can put anything in there. But somewhere you're making the components into the actual code and you're just calling it a BT when the logic has long since moved out of the tree.
@@LemuriaGames True. I am not saying a BT is a GOAP, I am saying, it doesn't take much to turn BT into a planner. You just add weapon select logic, and if you use A* pathfinding through the strategy space, in that weapon select logic, I think it becomes GOAP at that point.
AI Tree is pretty good, much close to unreals behavior trees. Behavior designer starts getting very cumbersome to work with after having a couple of deep branches. Also EQS is a really important point in that one, because without EQS, you need a lot of leg work on AI positioning.
I prefer the goap like structure. I have pseudo coded a structure similar to a goap-like structure for my simulation games and I think those kinds of structures make the most sense for complex behaviors. I just didn't realize the structure had a name. lol. Now it will be easier to find how others have implemented it. Thanks for the exposure to all this info!
This is a great video with a really good overview of the different AI behavior models. It has enough detail to be very informative on each topic but not so much to be overwhelming with info, love it.
The vast majority of AI enemies can simply use a state machine if all they need to do is pathfind and attack. Behavior Trees & GOAP are more for cases where your AI unit needs to interact with other systems.
4:35 don't waste your time at school, ESPECIALLY with something like game dev. It's incredibly expensive and most industry jobs ARE NOT worth it and are probably NOT prevalent in your area. Besides, the only way you will ever learn something is if you have enough interest in it to the point where you can teach yourself.
I am using a combination of hierarchical state machines and behavior trees for my XCOM-style prototype. Basically I run a different tree in every state. I coded both (borrowing A LOT from existing literature), so no fancy UI, all in code, but still quite modular and I am very happy with the result. I recommend looking into this sinergy. Hierarchical SMs can really address a lot of the scalability problems of FSMs.
I agree HFSM helps a lot with scalability, but still I find it cumbersome to set up all of the transitions as is required for responsive AI decision making in a FSM/HFSM. It can be very effective. My preference for more complex leans towards BT/GOAP though because I find them easier to work with as the complexity increases.
I you look closely you will see that BT is state machine. Difference is transition rules fused with tasks and transitions go to next branch. That simplification of transitions sacrifice flexibility for better readability.
Thank You for making this video 💯👍🏻. So far I've been using my own solution based on FSM with some additions to address some special use cases, but I want to expand those functionalities, so I really have to choose a more developed, existing solution. So I will rewatch your video in the future as I will have to choose an A.I. Framework for a Game with enemies and bosses functioning under a "more interesting" A.I. Behaviour. It's very good to be able to count on you as a teacher for these "cool things" related to Unity, Game Dev, A.I., etc 😄
I use state machines all the time. The exponential number of transitions is only a theoretical possibility. The number of transitions actually needed is usually much much lower. Also, some transitions are so fast that they can be triggered from any state without issues.
Behavior trees are still state machines, however it organizes the states by hierarchy and goes into sub-states. You can massively increase the scalability of a state machine by creating sub-states. There is no reason anyone would have to define every single transition condition between every state.
I didn’t dive into Utility AI yet, but with my super basic understanding it sounds like that’s the ultimate combination for advanced AI like RTS “main” AI. I was thinking about the RTS AI a lot for this video. I cut out all my discussion on it because I thought it needed more investigation before I talked a lot about it. Even with GOAP I was struggling to see how the RTS AI would be able to multitask effectively outside of throwing several options against the GOAP system and picking the “best” one. From what I understand this is a function Utility AI can handle more effectively when combined with GOAP.
@@LlamAcademy I'm not familiar enough with GOAP to know how it interacts with with Utility AI. Maybe with the priority system but, Utility AI seems to use that too. I've spent a week or so learning and documenting WiseFeline which I really like. If you ever want to do a video on the topic and want an easier path to get started, I'd be happy to share my notes and answer any questions. By the way, what I like about Utility AI is that you basically have Actions, Considerations, and Targets as your main concepts. There are typically multiple Considerations on each Action which get added together. Based on the resulting value, it is determined what Action should be taken. While an Action doesn't always have to go to a "Target' that is often what the Action does. So... you have a pretty simple structure which just keeps scoring itself. As things change on your screen, the scores will continue to determine what the proper action is and what target to go to or what thing to do. Hope this helps.
@@LlamAcademy Yeah GOAP doesn't seem like a good fit with an RTS. However, Utility is one since you don't need to run the A* algorithm for each decision. The dark art with Utility AI is the costing function to determine your considerations and actions. FYI, there is a newer form of GOAP that organizes collections of actions into tasks. So it saves some processing by fitting together these collections of actions to meet the goal. I unfortunately do not remember what is it called off the top of my head but it was used in the Transformer: Cybertron games from a few years ago.
Nice video! I did not know about GOAP. Your video is missing talking about of logical / planning languages such Prolog and PDDL to have a full analysis :)
i'm new to making games in unity right now i'm using a mix between scripts to checks for conditions and animator state machine to perform them. enemy ai always trigger my procrastinator within ;D now i'm gonna take a look at behaviour trees
This is awesome, right before I want to start with enemies :) Btw. do you mind doing something like this for Player Controllers as well? I'm currently struggling combining my movement, camera and weapon scripts, which are based off of (hierarchical) state machines and how they communicate with another. Like how my movement influences the camera and if I should be able to shoot while sliding. Still can't find a good solution and am now working with observer based patterns to send signals between those objects.
Great question! GOAP and Utility AI are different. DukeZhou on Stack Exchange does a pretty good writeup of how they are different: ai.stackexchange.com/questions/8651/what-is-the-difference-between-goal-based-and-utility-based-agents
@@LlamAcademy sir tell me how can i create enemy strafing bheavior when player is also strafing. I mean just watch sekiro vs isshin boss fight on yt you will get qn idea what i am talking about
Dominos like the tabletop game with the little white pieces and black dots? I'm not an expert at this game but I think it's relatively simple based on what you have available and the available moves. With my current knowledge of that game I'd honestly probably start with Unstructured AI because I think it's very simple. You have it analyze your "hand" and the board and pick the "best" or a random available move. Even to add difficulty you could just weight the available moves on how good they are and have different difficulty levels prefer better/worse moves.
GOAP is OP, but writing it is nightmare. Especially optimizing the planner. If you make a strategy game (like me), basic approach with table of variables might easily kill your computer due to thousands of entries for each agent.
Thanks for mentioning my project! Means a lot ❤
Thanks for providing such an awesome tool for us in an open source format 🧡
you're my saviour
Go to this video's Sponsor: snhu.edu/llamacademy to learn more about an accredited degree program in Game Design and Development! Special thanks to SNHU for sponsoring this video!
In case you skipped the description, here you can find some free and paid solutions for State Machines, Behavior Trees, and GOAP!:
⚫ UFSM - A free and open source Hierarchical Finite State Machine framework: github.com/Inspiaaa/UnityHFSM
⚫ Fluid Behavior Tree - a Free and Open Source Behavior Tree Framework: github.com/ashblue/fluid-behavior-tree
⚫ Opsive Behavior Designer: assetstore.unity.com/packages/tools/visual-scripting/behavior-designer-behavior-trees-for-everyone-15277?aid=1101l9QvC
⚫ AI Tree: assetstore.unity.com/packages/tools/behavior-ai/ai-tree-229578?aid=1101l9QvC
⚫ Node Canvas: assetstore.unity.com/packages/tools/visual-scripting/nodecanvas-14914?aid=1101l9QvC
⚫ Free GOAP Framework: assetstore.unity.com/packages/tools/behavior-ai/goap-252687?aid=1101l9QvC | github.com/crashkonijn/GOAP
Some of these links may be affiliate links, which at no additional cost to you, gives me a small portion of the purchase.
i find the goap cons funny, because when i started imaging ai for my game, its the first design i thought of, without ever hearing about it.
You are too powerful 💪 great first idea
Thanks for the nice video! As some other comments have mentioned, combining hierarchical state machines with behavior trees can be very powerful. I've done this with behavior trees as the base framework of the AI, and the state machines existing as "Selector" nodes in the tree. When the state machine node is ran, it runs the statemachine logic and selects the branch that corresponds to that state. If a single state's tree becomes too complex, it can be broken down into smaller substates each having their own behavior tree and a new statemachine node at the root of them.
I suppose it all comes down to preference but I prefer to handle scripted decision making with heirarchical state machines as you can have short lists of all the conditions that cause state transitions in one spot for each state or substate. Whereas behavior trees can become quite unruly if you have a single large tree jumping around to different points in that tree. (You have to remember which parts of the tree jumps to which other parts and how those nodes decide to jump. Essentially, the transitions feel scattered across the entire tree which has driven me crazy in the past)
So, I try to keep behavior trees smaller and more readable by keeping them as directed acyclic graphs. Basically, just a convenient way to form sequences of actions with a given flow, while HFSM handles the transitions.
If the desired behavior needs to more complex or emergent behaviors are desired, I agree that other frameworks like Utility AI or some kind of planner like GOAP or HTN can be nice. Really just comes down to what kind of behaviors you want from the AI and weighing the pros and cons, but I often like using the most simple frameworks first, and only delving into the more advanced frameworks for the specific AI states that need to be more complex.
Great suggestion. I typically start with the most simple approach and move up in complexity only when required as well.
Wish I could give more, keep up the great work!
Wow! Thanks so much Travis! 🧡
GOAP can implemented in a behavior tree. Just add the concept of a Triage Sequence.
You can make a weighted sequence of goals, and update their importance over time.
the most important goal is run, and the dependency chain of actions are just behavior tree sequences.
I thought so, but upon looking closer, no I don't think so. The decoupling of actions is something unique to GOAP. Essentially, if killing the player requires getting a weapon first, then in a BT you need to encode that into the "kill the player" sequence. In GOAP, if you have some abstraction, it doesn't have to. For example, in GOAP you can give items a "damage" variable and tell the AI that to kill a player it needs something that does damage. Now much later in development, you add a barrel and decide that if it falls on a player, it should do damage, so it gets its "damage" value set to something not zero. Suddenly, without any changes, the AI may decide to pick up a barrel and throw it. I don't see how to do that in a behaviour tree.
Behavior trees can do Utility decisions (an AI framework ignored in this video). Maybe Utility AI is kind of a between?
@@LemuriaGames the behavior tree agent would have a list of items with WeaponData components, that labels its fire rate, damage, range, reload speed, etc...
and the killPlayer Goal would triage those items with a pickBestWeapon function. Adding components allows decoupling, and anything that doesn't specifically fit into the concept of a behavior tree node can be a function called by the behavior tree.
so triaging with components makes a behavior tree into a planner.
@@RPG_Guy-fx8ns Yes, like I said: A kind of hybrid. Since BTs are essentially just a bunch of boolean decision points, you can put anything in there. But somewhere you're making the components into the actual code and you're just calling it a BT when the logic has long since moved out of the tree.
@@LemuriaGames True. I am not saying a BT is a GOAP, I am saying, it doesn't take much to turn BT into a planner. You just add weapon select logic, and if you use A* pathfinding through the strategy space, in that weapon select logic, I think it becomes GOAP at that point.
I can always count on your channel for learning something new or even just clarifying some general concepts, even after years of gamedev.
AI Tree is pretty good, much close to unreals behavior trees. Behavior designer starts getting very cumbersome to work with after having a couple of deep branches. Also EQS is a really important point in that one, because without EQS, you need a lot of leg work on AI positioning.
Agreed. EQS is a really nice benefit for using AI Tree over some of the other mentioned behavior tree solutions!
внатуре
@@DELPRODBEATS really
@@DELPRODBEATS pls speak English, Dunduk
Whats EQS
I prefer the goap like structure. I have pseudo coded a structure similar to a goap-like structure for my simulation games and I think those kinds of structures make the most sense for complex behaviors. I just didn't realize the structure had a name. lol. Now it will be easier to find how others have implemented it. Thanks for the exposure to all this info!
I’m the same man.
This is a great video with a really good overview of the different AI behavior models. It has enough detail to be very informative on each topic but not so much to be overwhelming with info, love it.
GOAP actually sounds like the most "natural" of the solutions, closer to how we make choices as opposed to having some kinda tree in the mind.
@@3zzzTyle it is but what is most “natural” is actually rarely the best path to take in game development
The vast majority of AI enemies can simply use a state machine if all they need to do is pathfind and attack. Behavior Trees & GOAP are more for cases where your AI unit needs to interact with other systems.
4:35 don't waste your time at school, ESPECIALLY with something like game dev. It's incredibly expensive and most industry jobs ARE NOT worth it and are probably NOT prevalent in your area. Besides, the only way you will ever learn something is if you have enough interest in it to the point where you can teach yourself.
I am using a combination of hierarchical state machines and behavior trees for my XCOM-style prototype. Basically I run a different tree in every state. I coded both (borrowing A LOT from existing literature), so no fancy UI, all in code, but still quite modular and I am very happy with the result. I recommend looking into this sinergy. Hierarchical SMs can really address a lot of the scalability problems of FSMs.
I agree HFSM helps a lot with scalability, but still I find it cumbersome to set up all of the transitions as is required for responsive AI decision making in a FSM/HFSM. It can be very effective. My preference for more complex leans towards BT/GOAP though because I find them easier to work with as the complexity increases.
I agree. HSM and BT have different strengths. A mix of both like you have seems to work very well.
I you look closely you will see that BT is state machine. Difference is transition rules fused with tasks and transitions go to next branch. That simplification of transitions sacrifice flexibility for better readability.
Thank You for making this video 💯👍🏻.
So far I've been using my own solution based on FSM with some additions to address some special use cases, but I want to expand those functionalities, so I really have to choose a more developed, existing solution.
So I will rewatch your video in the future as I will have to choose an A.I. Framework for a Game with enemies and bosses functioning under a "more interesting" A.I. Behaviour.
It's very good to be able to count on you as a teacher for these "cool things" related to Unity, Game Dev, A.I., etc 😄
I use state machines all the time. The exponential number of transitions is only a theoretical possibility. The number of transitions actually needed is usually much much lower. Also, some transitions are so fast that they can be triggered from any state without issues.
I would like to see some tutorials on trees or goap, in terms of some action or strategy game
Stay tuned 😉
underrated af
if you want a video subject I think it could be interesting to do a series on the development of a hearthstone-style combat card system
Thanks!
Very good explanation. Can you make a tutorial on G.O.A.P?
Thanks! I'm working on something with GOAP now 😉 stay tuned
@@LlamAcademy looking forward to it! :)
Behaviour Designer (asset shown on video) is on sale right now btw.
very good! Relevant content.
Behavior trees are still state machines, however it organizes the states by hierarchy and goes into sub-states. You can massively increase the scalability of a state machine by creating sub-states. There is no reason anyone would have to define every single transition condition between every state.
I also like the Free Behaviour tree runner made by The kiwi coder, a mate developper on unity.
got some bugs, but does the job, higly customizable.
Great video. Any thoughts on Utility AI or a GOAP/Utility hybrid?
I’ve been using the WiseFeline Utility AI. I haven’t combined it with GOAP but, I really like it’s flexibility.
I didn’t dive into Utility AI yet, but with my super basic understanding it sounds like that’s the ultimate combination for advanced AI like RTS “main” AI. I was thinking about the RTS AI a lot for this video. I cut out all my discussion on it because I thought it needed more investigation before I talked a lot about it.
Even with GOAP I was struggling to see how the RTS AI would be able to multitask effectively outside of throwing several options against the GOAP system and picking the “best” one. From what I understand this is a function Utility AI can handle more effectively when combined with GOAP.
@@LlamAcademy I'm not familiar enough with GOAP to know how it interacts with with Utility AI. Maybe with the priority system but, Utility AI seems to use that too. I've spent a week or so learning and documenting WiseFeline which I really like. If you ever want to do a video on the topic and want an easier path to get started, I'd be happy to share my notes and answer any questions.
By the way, what I like about Utility AI is that you basically have Actions, Considerations, and Targets as your main concepts. There are typically multiple Considerations on each Action which get added together. Based on the resulting value, it is determined what Action should be taken. While an Action doesn't always have to go to a "Target' that is often what the Action does. So... you have a pretty simple structure which just keeps scoring itself. As things change on your screen, the scores will continue to determine what the proper action is and what target to go to or what thing to do.
Hope this helps.
@@LlamAcademy Yeah GOAP doesn't seem like a good fit with an RTS. However, Utility is one since you don't need to run the A* algorithm for each decision. The dark art with Utility AI is the costing function to determine your considerations and actions.
FYI, there is a newer form of GOAP that organizes collections of actions into tasks. So it saves some processing by fitting together these collections of actions to meet the goal. I unfortunately do not remember what is it called off the top of my head but it was used in the Transformer: Cybertron games from a few years ago.
I have been learning about Utility AI. This video series gave me a good start ua-cam.com/play/PLDpv2FF85TOp2KpIGcrxXY1POzfYLGWIb.html
Nice video! I did not know about GOAP. Your video is missing talking about of logical / planning languages such Prolog and PDDL to have a full analysis :)
Thanks! Maybe PDDL deserves its own full video 🤔
@@LlamAcademy yes PDDL is not very known
i'm new to making games in unity
right now i'm using a mix between scripts to checks for conditions and animator state machine to perform them.
enemy ai always trigger my procrastinator within ;D
now i'm gonna take a look at behaviour trees
Great video. Have always been wondering about the different AI Behaviour Framework.
Curious to know how other games does their AI
Don't underestimate the "Any" state as it can make your spiderweb look more like a tree, and avoid n² 😅
When it comes to state machines, I would never do the core transitions in the animator. I do it through code and then just use the ‘Any State’ node.
What about utility AI?
Can there exist hybrid systems like GOAP + State Machine?
Yup!
This is awesome, right before I want to start with enemies :)
Btw. do you mind doing something like this for Player Controllers as well?
I'm currently struggling combining my movement, camera and weapon scripts, which are based off of (hierarchical) state machines and how they communicate with another.
Like how my movement influences the camera and if I should be able to shoot while sliding.
Still can't find a good solution and am now working with observer based patterns to send signals between those objects.
I second this, same problems
7:36 thats why i call animations from code
Would GOAP be the same as Utility AI?
Great question! GOAP and Utility AI are different. DukeZhou on Stack Exchange does a pretty good writeup of how they are different: ai.stackexchange.com/questions/8651/what-is-the-difference-between-goal-based-and-utility-based-agents
Which one does GTA or Elden ring use? Anyone know?
Behaviour Designer tutorial series please?
Something with designing Behavior Trees is in the works 😉
You're a savior Liam@@LlamAcademy
sir how can we create advanced AI like souls game like sekiro dark soul please please
I haven’t played that so not sure 🤔
@@LlamAcademy no need to play just watch sekiro enemy ai video that enough
@@LlamAcademy sir tell me how can i create enemy strafing bheavior when player is also strafing. I mean just watch sekiro vs isshin boss fight on yt you will get qn idea what i am talking about
I feel like goal oriented is simply an abstraction of AI.
what do you think will be a good framework for a game like domino ?
Dominos like the tabletop game with the little white pieces and black dots? I'm not an expert at this game but I think it's relatively simple based on what you have available and the available moves. With my current knowledge of that game I'd honestly probably start with Unstructured AI because I think it's very simple. You have it analyze your "hand" and the board and pick the "best" or a random available move. Even to add difficulty you could just weight the available moves on how good they are and have different difficulty levels prefer better/worse moves.
@@LlamAcademy , got it thanks!
fluid behavior tree and its free ;-)
Thanks! I'll add that as a reference in the description!
thanks!
a bit overkill on the red paint there.
Yeah....I got it done the day before I recorded this so it was still on my scalp 😅
How about HTN?
GOAP is OP, but writing it is nightmare.
Especially optimizing the planner. If you make a strategy game (like me), basic approach with table of variables might easily kill your computer due to thousands of entries for each agent.
I don't see how goap is different from behavior tree, it just depends how you set up the order of priorities in the behavior tree
nice overview
Thanks 😊!
I just trigger all my animations from anystate. no exponential transition needed... may not be perfect, but easy.
oh snap its harry from 3rd rock from the sun
wait you dont even get these things by default in Unity? 😂
spawn - patrol - chase - attack -off-chase - dieing with anim and stuff
A blend tree would fix the animator mess
miss a lot's of stuff hahahha
What about HTN?