AI Planner - v0.2-preview released

The new 0.2.2 version requires 2019.3 but it's out of beta and on version 2019.3.0f5

https://docs.unity3d.com/Packages/com.unity.ai.planner@0.2/manual/index.html#technical-details

 

Is there a recommended way to update a Decision Controller's 'worldObjectQuery' at runtime? I'm working with many agents, each with their own decision controller. I want each to include their own 'house' in their world query once they acquire it.

Similarly, I was thinking of expanging the search radius if an agent can't find what it's looking for.

Thanks again

 

Is it possible to use AI Planner in ECS ?

 

For starters, if you have a specific object that you only want included by a single decision controller, then you have a few options:

• Set Local Objects on the decision controller
• Set a World Query to "From Object" and select the object you want (you'll have to set other constraints with traits, so that other decision controllers don't pick up the same house)

Re: updating the state of the planner. You have two options:

• Automatically when an action completes by setting the Next State Update field in the decision controller to "Use World State"
• Manually, outside of action completion by updating the traits on your GameObject's TraitComponent.GetTraitData<>.SetValue and then forcing an update to the planner state with IDecisionController.UpdateStateWithWorldQuery:
  https://docs.unity3d.com/Packages/c...I.Planner.Controller.IDecisionController.html

 

Please see this post near the bottom: https://forum.unity.com/threads/ai-planner-v0-2-preview-released.789809/#post-5295789

 

Awesome work guys

Can't wait to integrate this into my game I think it would be perfect
Any ETA on pure ECS? Even a preview?
Or tips on how to do it myself? I'm guessing setting up a hybrid ECS to GameObjects (AI components) would be easiest?
You say months earlier and if thats still the case I'll hold off on developing my own complex AI or trying to integrate it with ECS myself

 

Thanks for your response, but either you misunderstood me or I'm misunderdstanding you.

I would like to add the 'From Object' to world query while through script as the game is running, not through the editor. I worked out an alternative for the time being using a trait on my Agent's TraitComponent.

I've been having another issue, which may be a bug. An action, 'GoToWork' doesn't seem to be updating the planner state, meaning my agent plans to 'GoToWork' repeatedly, slowing everything down. As you can see in the screenshot., my Plan Size is only 3 but the 'Max Plan Depth' is 55. The same action leads to the same state leads to the same action ad infinitum. Despite me using a Max Plan Size of 25, there is only one state so there seems to be no limit on plan depth. Once the action is carried out, the world state updates correctly
and things continue as they should, but while an agent is 'going to work', my frame rate drops from around 250 to 20 or thereabouts, as it processes the endless loop.

In the screenshot I've included both the Plan Visualiser, the state data, and the planner action

 

found a little bug, i called a Int Field in a Trait "Amount", this causes an Error i guess because there is already a var with the same name (?)

i wonder if its possible to access Trait values by code?
Also generate Traits and Plans and so on at runtime ?

Also i wonder how to do this, for example.
I have a Trait int Food = 0.
I have the Trait bool Hunger = true
then i have an Action Eat it has an effect Food += 1
and then i want to have an Effect where i can say If Food == 10 then modify Hunger = true.
And then i can use Hunger == false as a Precondition.

and i think it should work like this ?
Atm i could only set a Precondition Food <= 10 or something to make it stop when satisfied ?
Im still new to this trying to figure it out
thx

 

Getting

Code (CSharp):

1. Internal: JobTempAlloc has allocations that are more than 4 frames old - this is not allowed and likely a leak

when running the Escape Room Sample on 2019.3.0f6.

 

I'm attempting to create a modular Eat action where the value recieved from the action is modulated by the animal's food type preference values.

So, something along the lines of this:

Enum Trait: EFoodType containing members Grass, Bush, Meat, Etc
Trait FoodType with EFoodType field named "Type"
Float Trait FoodPreference containing members Grass, Bush, Meat with associated values such as 1, 0.3, 0.05.
Float Trait FoodValue with a value of X that varies depending on the plant.

The Eat action in the Effects Modified section would do something like this:
Agent.Hunger -= Food.FoodValue * MatchedValueOf(FoodType.Type, FoodPreference.Type)

So the result if the animal ate Grass would be a value of 1 or a value of 0.05 in the case of eating meat. Where as another animal would recieve a value of 1 for Meat and 0 for Grass etc.

What would be a good way to go about this? I don't wish to lock off the other potential food types from being considered by doing a FoodType == FoodType in the precondition, just modulate the recieved value based on a preference.

I'm probably missing something absurdly obvious but the only approach I've come up with is to create duplicate eat actions for each plant and animal type, but that seems less than great as well.

Also, it seems like it might be useful to have additional operands in the effects section such as *=.

 

Yes, this happens due to some constraints in the allocators because we make use of long-running jobs. Once the planner ends up being Burst-compiled these warnings go away because the jobs complete before 4 frames. We'll look into this further to avoid spamming the console.

 

Thanks for the suggestion of the operand. I've added it, so it will be in the next release.

Re: your domain, I think you're on the right path, but I wasn't clear with your description whether FoodPreference is a trait that has a FoodType and a value multiplier field (e.g. FoodMultiplier). If not, then I'd suggest that that is what you add to the FoodPreference trait.

This will be an odd way to model the domain, but it's similar to an inventory. Until we support something like collections your food preferences will need to be "free-floating". Think of having 3 (or N for the number of FoodTypes) trait-based objects called AgentGrassPreference, AgentBushPreference, AgentMeatPreference, etc.

On each of those objects, you'd have the multiplier in the float field for that trait. Then in your action parameters you'd create a 'foodPreference' parameter for an object that has the FoodPreference trait.

In the preconditions you'd simply make sure that
foodPreference.FoodType == food.FoodType

Unfortunately, we don't have temporaries, so I'd suggest adding a scratch float field to foodPreference, so you could do the following:
foodPreference.FinalFoodValue = food.FoodValue
foodPreference.FinalFoodValue *= foodPreference.FoodMultiplier (once this becomes available)
Agent.Hunger -= foodPreference.FinalFoodValue

This would also be a good place for a CustomActionEffect if you want to write one:
https://docs.unity3d.com/Packages/com.unity.ai.planner@0.2/manual/CustomPlannerExtensions.html

Does this help?

 

is the value of trait parameter can be accessed in decision controller action callback?. or it must be do in someway?.
i have already try this

Code (CSharp):

1. var myTraitComponent = GetComponent<ITraitBasedObjectData>();
2. myTraitComponent.GetValue(TraitTypeName.FieldName);

but it return default value of the parameters, it must be updated right?. because i have already set the value will be modified in action effects.

 

This is what i get when i name an int32 "Amount", its fine when i name it "amount" or something else.
And it points to this line in SetField(): Amount = (System.Int32)value;

 

Hey Mplantady thx for your help !
I tried the first version wich is having a fixed value = 10 in the precondition, this works, thanks!
But when i make a Trait i.e. "Satisfaction" i set its basic value to 10 (i also added it to my agent trait component, and in the action parameters) and then i ask for it in the preconditon "Food"< "Satisfaction" nothing happens.
In the visualizer it shows only the first Node.
Or is there another way to make a "global object" like you said?

Another question i have is there a way to make Actions take a certain time?
In my game i was trying something similar to this project, but im by far not yet as good of a coder, wich is why i love this project . But there i had callbacks for start/mid/end of an action and also a time defined and also then it could choose a path of actions not only by shortest nodes but also by shortest time? On the other hand it makes sense to make it all with preconditions, its done when its done. But there must be a way to wait a certain time i.e. a cooldown after an attack ?

 

Thanks amirebrahimi, that is a helpful suggestion for working around the lack of collections for now. And yup, that is how I setup the FoodPreferences. I'll look into the custom actions. Thanks!

 

If you are asking about how to get the updated value of a trait coming from the planner through an action, then take a look at the Player prefab in the Match 3 sample. It makes use of the Cell trait coming back from the SwapRight and SwapUp actions: https://github.com/Unity-Technologi...ob/master/Match3/Assets/Scripts/Player.cs#L80

Hope that helps.

 

I think I need a hand understanding the Cost/Rewards piece of AI Planner. I'm sure it's logical, but it seems kind of arbitrary in my mind. What does it mean if I give an action a reward value of 1? What does 1 mean? Is there a unit? If I have two characters with two different goals, but they're considering the same set of actions, what reward are they "thinking" of, if that makes sense? Thanks!

 

Super helpful! Thank you, look forward to your blog post! As for recommended/best practice, would it be fair to say I would just want to make sure the values relative to each other are as accurate as I can guess (ie: it's more important to get to the goal than to get coins, but exactly how much more is a little bit of an art), then test it out, and tweak the values from there?

 

Thank you for the additional info Trevor, it is useful and I look forward to the blog post.

Regarding the use of the utility approach in AI Planner; one of the nicer aspects of a utility based ai is the non-linear nature of the responses. Is that currently driven by the rewards heuristic or is there a good way of modeling that now?

Also, when considering an action, does the current basic heuristic take the trait effect modifiers into account or is it solely driven by the action reward value?

A few random things I found myself wanting while working with the system:

• Line of sight based object filtering.
• From collection object filtering.
• Action duration
• Effect modifier utilizing duration
• Add trait @ action start (i.e. in use)
• Remove trait at action end. (I understand there is the ability to add the group bit so perhaps that is the preferred flow)
• Simultaneous or stacked actions (i.e. multi-tasking ala Sims4)
• Effect utility modifier
• Add/Remove conditions
• Reward utility modifiers
• Multiple, potentially competing, plans.
• Near comparison (as opposed to == ) for vector based comparison.

I understand that some of the above may be more project specific extensions that the end user is expected to make but thought I'd toss them out just in case.

Pretty neat and solid system so far, really looking forward to the future updates.

 

Fantastic, thank you for the additional information!

 

Is there any updates on when the ECS version of this will be out? Last comment mentioned it was on the pre-GDC roadmap. Thanks.

 

Sounds good. Thanks for the update.

 

Hello @TrevorUnity . Are you guys still on track for the DOTS release this month? Thanks.

 

Appreciate the prompt response. We are trying to plan around integration with the AI Planner and using DOTS for us is a must for performance reasons. Is there a time table on this newest release or some way we could start building our own DOTS modules today in preparation? We really like the tech for our game. Thanks again.

 

Responded via PM.

 

yes please!

 

Hallo
AI planner does not appear in the package manager
how can I install it?
Do I need some licence upgade for the planner?
I have unity 2019.3

Thanks

 

Hallo
Ah ok that was the problem thanks!
I will try to apply it not to a game but to an industrial environment simulation
I am very curious to see the performance
btw Great idea to offer a planner for games!

Flavio

 

Hallo
I tried to setup a simple scene where a cube can move on an grid of locations controlled by the AI Planner.
I can build the game as an exe file and it runs(even if cube does not move for some reason)
The problem is that if I try to play it inside the editor I got this message here



Something in the generated code went wrong

Tried the following:
-remove the 'Temp' folder
-remove the 'obj' folder
-close and restart Unity and visual studio

I am using
Unity Version 2019.3.10f1
AI Planner Version 0.2.3

Any Idea?
Thanks!

 

Hi guys, question: is there any plan to bring this package up to 2020 beta or even alpha territory? I'm planning to work on an early prototype (4X civ-like game with a couple of twists and different setting, different themes and all), I would start to use full DOTS spectrum and the Hybrid Renderer V2 with URP support.
It would be awesome to include the AI planner too.
I will play around with it with lesser versions, obviously, but is there any rough timeline what can I expect?
And obviously I'm not talking about production ready state by any means, only preview.

Currently I'm getting this error trying Planner 0.2.3-p1 with Entities 0.11,0-p7 and on 2020.0a13.
Library\PackageCache\com.unity.ai.planner@0.2.3-preview.1\Runtime\Controller\BaseTraitBasedPlanExecutor.cs(114,91): error CS0117: 'ScriptBehaviourUpdateOrder' does not contain a definition for 'CurrentPlayerLoop'

I guess this is expected.

 

Hallo Trevor
I removed all the generated .prj project files but no luck.
I noticed it is not regenerating them by the way, I guess the process stops before the generation....

The project is here
https://github.com/FlavioBortolan/ai-planner-samples/tree/master/Planner 7

Thanks!

 

Thanks a lot
Now compilers errors have gone
Let's see if I manage to get the cube to its destination
Thanks
Flavio

 

Hi there, we're really interested in using AI Planner but have a couple of questions we couldn't figure out:

Our game involves subjective world state – each AI has their own memories of the world state, and they only actually interact with the real world when they can (for example) visually confirm the object is where they thought it was located. So all the AI decisions are made based on the agent's memory (a custom data structure we have), not the ground truth of the state of GameObjects.

Is there a way to implement that with AI Planner? Perhaps with some kind of special World Query?

Also, we are hoping to use ECS for this as we would like to simulate 500-1000 agents simultaneously, is AI Planner designed to handle that? In theory, our search space for each individual AI is relatively limited.

Thanks!

 

Does the Planner use the job system? or is it planned?
I would love to use it but I'm aiming for a LARGE scale rts with DOTS
Some sort of load ballencer will probably help, but being able to schedule the planner as a job (which is able to take a few frames) would probably be necessary