AI Planner - v0.1.1-preview Release!

exciting news! thanks, taking a look now.

 

Thanks for good work, waiting for you update.

 

Not working currently on 2019.3.0b1, throws a bunch of

Code (CSharp):

1. MissingMethodException: void UnityEngine.UnityAPICompatibilityVersionAttribute..ctor(string,string[])

Downgraded and it seems to work now.

 

Package documentation are now available.

 

Otto Does not work with 2019.2.3f1

Assets\AI.Planner\Workaholic-Custom\CustomNavigateEffects.cs(3,18): error CS0234: The type or namespace name 'Domains' does not exist in the namespace 'AI.Planner' (are you missing an assembly reference?)

 

Do you have an Assets\AI.Planner\AI.Domains.Dll in your project?

 

Assets\AI.Planner\Workaholic-Custom\CustomNavigateEffects.cs(3,18): error CS0234: The type or namespace name 'Domains' does not exist in the namespace 'AI.Planner' (are you missing an assembly reference?)
2019.2.2~2.3 Same error

 

Yes the DLL is in there, it is what is currently in Otto

https://drive.google.com/open?id=12a_T8w_rPdyttXwxcbKB4YRTNQKLTyVJ


Assets\AI.Planner\Workaholic-Custom\CustomWorkEffects.cs(3,18): error CS0234: The type or namespace name 'Domains' does not exist in the namespace 'AI.Planner' (are you missing an assembly reference?)

Assets\AI.Planner\Workaholic-Custom\CustomWorkEffects.cs(8,85): error CS0246: The type or namespace name 'StateData' could not be found (are you missing a using directive or an assembly reference?)

Failed to load 'C:/Users/vinci/Desktop/otto-master/Assets/Models/Otto/Animation/Item_GrabPocket.anim' because it was serialized with a newer version of Unity. (Has a higher SerializedFile version)

This last one is weird because the Git hub says Otto was updated to 2019.2.2f1 but I am opening with 2019.2.3f1

 

Some how

From this screenshot it looks like the DLLs aren't importing correctly because it shouldn't show it as a Native DLL (it's managed). What is the file size of the DLL? It should be around 26KB.

 

Well I got it working, I forgot that if you download a zip from GIT it will strip big files.

Otto must be download via command line or via something like source tree.

Plus I had to rebuild the library. But Otto is working! Even 2 at the same time!

 

I've started a thread within the GitHub community to encourage the ability to disable downloading of .zip files for owners of repositories, but haven't heard back from anyone at GitHub yet. Unfortunately, this causes many issues. In the future, I think we will include samples with the package. Otto is a bit too large to be a sample, which is why we didn't include it.

 

How did you rebuild the library? I have cloned the project from the command line and have confirmed my Domains file is present and is 26kb, but I still get:

Assets\AI.Planner\Workaholic-Custom\CustomNavigateEffects.cs(3,18): error CS0234: The type or namespace name 'Domains' does not exist in the namespace 'AI.Planner' (are you missing an assembly reference?)

 

Close the project, go to the root folder, where the assets, project settings and library folders are. . . . delete the library folder. Re open the project

 

Ok next problem. . . when I upgrade a project that had the older planner version in it I was not able to create all the scriptable assets, so I removed all the old stuff.

It started complaining

Microsoft (R) Visual C# Compiler version 2.9.1.65535 (9d34608e)
Copyright (C) Microsoft Corporation. All rights reserved.

error CS2011: Error opening response file 'C:\ASC\DragonsLore\game\Assets/AI.Planner/csc.rsp'

It also did not generate a new AI.Planner folder

In a fresh project I do see the AI section in the menu and how it generates the assemblies ( if a folder is already corrected)

But it does seem like this missing response file has broken my old project =(


Update: I reverted that file and it allowed the package to complete installation and create all the assets. BUT I still have to have that file, does anyone know where it is being set?


Liking the new interface.

I noticed in the documents we need a new class

public class YourAgentClass : BaseAgent<YourAgentClass, DomainObject, StateEntityKey, StateData, StateDataContext, ActionScheduler, Heuristic, TerminationEvaluator, StateManager, ActionKey>
{
}

However ActionScheduler is tied to the namespace AI.Planner.Actions.WorkaholicAgent, so in fresh projects without Otto the document example doesn't compile

Update

Ok so generating assemblies after the domain is set up will create a namespace to an agent's actions which can be used in your BaseAgent class.

I highly recommend everyone recreate Otto in their project, it is very informative.

 

We got rid of the csc.rsp files, since the assemblies are now compiled outside of the project. You should just be able to remove csc.rsp. However, you might also need to check if you have any <X>_DOMAIN_GENERATED or <X>_ACTIONS_GENERATED, etc. in your build settings defines. Remove those.

 

That definitely helped. I think some other assets got lost somewhere as I have a number of "Missing Prefab"s in the scene, and everything is black except for the UI (bars, etc) and Otto, who does not have any textures. However when I hit play, he animates, and the AI Planner does run. I recall what the scene looked like previously so I pretty much get the gist of what is happening, but I'm not sure if anyone else ran into this.

In any case, I can see the AI Planner planning and guiding Otto's actions, which is cool. Also I agree, the new interface is improved. Looking forward to people posting their processes on YouTube.

 

Seems like the repo did not pull all the large files properly. If you do a `git status` what does it show? Have you tried doing a `git lfs fetch`?

 

Well I have been rebuilding Otto in another project, it seems I must have missed something When I run I get

My Otto object first throws this.
ArgumentException: Object AI.Planner.Domains.DomainObject does not exist within the state data Domain Object: (<< 2 >>)

Then it spams this

NullReferenceException: Object reference not set to an instance of an object
UnityEngine.AI.Planner.DomainLanguage.TraitBased.BaseAgent`10[TAgent,TObject,TStateKey,TStateData,TStateDataContext,TActionScheduler,THeuristic,TTerminationEvaluator,TStateManager,TActionKey].Update () (at Library/PackageCache/com.unity.ai.planner@0.1.1-preview/Runtime/Agent/BaseAgent.cs:273)
Workaholic.Otto.Update () (at Assets/Scripts/AIPlanner/Otto/Otto.cs:32)

Which fails in BaseAgent:Update()
{

m_Controller.Update();

m_JobHandle = m_PlannerScheduler.Schedule(default);
// FIXME: Ideally, we don't complete immediately, but without doing this the entity debugger throws errors
m_JobHandle.Complete();

}

m_Controller is not defined.

Not sure where I went wrong.

 

Yeah it was clearly how I set up something. . . . I got otto doing his thing in my level so I need to try again

 

I create 2 action with 2 different operation action-action: "a1" with operation action "op1" and action "a2" with operation action "op2". Both operation action will displays "op1" or "op2" Debug.Log("op1") , Debug.Log("op2") dependent in which action was taken.Also Traits was created with one bool field "check".
action 1 precondition(field "check" == false) and effect to change field "check" = true.
action 2 precondition (field "check" == true) and effect to change field "check" = true.

In normal case the action 1 will run then follow by action2 then action 1 followed by action 2 and so on and output op1 op2 op1 op2 op1 op2.
But i get only action 1 was run continuously.
Demo Scenes in attachment.

 

The various iterations in usability have been great so far! I'm starting to get how it all works. Operational Actions are where I'm hung up at the moment too. The ones that come with Otto are very informative but they're a read for sure. Makes sense why - it just takes some time.

 

So I spent the weekend getting deep with the Otto example. . . here is what I learned:

1) A Planner action has a state machine paradigm:
BeginExecution()
ContinueExecution()
EndExecution()

2) These methods control the animator state machine (Mechanim). The animator is composed of sub-state machines at the root level like sleep/work/eat.

3) The BeginExecution() from the Action calls a trigger property to switch the animator to a sub state like sleep/work/eat.

4) Each sub state machine is composed of 3 categories or states, BeginState ( ie jump into bed), StateUpdate (sleeping loop), EndState(get out of bed)

5) Many of these states use StateMachineBehaviors (scripts attached to Animator states) to initialize, and update and an animator property.

6) These animator properties are then used in transitions from the StateUpdate to EndState in the sub state machine.

7) Many StateUpdates also transition to itself based on a boolean property.



The main point being. . . to truly understand the Otto example you need a deep under standing of Mechanim

 

Just to keep in mind.
ContinueExecution() and EndExecution() mostly depend on return value from OperationalActionStatus Status(...)

 

Thanks for posting a summary! I like that character animation/action is decoupled from the logic/planning. I need to look more into the StateMachineBehaviors, it sounds like they give a finer grain level of control? (ie, if we wanted to make Otto roll around into various sleeping positions while in the Sleep state, we could do it here?)

 

This is a great package.

I do wonder what approach should be used to modify an agent at run-time? Using Otto as an example, how should he change to a different type of work or how should he be allowed to eat additional foods at run-time?

 

Thank you, @TonyGrahn. We're glad this is useful to you.

Re: modifying an agent at run-time -- it depends on what you want.

If you want Otto to eat an additional type of food, then it would be possible to modify the original ConsumableType enumeration to include another type of food. You'd need to add a new Dispenser game object in the scene and model it after the FoodDispenser or BottleDispenser. You'd also need to add an Inventory_<NewItem> domain object to Otto as a way to keep track of these items. After doing this work and setting up any visuals in the game the current actions should work fine.

Now, if you're talking about a mode-switch where Otto would behave in an entirely different way or do a different type of work, then you'd do a bit more work to create a separate planning definition and add multiple Otto scripts to the AI_Prefab GameObject. Then you could toggle between the two depending on what mode you wanted. However, I wouldn't suggest doing this with the current version. We're about to introduce a better workflow that should enable just that in a much easier way.

Does this answer your question?

 

Yep, that sounds like a good use of a state machine where you want specific, designed behavior without a need for decision making.

 

Yes, this very much answers my question. I noticed that trait, enum, and action definition assets were compiled from generated C# at build time and was worried this might make large behavior changes impossible at run-time. But if I understand you correctly, creating separate planning definitions for every permutation of behavior and switching between them is a workable solution. I also see that custom reward functions are possible so if those are going to be in the final version, maybe they could be useful for this too.

Great work, I understand it is still under very active development and you're working on workflow so I know not to get too emotionally attached to code I hack together playing with Otto.

 

can't find it

 

What version of Unity are you using?

 

2018.4.0f1

 

also I tried 2019.3b4 (which shows the package) and got this error

Code (CSharp):

1. An error occurred while resolving packages:
2.   Package com.unity.entities@0.1.1-preview has invalid dependencies:
3.     nuget.mono-cecil: Resolved version [0.1.5-preview] does not satisfy requested version [0.1.6-preview]
4.  
5. A re-import of the project may be required to fix the issue or a manual modification of D:/PROJECTS/Unity AI Planner/Packages/manifest.json file.
6.  

 

found out that entities 0.1.1 ain't compatible with this, gotta downgrade to 0.1
I thought the package manager handled these?

 

v0.1.1 requires 2019.2

 

This is very interesting tool!

One question about its future direction: will it be ever possible to load plan definitions at runtime from a text file, rather than requiring the editor and code generation? For context: we make games that are heavily procedural and rely on everything getting loaded from text files at startup, for moddability as well as to make scripting easier on the design side (especially in AOT environments where recompilation is painfully long - but also, scripting in text makes collaboration much easier).

Unfortunately it doesn't look like that use case is currently supported by this package, because of reliance on codegen, and I'm wondering if that's something that's on the roadmap, or if that's a use case that is going to be unsupported?

 

It seems can't input float value into this field. It auto changed to Int value, even if I copied a float here.
PS: Target.Progress is float type.

 

Thank you for the details! Agreed that live codegen would be bad, specifically since it's not an option on AOT platforms. Being able to define actions entirely through text files, load them up at runtime, and have the planner use those instead of codegen, would be highly desirable.

Looking forward to future developments!

 

Great, thank you !

 

Hi, another question, my current working project is hybrid ECS based. Is AI Planner working well on it?
I am learning the otto example, it seems DomainProviders register GameObjects to WorldDomainManager, but ECS World is not GameObject based.

 

Is there a tutorial for this?

 

Is it plan for allowing action to be used by multiple different agents without using inheritance ?

 

The documentation and Otto project are very helpful resources, but it might be helpful (just speaking for myself at least) to see someone go through the process of setting up something simple like Otto from the ground up, maybe in a video. Certainly there is a lot of code written for the operational actions which would be tedious to watch someone type out, but if there could just be a "here is where I started, here is what I did next and why, here are some things to keep in mind" type video, that would be very useful to me. All that said, I know that AI Planner is still in flux so it might be asking for trouble to make anything quite that concrete at this point.

I think the operational actions are where I struggle the most, simply because the ones with Otto are a little bit intricate. Maybe a better compromise would be more detail about operational actions in general? Like a workflow for making them, or even just some best practices.

Another suggestion might be making/publishing a different demo scene to illustrate a different use of the tool, though I fully admit that even a basic 2D demo with crude primitive shapes would probably take time that your team doesn't want to necessarily spend writing more example code that other users would invariably expect you to maintain.

Or, here's another idea - maybe instead of one "here's how all of AI Planner works" video, maybe a short series of videos about each piece? Actions, agents, etc. Maybe each video shows a couple examples of using each component, and maybe there's a little summary in there about what code generation does and such for folks without formal CS backgrounds.

In my own example, I'm trying to make a scene which isn't all that different from Otto, but for a few things:
1. There are two agents instead of one
2. Trying to experiment with using a lot of interactable items (traits), maybe closer to 30-40

I'm particularly curious to know more about multi-agent systems, though I'm not sure if the tool is at that stage yet.