AI Planner - v0.1.1-preview Release!

We will have a variety of smaller samples soon. We already have these internally, but they require more documentation. Re: the link you've provided -- I believe if you can map the steps over to the planner that you should be able to re-create that example with the AI Planner.

 

The planning algorithm isn't ECS, but rather ECS-friendly. More importantly, it is jobified (i.e. threaded), so that it would be future-proofed to work with ECS. We have not done any rigorous performance comparisons, but it is something we hope to do and to compare it to other approaches and/or other planners.

The planner algorithm is not A*, but rather a bounded heuristic-based graph search. However, they do share the same computation complexity order. It's hard to compare it with the forum post you mention because it isn't an apples-to-apples comparison and would require a deeper dive into that implementation.

GOAP makes use of A*, but has a predicate-based state representation. So, to your question about it being heavier, the answer unfortunately is one of "it depends". The choice of state representation and the calculations you do in the heuristic are what factors in.

 

For your own EnemyAIAgent are you attempting to re-use the same domain that Otto uses or are you creating a new one? Do you have an assembly definition (i.e. asmdef) that your EnemyAIAgent.cs file is contained within?

 

I think there is a bug in the ConditionalCompilationUtility that we use. Can you try adding the defines to your project settings directly:

PLANNER_DOMAIN_GENERATED
PLANNER_ACTIONS_GENERATED

If that doesn't stop the continual compilation loop, then go into Packages/com.unity.ai.planner/Runtime/Utility/ConditionalCompilationUtility.cs and add a return at the beginning of the static constructor (line 52)

 

Explicit support for modeling other agents is not currently in our roadmap, as the forms and uses of such predictive models are rather domain/game-specific. It's unlikely that we'd be able to provide a universal solution, particularly as the scale of the problem can vary greatly. For example, should the model predict (1) the effects of the other agent's current action, (2) the effects of each stage of the other agent's current plan, or (3) the possible set of plans the other agent may switch to in the process of interacting with other agents/the game/etc? From the perspective of agent A, should the model of agent B include B's model of agent A, forming a recursive modeling loop? It depends on the application, as well as the ability to tractably solve such problems at sufficient scale, unfortunately.

That said, I believe we have an opportunity to release small samples which illustrate various strategies for handling the uncertainty of other agents' actions, including that of human players. I'll give a brief overview of how you might tackle this problem with our current tools:

In our trait-based language, you can model another agent as an object with an "Intent" trait, which can hold data such as an enum of possible intents (e.g. steal an object, attack an agent, hide, etc) as well as other metadata you might want to include. Then, you can condition the effects of your agent's actions to add additional effects of other agents accordingly (see: ICustomActionEffects).

If you want to share intent across agents, you might consider a blackboard or some other solution in which agents can publicly post intent to be consumed by other planning agents. If you want some uncertainty over intent---such as when modeling a human player---you may consider associating each intent with some probability and assign probabilistic outcomes to each action (e.g. 30% player consumes a potion, 70% attacks an agent). You can take this one step further by creating a module which attempts to infer player intent, adjusting the intent probabilities accordingly.

 

We are putting the final touches on the next release right now, actually. Should be out within a week.

The release cycle window is largely dependent on the scale of the new features. This next update has changed the package quite considerably, as we wanted to improve the authoring experience quite a bit. Internally, we've been working with customers in order to receive feedback, try new features, fix bugs, etc. The following release will likely be smaller in scale and have a shorter timeline.

The planning team is 2.5 engineers, currently (keep your eye on the Unity job board in 2020 ). We're also involved with other projects within the AI team, not all of which have been publicly discussed, so expect even more exciting news from the team in the coming year!

 

Locking this thread, since 0.2.0 is the new release: https://forum.unity.com/threads/ai-planner-v0-2-0-preview-released.789809/