AI Systems — Behavior Trees & Blackboard

The interview question is "State machines vs. behavior trees — when and why?" The answer is not which is better — it is which fits the problem.

The answer framed around designers

The real deciding factor is often who needs to author it and what they need to control:

• State machine when: the AI has a fixed set of authored beats (patrol → investigate → alert), the sequence matters precisely, and a designer needs to control exact transitions. Boss phases are the canonical example.
• Behavior tree when: the AI needs to react to many dynamic conditions (combat, cover, flanking, retreating), behaviors need to be reused across multiple enemy types, or the behavior set will grow over time.

Unreal's Behavior Tree system evaluates a tree of nodes top-to-bottom, left-to-right each tick, selecting the highest-priority branch whose conditions are satisfied.

Node types

Composites in detail

A simple combat BT

Root
└── Selector
    ├── Sequence  [HasTarget?]          ← Decorator gate
    │   ├── Service: UpdateTargetLocation
    │   ├── MoveTo(TargetLocation)
    │   └── Task: MeleeAttack
    ├── Sequence  [HeardSound?]
    │   ├── MoveTo(SoundLocation)
    │   └── Task: Investigate
    └── Task: Patrol

Abort types on decorators

When a decorator condition changes mid-execution, abort types control what happens:

Writing custom Tasks and Decorators

// Custom task — subclass UBTTask_BlueprintBase in BP,
// or UBTTaskNode in C++
UCLASS()
class UBTTask_MeleeAttack : public UBTTaskNode
{
    GENERATED_BODY()

    virtual EBTNodeResult::Type ExecuteTask(
        UBehaviorTreeComponent& OwnerComp,
        uint8* NodeMemory) override;
};

The Blackboard is the AI's shared memory — a key-value store that the Behavior Tree reads and writes through decorators, services, and tasks. It decouples the AI's data from its logic.

Common Blackboard keys

Writing to the Blackboard

// From C++ (e.g. in a Service or AIController)
UBlackboardComponent* BB = OwnerComp.GetBlackboardComponent();
BB->SetValueAsObject(TEXT("TargetActor"), PerceivedActor);
BB->SetValueAsVector(TEXT("TargetLocation"), TargetActor->GetActorLocation());

The AAIController is the bridge between the Behavior Tree and the Pawn. It possesses an AI Pawn, runs the BT, owns the Blackboard, and owns the AI Perception component.

UCLASS()
class AEnemyAIController : public AAIController
{
    GENERATED_BODY()

protected:
    UPROPERTY(EditDefaultsOnly)
    UBehaviorTree* BehaviorTree;

    virtual void OnPossess(APawn* InPawn) override
    {
        Super::OnPossess(InPawn);
        RunBehaviorTree(BehaviorTree);
    }
};

AI Perception

The UAIPerceptionComponent handles sight and hearing — it fires delegates when the AI sees or hears something, and the AIController updates the Blackboard in response.

// Bind to perception events in AIController
PerceptionComponent->OnTargetPerceptionUpdated.AddDynamic(
    this, &AEnemyAIController::OnPerceptionUpdated
);

void AEnemyAIController::OnPerceptionUpdated(
    AActor* Actor, FAIStimulus Stimulus)
{
    if (Stimulus.WasSuccessfullySensed())
    {
        GetBlackboardComponent()->SetValueAsObject(
            TEXT("TargetActor"), Actor);
    }
}

EQS lets AI ask spatial questions of the world — "What is the best cover position?", "Where can I flank the player?", "Which patrol point is closest and not visible to the player?" — and get back a scored set of candidates.

How it works

1. Generator — produces candidate points (grid around actor, patrol points, actors of a class)
2. Tests — score and filter candidates (distance, line of sight, dot product to target)
3. Result — the best scoring candidate is stored in the Blackboard and used by the BT

Running EQS from a BT Service

The most common pattern: a BT Service runs an EQS query on a timer and writes the best result to the Blackboard. The Behavior Tree then uses that location for movement tasks — no per-frame spatial logic in C++ required.