Behavior Trees and Blackboards

Execution model

UE Behavior Trees are event-driven. Search/execution proceeds left-to-right and top-down; left branches represent higher priority. Blackboard observers and decorators can trigger reevaluation and abort running work.

Asset/data contract

• Blackboard keys are typed working memory: target actor, last-known location, state flags, timestamps, current Smart Object handle wrapper, or requested action.
• Use a Blackboard parent only for genuinely shared keys.
• Keep durable/shared game truth in the owning gameplay system, not only in a Blackboard.
• Name keys semantically and distinguish observation (SeenActor) from conclusion (CombatTarget).

Node responsibilities

• Composite: controls branch traversal.
• Decorator: gates execution and optionally observes changes/aborts branches.
• Task: performs one bounded action and reports success/failure/in-progress.
• Service: periodically updates data while its branch is active.

Observer aborts

• Self: abort the decorated running branch when its condition fails.
• Lower Priority: abort branches to the right when this higher-priority condition becomes valid.
• Both: applies both behaviors.

Use the smallest abort scope that expresses intent. Broad aborts create churn and can repeatedly cancel expensive actions.

Task contract

1. Validate controller, pawn, keys, and required subsystem.
2. Start the bounded action.
3. If latent, retain only the handles/state required for completion and return In Progress.
4. Complete exactly once with success or failure.
5. On abort, cancel delegates/timers/moves/abilities/claims and release resources.
6. Prevent stale callbacks from finishing a newer task instance.

Prefer native Tasks/Services in measured hot paths; Epic notes Blueprint versions are less performant.

Services

Set an interval and random deviation from reaction needs and population budget. Do not use a service as disguised per-frame Tick. Prefer Perception/Blackboard events for facts that already publish changes.

Common architecture

Selector
  Combat (Target valid) [abort lower priority]
    Sequence: choose position -> move -> execute action
  Investigate (LastKnownLocation set)
    Sequence: move -> search -> clear
  Patrol
    Sequence: choose point -> move -> wait

Keep sensing outside the tree where possible; the tree reacts to well-defined working memory.

Debug acceptance

• live Behavior Tree shows the expected active branch;
• Blackboard values change at the responsible event;
• aborts cancel latent work and release claims;
• every task result is handled;
• no high-priority condition causes search/restart oscillation;
• worst-case agent count stays inside the AI frame budget.