AI Behavior Trees Utility AI skill

Behavior Trees & Utility AI

Two complementary ways to structure NPC decision-making, plus how to combine them. A behavior tree (BT) expresses structured, prioritized, reactive logic as a tree that is "ticked" each step. Utility AI answers "how much do I want each option right now?" by scoring actions with normalized curves and picking the best. Ship believable agents by using a BT for structure and Utility AI where graded trade-offs matter.

This skill is the implementation companion to game-ai (which helps you choose between FSM / BT / steering / pathfinding). Read game-ai to pick a model; read this to build the runtime.

When to use

• Use to build a reusable BT runtime: a Blackboard, Node base, action/condition leaves, Sequence/Selector/Parallel composites, and decorators (Inverter, Cooldown, Repeat).
• Use to build a Utility AI decider: response curves, considerations, and an evaluator that scores and selects actions (max, softmax, or weighted-random for variety).
• Use to build hybrid AI — a BT whose leaf delegates the "which attack / which target" choice to a utility evaluator.

When not to use: to choose between FSM, BT, steering, or pathfinding, and for A*/navmesh routing, use game-ai. For Unreal's asset-based BehaviorTree/Blackboard, BTTask/BTService and AIController, use unreal-behavior-trees. For the navmesh agent that moves the NPC, use unity-navmesh or the engine's navigation node.

Core workflow

1. Pick the model. Structured, prioritized, interruptible behavior → BT. Continuous "score every option" decisions (targeting, needs, item choice) → Utility. Both → hybrid.
2. Design the Blackboard first. One typed key/value store per agent is the shared memory that decouples nodes; leaves read/write it and never hold references to each other.
3. Write leaves. Conditions return Success/Failure immediately; actions return Running across frames until they finish. Keep leaves small and side-effect-explicit.
4. Compose. Selector = OR/fallback (first non-failure wins); Sequence = AND (stop at first non-success); Parallel for concurrent branches. Wrap with decorators for policy (invert, cooldown, repeat, force-success).
5. For Utility: enumerate considerations, map each raw fact through a normalized 0..1 curve, combine (weighted product with compensation, or weighted sum), then select the max — add hysteresis so agents don't flip-flop on ties.
6. Tick deliberately. Tick the tree/evaluator once per decision step (often slower than render). Preserve Running state between ticks; verify by drawing the active path and the per-action scores on screen while tuning.

Architecture at a glance

A behavior tree evaluates top-down, left-to-right; each node returns a status up to its parent:

flowchart TD
    Root["Selector (root)"] --> Combat["Sequence: Combat"]
    Root --> Patrol["Action: Patrol"]
    Combat --> See["Condition: CanSeePlayer?"]
    Combat --> InRange{"Selector: Reach"}
    Combat --> Attack["Action: Attack (Running)"]
    InRange --> Close["Condition: InAttackRange?"]
    InRange --> MoveTo["Action: MoveToPlayer (Running)"]

Utility AI is a scoring pipeline — every candidate action is scored, then one is selected:

facts (distance, health, ammo…)
      │  each fact → a normalized 0..1 response curve (consideration)
      ▼
score(action) = weight · combine(consideration_1 … consideration_n)   # product+compensation or sum
      ▼
select: argmax  ·  or softmax / weighted-random for variety  ·  + hysteresis to avoid jitter

Status is a three-value enum shared by every node — this is the contract that makes the tree composable:

public enum Status { Success, Failure, Running }

public abstract class Node
{
    public abstract Status Tick(Blackboard bb, float dt);
    public virtual void Reset() { }   // idempotent cleanup after completion or abort
}

// Selector = fallback/OR: return the first child that is not Failure.
// `Current` (from Composite) remembers a Running child so the next tick resumes there.
public sealed class Selector : Composite
{
    public override Status Tick(Blackboard bb, float dt)
    {
        for (; Current < Children.Count; Current++)
        {
            var s = Children[Current].Tick(bb, dt);
            if (s == Status.Running) return Status.Running;            // resume here next tick
            Children[Current].Reset();                                // terminal child can re-enter cleanly
            if (s == Status.Success) { Current = 0; return Status.Success; }
        }
        Current = 0;
        return Status.Failure;                                         // every child failed
    }
}

The reciprocal Sequence (AND — stop at first non-Success), the reactive variants that let a higher-priority branch interrupt a running one, Parallel, the Blackboard, the leaf base classes, and every decorator are in references/behavior-tree-core.md. The leaves use Unity's Vector2 member names (magnitude, normalized, sqrMagnitude); in Godot C# use Length(), Normalized() and LengthSquared().

Utility scoring in one snippet

// A consideration maps one raw fact to 0..1 through a response curve.
float Score(Blackboard bb)
{
    float distance01 = Curves.InverseLerp01(bb.Get<float>("distToPlayer"), 20f, 2f); // near = 1
    float health01   = Curves.Sigmoid(bb.Get<float>("health01"), k: 8f, mid: 0.4f);  // hurt = low
    // A 0 in any factor still vetoes the action; compensation stops each extra
    // consideration from dragging every score toward 0. No allocation per call.
    return Curves.Compensate(distance01 * health01, count: 2);
}

The full curve library (linear, quadratic, exponential, logistic/sigmoid, smoothstep), the Consideration/UtilityAction types, and the UtilityEvaluator selection strategies are in references/utility-ai-system.md.

Pitfalls

• Resetting a Running subtree every frame restarts it. Return Running and preserve its state until completion or abort. Parents reset terminal children immediately; BehaviorTree resets the terminal root. Make cleanup idempotent for ancestor resets.
• Deep trees re-evaluated wholesale each tick waste time and cause thrash. Prefer shallow trees and conditional aborts (a higher-priority condition can interrupt a lower branch).
• A guard condition stops being checked once the action after it is Running. A memory Sequence resumes at the running child, so [CanSeePlayer?, Chase] keeps chasing after sight is lost. Use a ReactiveSequence (re-checks from the first child each tick) or have the action itself fail when its blackboard input disappears.
• A cooldown that clears on Reset() lets an agent bypass it by flickering between branches. Keep the ready-time when the branch is aborted.
• Un-normalized considerations. If one curve outputs 0..100 and another 0..1, the big one dominates. Every consideration must return 0..1.
• Utility jitter on near-ties. Add hysteresis: give the currently-running action a small bonus so the agent commits instead of oscillating.
• Allocating nodes, closures, or arrays every tick creates GC spikes. Build the tree once at spawn; keep per-tick work allocation-free.

References

• references/behavior-tree-core.md — Blackboard, Node/leaf base classes, action & condition leaves, Sequence/Selector/Parallel, and the decorator library (full C#).
• references/utility-ai-system.md — response-curve library, Consideration, UtilityAction, and the UtilityEvaluator (argmax, softmax, weighted-random, hysteresis).
• references/practical-examples.md — a guard Patrol→Combat BT, a villager needs-based Utility AI, and a hybrid agent, as drop-in templates.
• references/best-practices-and-pitfalls.md — memory management, profiling, avoiding deep trees, event-driven aborts, and combining Utility AI with BTs (hybrid architecture).

Related skills

• game-ai — choose between FSM / BT / steering; A* and navmesh pathfinding.
• unreal-behavior-trees — Unreal's asset-based BT/Blackboard, tasks, decorators, services.
• unity-navmesh — the NavMeshAgent that carries out "move to" intents.
• physics-tuning — agent radius, movement, and collision response for the motion layer.
• tower-defense, fps-shooter, rpg — genres that compose this decision layer.