How APS thinks#

Read this once and the rest of APS explains itself. It is the model every page builds on.

The system in one picture#

1. GatherWho could it perceive? Every Pawn in range, plus actors you registered.
2. SensesVision, hearing, smell, touch, vibration, damage, pain, echolocation.
3. FusionBest sense plus corroboration.
4. Memory engineSmoothing, decay and the seven states.
5. Belief recordOne per target: everything it believes.
6. BrainWhere you are, threat, emotions, squad sharing.
7. Events and nodesFor your Blueprint or Behavior Tree.

The shape to remember: senses produce numbers, the core turns numbers into belief, belief fires events, and your code decides what to do. Everything is set on one asset per kind of AI, the Perception Profile.

1. Confidence#

Every sense, on every perception tick, gives a number from 0 to 1 for every target it can perceive: how strongly that sense believes the target is there right now.

• Vision at 5 m in bright light, target fully in view: around 0.9.
• Vision at 30 m, only a shoulder showing past cover: around 0.15.
• Hearing a footstep two rooms away through a wall: around 0.2.
• Being shot: around 1.0.

Confidence is not distance and it is not visibility. It already accounts for distance, angle, light, cover, how fast the target moves and the sense's own quirks. The Confidence page shows each of these in the demo.

2. How senses combine#

Senses don't average and they don't multiply. The strongest active sense sets the level, and each other active sense adds a small flat bonus.

Two rules follow, and they matter:

• A silent sense costs nothing. Giving a guard smell can't make it worse at seeing. Senses only ever add.
• Agreement is what makes multi-sense AI feel smart. Something heard and half-seen is surer than either alone, as it is for people.

Each sense's weight is in the profile's Sense Weights: a guard dog might have Smell 1.5 and Vision 0.6.

Rise, hold, decay#

The fused value is smoothed into Smoothed Confidence, and that is the value everything else uses: the states, the events, your meter.

The rule that surprises people: while any sense is active, confidence never falls. Ray sampling makes the raw value dip a few percent between ticks; holding stops that noise from making the AI flicker.

Smoothed = Lerp(Smoothed, Fused, clamp(ConfidenceRiseRate × dt, 0, 1))
Smoothed += DwellRate × Fused × dt
DwellRate = Lerp(StandingDwellRate, MovingDwellRate, clamp(TargetSpeed / 600, 0, 1))

no curve:    Smoothed ×= exp(−DefaultDecayExponent × SenseDecayMultiplier × dt)
with curve:  Smoothed = Lerp(Smoothed, LastActiveConfidence × DecayCurve(TimeSinceLastSensed),
                             clamp(SenseDecayMultiplier × dt, 0, 1))

3. The seven states#

Each target sits in exactly one of seven states, and each change fires a Blueprint event. The Confidence page plays them live.

The event fires on every entry into a state, going up or down. When a Tracked target steps back to Detected, On AI Detect fires again; dropping back to Undetected fires On AI Forget. Keep a "first time" flag in your Blueprint if you only want the first one.

Grace time and direct cut#

When every sense goes quiet, the AI doesn't drop you at once. Each sense has a grace time, and says whether losing it may cut straight to Lost:

Direct Cut jumps straight to Lost. Without it, confidence steps down through Tracked, Detected and Suspected first, which reads as the AI slowly losing certainty. All of these are on the profile's Loss settings (Vision Loss, Hearing Loss and so on).

Which sense's row applies: the sense that had the target last, with one overlap. While Vision (or Echolocation) still perceives the target, its row applies even if another sense leads. When it loses the target but hearing still has it, vision's row is kept for vision's own grace time, so a moment's overlap doesn't turn ducking behind cover into a slow fade. After that, hearing's row applies: its 3 s grace and its gradual fade.

4. Why it lost you#

When a target goes Lost, APS records why. Read it with Get Loss Reason and branch your search on it.

The reason comes from the last sense that had the target. If a guard could still hear your footsteps fading after you slipped out of sight, the loss is SoundFaded, not Occluded. Vision answers from what blocked its last look: a wall, too few visible points or too little exposure is Occluded; distance or the edge of the cone is OutOfRange; a target in view but too faint, cloaked or blotted out by pain is SensorDropout.

This one branch is the difference between an AI that runs to the corner you hid behind and one that wanders in a circle. The Memory page shows it in the demo.

Get Loss Direction is recorded at every loss: the target's measured direction of travel at that moment. It is zero when the target wasn't moving, so check Is Nearly Zero before using it.

5. Awareness#

The states are per target. Awareness is per AI: one level from its top target's confidence.

Awareness reuses the state thresholds, so tune those and awareness follows. Use it for alert music, weapon-ready poses, a HUD meter: Get Awareness Level, Get Awareness Level for Target, and On AI Awareness Changed.

6. Attention: the target it commits to#

An AI can hold beliefs about many targets at once (up to Max Tracked Targets, 16 by default). It can only chase one.

• Get Top Target is whoever scores highest this update. It can flicker when two are close.
• Get Attention Target is the one the AI commits to. It is sticky on purpose:
  • Attention Stickiness Time (2 s): the least time before it will even consider switching.
  • Attention Switch Threshold (0.2): how much better a rival must score to take its attention.

Once the stickiness time has passed, attention also moves to the top target when its own target is Lost, Remembered or Expired, however small the gap. You can force it at any time with Force Attention Target (when the player shoots this AI, say). On Attention Changed gives the old and new target so you can re-plan.

Only targets the AI has noticed (Suspected or more) take attention or count as the top target; one that falls back to Undetected lets attention go. A focus set with Force Attention Target on someone not noticed yet holds for the stickiness time.

SortScore = Confidence × Sort Weight Confidence   (0.5)
          + Recency    × Sort Weight Recency      (0.3)
          + Proximity  × Sort Weight Proximity    (0.2)

7. Where it thinks you are#

Even with every sense silent, APS keeps a model of where you are, and it grows less sure over time.

The uncertainty radius is a ready-made search radius: feed it to an EQS query or Get Random Reachable Point in Radius and the search widens the longer you stay hidden.

Prediction is off by default, because for a slow patrolling guard it makes the AI feel psychic. Turn it on for pursuers. With it off, the predicted position always equals the last known one.

Predicted = LastKnown + v·T + ½·a·T²,   T = clamp(Prediction Horizon, 0.1, Max Prediction Time)

8. Memory#

Two layers, plus a diary:

• Short term: the smoothed confidence fading after a loss (the slide from Lost to Remembered), shaped by Decay Curve, or Default Decay Exponent without one.
• Long term: Long Term Memory Strength, a second number on the belief (read it with Get Belief Data). It builds slowly while the AI senses you and fades far more slowly than confidence. APS itself doesn't act on it: read it to keep an AI twitchy for a minute after an encounter.
• Episodes: the last 5 encounters with each target, read with Get Recent Episodes or Get Last Episode: how it was lost, where, which way it was looking, which way you ran. One is written each time it loses you straight from Detected or Tracked (by default a Vision or Echolocation loss).

Memory Refresh Bonus gives a head start when it finds you again, and Min Time in Lost makes it search for at least that long before giving up. The Memory page covers all of it, with the demo's Memory wing.

9. Threat and emotion#

Threat is a score per target, from 0 to 1, mapped onto None, Low, Medium, High, Critical. It combines how sure the AI is, how much damage the target has done, how many senses agree, who it is, and whether it is in contact right now.

Emotion is five values on the AI itself, each 0 to 1: fear, aggression, curiosity, alertness, panic. Each rises and falls with what is happening; the strongest one above Emotion Active Threshold is the dominant emotion. Set every emotion's cap to 0 and a robot feels nothing, without touching code. The Emotions & pain page explains every rule.

ThreatScore = SmoothedConfidence          × Threat Weight Confidence       (0.35)
            + clamp(DamageReceived / 100) × Threat Weight Damage Received  (0.30)
            + (ActiveSenses / 8)          × Threat Weight Senses           (0.15)
            + RelationshipModifier        × Threat Weight Relationship     (0.10)
            + ApproachModifier            × Threat Weight Approach         (0.10)

10. The perception tick#

APS doesn't run every frame. The component ticks, but perception only runs every Base Update Interval (0.1 s), multiplied by the AI's level of detail (below) and by any Quality Profile.

1. GatherCandidate targets, at most 5 times a second.
2. Prepare sensesAccumulators, contacts, health.
3. Own sensesPain fires its event.
4. Each targetEvery sense, then pain dulling, fusion, fairness rules and evidence.
5. MemorySmoothing, decay, state changes.
6. Where you areThe spatial model.
7. ThreatAnd relationships.
8. Player modelWhat it learns about you.
9. BrainEmotions, awareness, attention, squad sharing.
10. EventsYour Blueprint events fire.
11. HousekeepingEvery `Maintenance Interval` (0.5 s).

Two consequences:

• Events fire on the perception tick, not the frame. At 0.1 s an event can be up to 100 ms late. That is fine for AI behaviour, and it is why APS scales.
• Perception runs on the server only. Clients skip it; see Multiplayer.

11. Beliefs about places#

A belief doesn't have to be about someone. A gunshot with no known shooter, a body, a forced lock: these are beliefs about a place, with the same states and their own confidence.

• Emit Sound at Location and discovered evidence create them for you.
• Report Location Belief (location, tag, confidence, accuracy) creates one from your own logic. Reports close together with the same tag merge into one.
• On Location Belief Changed fires on every change, with the location, tag, state and confidence. Get Strongest Location Belief and Get Location Beliefs read them back.

Places never take the AI's attention, change its awareness or stop On AI All Clear, so a noise can't steal focus from a person or keep the AI from standing down.

12. Level of detail#

Every AI is scored for significance, 0 to 1, each subsystem tick, and the score picks its level of detail (its tier). No setup needed.

Two things follow, and both surprise people:

• An idle AI never reaches tier 0. With nothing going on its score tops out at 0.6: tier 1 out to 50 m, tier 2 to 150 m, tier 3 to 225 m, suspended beyond.
• An engaged AI almost never drops out of tier 0. Once it is Suspicious or more alert, it stays at full rate out to about 260 m, and tier 1 beyond. A chase doesn't stutter because the camera is far away.

Viewer means every local player camera, so split-screen and listen servers work; a dedicated server has no viewer and scores every AI as if at zero distance. Tiers also filter sounds: a sound type whose Max LODTier is 1 is ignored by a tier 2 AI. Replace the score with your own APS Significance Policy: see Performance & crowds.

Significance = 0.6 × (1 − distance to the nearest viewer / 30000 cm)
             + 0.55 if the AI is Suspicious or more alert
             + 0.25 × (awareness level / 4)

13. Who it can perceive#

• Pawns are automatic: no registration needed.
• Other actors (a security camera's target, a thrown item) call Register Perceivable Actor, or carry an APS Target Component with auto-register on.
• The gather range is the largest of Vision Max Range, Hearing Max Range and each sense's own reach (Vibration Detect Range, Echo Range and so on).
• An AI never perceives a Pawn with its own controller.
• Targets it already believes in stay on its list even out of range, so memory fades smoothly instead of freezing.
• Candidates come from a grid the subsystem rebuilds 5 times a second (15 m cells), so finding them costs the same however many AI and Pawns there are. The console variable aps.UseSpatialIndex 0 switches back to scanning every Pawn, for comparing.

14. Make it your own#