Confidence#
How sure your AI is that you're there, from 0 to 100%. Never just "seen" or "not seen".
- Every AI holds a confidence from 0 to 100% for each thing it perceives.
- It climbs while the AI can see you and fades when it can't. Distance, light, cover, stance, the angle and the weather all count.
- At set points your Blueprint gets an event: Suspected at 15%, Detected at 35%, Tracked at 60%.
Seven states, one event each#
As confidence rises and falls, the AI moves through seven states. Each step fires a Blueprint event, so your AI reacts in stages instead of all at once. Steps down fire too: a Tracked target that slips back to Detected fires On AI Detect again.
| State | Event | A typical reaction |
|---|---|---|
| Suspected | On AI Suspect | Turn the head, a "Hm?" bark, walk over slowly |
| Detected | On AI Detect | Alert, call it out, take cover |
| Tracked | On AI Track | Chase, shoot, flank |
| Lost | On AI Lost | Search: read Get Loss Reason to choose how |
| Remembered | On AI Remember | Patrol near the last known spot |
| Expired, or back to Undetected | On AI Forget | Back to normal |
The 15, 35 and 60% lines are the profile's Suspect Threshold, Detect Threshold and Track Threshold. A jumpy civilian and a calm soldier can run the same Blueprint with different numbers.
What changes it#
Everything vision checks is multiplied together. If any one thing is zero (too far, outside the cone, fully behind cover), the AI doesn't see you at all. Each clip below is the real demo room: the result card shows what the guard was sure of.
Distance#
Far away is a guess; close up is certain. The drop is gentle at first and steep near the end: at half the range 75% is left, at full range nothing.
Setting: Vision Max Range, shaped by Vision Distance Curve if you assign one.
Light and shadow#
APS knows how lit you are where you stand: in shade a guard sees you about three times worse than in the sun.
Settings: tick Use Per Target Light (Trace Sun Shadow under it is on already), and call Set Sun Direction once from your level. More on World & fairness. Darkness Min Detection is how much it still sees in total darkness.
Cover#
APS checks 5 points on you: head, chest, pelvis and both shoulders. The more of you it can see, the surer it gets. Lean your head over cover and it starts to notice; duck fully and it can't see you at all.
Stance#
Crouching helps only behind cover. Set how many of your points must show before you count, per stance: Min Visible Points Standing, Min Visible Points Crouched and Min Visible Points Prone (all 1 by default, so stance changes nothing until you raise them: 2 crouched and 3 prone is a good start).
Vision cones#
Three cones, checked in order:
| Cone | Angle | Strength | Range |
|---|---|---|---|
| Focal | 60° each side | full | full |
| Peripheral | 110° each side | 45% | 60% |
| Rear | 60°, from behind | 25% | 35%, only targets moving faster than 3 m/s |
The focal cone is always on. The peripheral and rear cones are optional: tick Enable Peripheral Cone and Enable Rear Motion Cone on the profile. The focal angle is Vision Half Angle Deg.
Eyes in the head#
The guard sees where its head looks, not where its body points. Glance-around animations really move its sight, and a turn takes time, so it can't see you before it has turned to look.
Settings: Use Eye Socket with Eye Socket Name (for example head), Eye Direction Mode, and Eye Turn Rate Deg Per Sec.
Glass, leaves and smoke#
Each surface lets through its own share of sight: glass all of it, foliage and smoke part of it, walls none. Hide in tall grass and you are partly visible, not all or nothing.
Setting: Surface Vision Transmission maps a physical material to the share it lets through.
Fog and weather#
Fog shortens how far a guard can see. Put an APS Environment component in a smoky room, a foggy valley or a dark cellar: inside it, sense ranges and light change. Set Weather Modifier changes visibility everywhere at once.
Fair spotting#
A guard that has never seen you needs a moment to believe its eyes: a quick dash across its view is forgiven. Once it knows you, it re-spots you at once.
Setting: First Spot Reaction Time. It is 0 (off) by default; 0.3 s feels sharp but human. On AI Telegraph fires a moment before it commits, for a "huh?" that warns the player.
Read it in Blueprint#
| You want | Use |
|---|---|
| React at each stage | The events above, on APS Core (or an APS Perception Listener) |
| A detection meter | Get Belief Data (Target) → Smoothed Confidence, 0 to 1 |
| See it while you play | Tick Show Visualizer on APS Core |
Go deeperThe vision formula
Every factor multiplies, so any single zero means no detection:
Confidence = DistanceFalloff × AngleFalloff × ExposureRatio × LightFactor
× MotionBonus × WeatherMod × VisibilityMultiplier × ConeScale| Factor | By default |
|---|---|
| Distance falloff | 1 − (distance / range)²: 75% at half range, 0 at full range |
| Angle falloff | cos(angle × 90°): full at the centre, zero at the cone's edge |
| Exposure | The weighted share of your sample points with a clear line |
| Light factor | Lerp(Darkness Min Detection, 1.0, light level) |
| Motion bonus | Up to 30% more for a target sprinting at 600 cm/s or faster |
| Weather | The world's weather modifier |
| Visibility multiplier | Your target's camouflage (APS Target Component) |
| Cone scale | 1.0 focal, 0.45 peripheral, 0.25 rear |
Assigning Vision Distance Curve, Vision Angle Curve or Motion Bonus Curve replaces the default shape. The angle is measured to your head (the first sample point), not your body's centre.
Go deeperThe five points, and their weights
| Point | Weight |
|---|---|
| Head | 1.0 |
| Chest | 0.9 |
| Pelvis | 0.7 |
| Left and right shoulder | 0.6 each |
Exposure is the weighted share of points with a clear line. Head and one shoulder clear, the rest blocked: (1.0 + 0.6) ÷ 3.8 = 42%. Unreal's own AI Perception traces only to the centre of your body, so it would report not seen.
The points come from the target's APS Target Component (5 by default, following its animation), otherwise from the profile's Default Visibility Samples, otherwise from a built-in set limited by Vision Sample Count. Min Exposure to Register ignores slivers. A point counts toward the stance limit only when more than half of the sight reaches it.
Go deeperHow several senses combine
Senses don't average. The strongest one sets the level and every other active sense adds a small bonus:
Fused = (best active sense × its weight) + (other active senses × Corroboration Bonus)Vision at 80% plus hearing and smell gives 80% + 2 × 10% = 100%: seeing you and hearing you makes it surer than either alone.
Go deeperThe thresholds and timings
| Setting | Default | What it does |
|---|---|---|
Suspect Threshold | 0.15 | Enter Suspected |
Detect Threshold | 0.35 | Enter Detected |
Track Threshold | 0.60 | Enter Tracked |
Min Track Duration | 0.5 s | Time in Detected before it can track: no instant lock-on |
Telegraph Threshold | 0.25 | When On AI Telegraph fires: keep it between Suspect and Detect |
First Spot Reaction Time | 0.0 s | On first sight only, how long it must see you before confidence builds |
Vision loses you quickly: when it stops seeing you, its share drops at once rather than fading. Its loss reason says how: Occluded behind something, OutOfRange beyond its range or cone, SensorDropout when it was too dark, cloaked or blinded.
Change it from script#
Set Target Confidence (Target, Confidence) raises the AI's belief once, at any range: a tip-off, an alarm, "it heard the shot". It never lowers belief, and afterwards the belief fades as usual unless a sense keeps it up.
For anything that should hold, or push belief down, use evidence: Add Target Evidence (Target, Evidence Tag, Operation, Value, and under the arrow Lifetime and Source).
| Operation | While it holds |
|---|---|
| Raise (floor) | Belief is at least Value. Gives the AI a belief about the target if it had none. |
| Clamp (ceiling) | Belief is at most Value, even while it sees you: a disguise that keeps every guard at Suspected |
| Lower (relative reduction) | Belief settles Value below what the senses alone say: doubt rather than a hard limit. Only while a sense reports the target; once all are silent, normal fading takes over. |
| Invalidate (force zero) | Belief is forced to 0 and the target steps back down: "you didn't see anything" |
- Adding the same Evidence Tag again replaces the entry, so re-applying a disguise refreshes it instead of stacking.
Lifetime0 holds until Remove Target Evidence (Target, Evidence Tag) or Clear Target Evidence (Target); above 0 it expires by itself.- Several pieces resolve together: the highest floor, the lowest ceiling, the reductions added up, and a ceiling always beats a floor.
- Get Target Evidence (Target) lists what is applied, with its
Source, so you can see why belief sits where it does.
The player puts on a guard's uniform
└─► for each guard: Add Target Evidence (Player, "Disguise", Clamp (ceiling), 0.15)
The player is seen taking it off, or attacks
└─► for each guard: Remove Target Evidence (Player, "Disguise")APS and Unreal's AI Perception#
| Unreal AI Perception | APS | |
|---|---|---|
| The result | Seen or not seen | 0 to 100%, through seven states |
| Line of sight | One trace to the body's centre | Five points: how much of you shows |
| Cover | Head over a crate: still "not seen" | Partly visible, partly sure |
| Light | Not taken into account | Light and shadow where you stand |
| Cones | One cone | Focal, peripheral and rear, plus the head's direction |
| Surfaces | Blocks or doesn't | Glass, leaves and smoke each let some sight through |