Profile settings#

Inherit every setting from this profile, keeping only the fields changed here.

A field counts as changed when it differs from the class default — exactly the fields the Details panel marks with a revert-to-default arrow. Leave a field alone and it tracks the parent; edit it and it is yours.

Start a child from a NEW profile rather than a duplicate of the parent. A duplicate carries every one of the parent's tuned values as changed fields, so it would inherit nothing and drift the moment the parent moves on.

To pin a value that happens to equal the class default, put it in Overrides below — that list is read after inheritance and always wins.

Composition is resolved once, when an agent starts or swaps profiles, into a private copy owned by that agent. Editing the parent therefore affects agents that spawn afterwards, not ones already running.

Use Show Composition to see exactly which fields are inherited.

Extra changes stamped on after inheritance is resolved.

Rarely needed on an asset — editing the field directly is the normal way to override one. This exists for values a Blueprint wants to compute, and it is the same list the component uses for per-agent tweaks.

Preset to stamp onto this profile when Apply Archetype is pressed.

Applying overwrites the fields the archetype cares about and leaves the rest alone. It is a starting point, not a mode — nothing reads this at runtime.

The profile Compare With reports differences against.

Sense classes to instantiate on this AI. Order = slot index.

A new profile starts with Vision and Hearing already here, so an AI works the moment you assign the profile. Add Smell, Touch, Vibration, Damage, Pain or Echolocation as you need them — or remove Hearing for an AI that only sees. An empty list means the AI perceives nothing at all.

Per-sense confidence weight. Missing = 1.0.

How often each sense updates, in seconds, by sense. A sense not listed uses its own: Vision 0.15, Smell 0.25, Echolocation 0.3, the others 0.1.

The least that pain or low health can dull vision to, 0 to 1. Other senses can go to 0.

Automatically route the owner's collision hits and overlaps into the Touch sense. Without this every bump has to be hand-wired from Blueprint.

Treat the AI's overlaps as a held Grab, not only its blocking hits as Bumps.

How several senses combine into one confidence value.

How dangerous a target is. The level is still derived from the Brain|Threat thresholds below, so a policy only supplies the score.

Which target the AI commits its focus to.

Vision: max sight distance (cm).

How far hearing reaches when a sound has no Sound Type of its own, and part of how far away AI look for targets at all (cm). Each Sound Type sets how far its sound carries.

Smell: max scent detection range (cm).

Vibration: max ground vibration range (cm). Passes through walls.

Echolocation: sonar pulse range (cm).

On: being hurt raises the AI's confidence in the attacker to the damage taken from them so far ÷ 100, at any range. Off: damage only fires On AI Damaged, and it adds nothing to threat or emotions either.

Half the width of the main (focal) vision cone, in degrees: 60 sees 60° to each side.

How many of the built-in sample points (head, chest, pelvis, shoulders) vision traces, 1 to 5. Only for targets without an APS Target Component and when Default Visibility Samples is empty.

How much the AI still sees in total darkness, 0 to 1: 0 is blind in the dark, 1 ignores darkness (infrared).

How high the eyes are above the actor's origin, in cm: used without Use Eye Socket, or when the socket doesn't exist.

Optional: replaces the default distance falloff. X is the distance as a share of the range (0 to 1); Y is how much confidence is left.

Optional: replaces the default falloff across the cone. X is the angle as a share of the cone's half-width (0 to 1); Y is how much confidence is left.

Optional: replaces the default movement bonus. X is the target's speed in cm/s; Y multiplies confidence (kept between 0.5 and 1.5).

Does nothing in APS 3.0. No code reads it: there is no noise floor. Use Suspect Threshold, and a Sound Filter Profile's Min Alert Level.

How much each wall between a sound and the AI muffles it: each wall multiplies the sound by e to the minus this (0.4 takes about a third off).

Does nothing in APS 3.0. No code reads it: hearing is always event-only (it hears only the sounds you emit).

Optional sound filter — which categories/alert levels this AI hears.

How fast repeated sounds build up confidence.

Seconds of silence before accumulated sound confidence starts decaying.

Confidence decay rate per second after hold time expires.

How fast smell confidence builds per second while in range.

How fast smell confidence decays per second after target leaves range.

Minimum target movement speed (cm/s) to produce detectable vibration.

Does nothing in APS 3.0. No code reads it: touch strength comes from the contact report and the contact type's multiplier.

Physics impulse magnitude that maps to Strength=1.0.

Confidence gained per second while a persistent contact is held.

Maximum extra confidence from persistent contact duration.

Echolocation only pings forward: it ignores targets more than about 101° off its facing.

Take the eye position from a mesh socket instead of Eye Height Offset: sight then starts where the head really is, as it animates.

Socket / bone the eyes sit on. Typical: "head", "eyes", "S_Eyes".

Where the cone's FACING comes from. Socket Rotation makes the cone follow the head bone.

How the socket's rotation becomes a look direction. On Automatic you can name a plain bone (head, eyes) and APS works out how it is oriented from the skeleton's rest pose: the cone starts along the character's forward and follows the bone.

A correction you set by hand, used only when Eye Socket Alignment is Manual Offset: it turns the socket's forward onto the character's forward.

With Eye Direction Mode Blended: how much of the facing comes from the socket, 0 (all actor) to 1 (all socket).

How fast the AI's view may turn, in degrees per second. 0 turns instantly (Unreal's own perception snaps to a Move To target before the body has turned). Set it near your character's real turn rate so it can't see you before it has turned to look.

Sample points traced when the target has no APS Target Component. Offsets are in the target's LOCAL space, so they rotate with it — a prone or turned target samples the right places. Leave empty to use the built-in head / chest / pelvis / shoulders set.

How much of a target must be visible (0 to 1, weighted by its points) before it counts as seen at all. 0: one visible point is enough. Raise it so a sliver of shoulder behind cover isn't seen.

How many of a standing target's sample points must be visible before it registers at all. 1 by default for every stance, so stance changes nothing until you raise these.

How many of a crouched target's sample points must be visible before it registers at all. 1 by default (stance changes nothing); raise it to 2 for stance-aware stealth.

How many of a prone target's sample points must be visible before it registers at all. 1 by default; raise it to 3 for stance-aware stealth.

The collision channels that block sight. Empty means Visibility only. A list replaces Visibility, so include it if your walls block it. Each extra channel is one more trace per point.

Bodies block sight: the Pawn channel is traced too, so squadmates hide what is behind them.

How much sight passes through each physical surface, 0 to 1: 1 fully see-through (glass), 0.4 partly (foliage, smoke). Surfaces not listed block completely.

Use a separate vertical half-angle, making the cone elliptical instead of perfectly round. Essential for games with verticality.

Half the height of the vision cone, in degrees, with Use Separate Vertical FOV.

Enable a wider, weaker peripheral cone outside the focal cone.

Half the width of the side cone, in degrees. It must be wider than Vision Half Angle Deg to add anything.

Confidence multiplier for detections made in the peripheral cone.

How far the side cone reaches, as a share of Vision Max Range.

Enable a rear cone that only registers MOVING targets — Thief's "false vision" cone. Stops players walking straight up behind an AI with impunity, without making the AI omniscient.

Rear cone half-angle measured from directly BEHIND the AI (degrees).

Target must move at least this fast (cm/s) to register in the rear cone.

Confidence multiplier for rear-cone detections. Keep low — this is a hint, not sight.

How far the rear cone reaches, as a share of Vision Max Range.

Sample light AT THE TARGET rather than using one global ambient value. Without this, a dark corner is indistinguishable from broad daylight — the global level only models day/night.

With Use Per Target Light: trace from the target toward the sun to see whether it stands in shadow. Needs Set Sun Direction on the APS Subsystem.

The light level a target in shadow gets, 0 to 1. Shadow only ever darkens.

How far the shadow trace reaches (cm).

Make the focal cone wide up close and narrow far away.

Half the cone's width up close (within Keyhole Near Distance), in degrees.

Half the cone's width at full range, in degrees.

Up to this distance (cm) the cone is at its near width; beyond it narrows toward the far width.

How the AI loses a target when vision is the sense that last had it: Grace Time (0.3 s) before it may count as lost, Direct Cut (on: straight to Lost) and how fast belief then fades (× 3).

Losing a target heard last: a 3 s grace, steps down through the states rather than cutting to Lost, normal fading (× 1).

Losing a target smelled last: a 20 s grace (scent lingers), steps down, slow fading (× 0.3).

Losing a target felt last: a 0.5 s grace, steps down, fading × 2.

Losing a target felt through the floor: no grace (stop moving and it's gone), steps down, fast fading (× 5).

Losing an attacker: an 8 s grace (it remembers being hit), steps down, slow fading (× 0.5).

Losing a target pinged last: a 0.6 s grace, straight to Lost, fading × 2.

What each extra sense that reports a target adds on top of the strongest one. Vision 0.8 alone is 0.8; with hearing too, 0.9. 0 uses the strongest sense alone.

How fast confidence rises toward what the senses report. Higher is snappier detection.

Does nothing in APS 3.0. No code reads it. Use Confidence Rise Rate, and each sense's Confidence Decay Multiplier in the Loss settings.

Extra confidence per second, while the AI keeps sensing a target standing still: the longer it looks, the surer it gets (never above what the senses report).

The same for a target moving fast (600 cm/s or more); speeds in between blend the two.

Does nothing in APS 3.0. No code reads it. Use Confidence Rise Rate, and each sense's Confidence Decay Multiplier in the Loss settings.

Optional: how much of its last confidence the AI keeps (Y) after so many seconds without sensing the target (X). Empty fades exponentially at Default Decay Exponent.

Optional: how much long-term memory strength is left (Y) after so many seconds without sensing the target (X).

How fast belief fades once no sense has the target, when there is no Decay Curve. Higher forgets faster.

Extra confidence, given once per loss, when the AI senses again a target it had Lost or Remembered: finding you again locks on faster.

Below this confidence a Remembered target Expires and its slot is freed.

Minimum seconds the AI must spend in Lost before it can transition to Remembered. 0 = no delay (default). Increase to make the AI search longer before giving up.

Distinct subjects one agent may hold memories about. The least recently written is dropped when this is exceeded.

Distinct tags per subject. The oldest is dropped when this is exceeded.

Seconds a memory (Remember) lasts without being written again. 0 keeps it for the whole session.

Confidence to enter Suspected state.

Confidence to enter Detected state.

Confidence to enter Tracked state.

The confidence at which the AI's awareness becomes Fully Aware.

Minimum seconds in Detected before transitioning to Tracked. Prevents instant lock.

The least time, in seconds, the AI stays focused on its attention target before it will consider switching: for example 0.5 for a camera, 2 for a guard, 5 for a zombie.

How much better a rival must score to take the AI's attention.

Keep track of where targets are: Last Known Position, Uncertainty Radius and Predicted Position.

Work out where a lost target is heading: Predicted Position moves ahead of the last known position, along its last movement.

How far ahead, in seconds, Predicted Position looks along the target's last movement.

The longest Prediction Horizon allowed, in seconds.

Rate at which uncertainty radius expands after losing target (cm/s).

Maximum uncertainty radius (cm).

If true, fast-moving targets produce larger uncertainty when lost.

How quickly the AI's measure of a target's speed follows changes, 0 to 1: lower is smoother, higher reacts faster.

Run vision's traces off the critical path. The exposure it reports is one vision update old (about 0.15 s). With more than one occlusion channel or any Surface Vision Transmission, APS traces normally anyway.

How often perception runs, in seconds. Lower is more reactive and costs more; doubling it halves the cost.

How often, in seconds, the AI tidies away beliefs that have expired.

How many beliefs this AI can hold at once. The slots are set up when it starts, so it is memory as well as time.

Let AI at tier 3 and 4 (far away and idle) update their component only every 0.1 and 0.25 s. Their perception timing doesn't change. For an idle AI those tiers start at about 150 m and 225 m from the nearest camera.

How much confidence counts in a target's sort score (which picks the top target and who keeps a slot in a crowd).

How much having sensed the target recently counts in its sort score.

How much being close counts in its sort score.

Beyond this distance (cm) being close adds nothing to the sort score.

Let distant AI (from Statistical Tier Threshold) perceive by chance instead of tracing: cheap, but it doesn't see walls, so a statistical AI can notice through them.

The level of detail tier from which an AI goes statistical. At 3 only distant, unengaged AI do (for an idle AI, beyond about 150 m).

Chance per second of noticing an ideally-placed target — close, lit, moving. Everything else scales this down.

Confidence granted when a statistical detection lands. Deliberately modest: a distant agent that half-noticed something should be suspicious, not certain.

The highest fear can climb to, 0 to 1. 0 turns fear off.

The highest aggression can climb to, 0 to 1. 0 turns aggression off.

The highest curiosity can climb to, 0 to 1. 0 turns curiosity off.

The highest alertness can climb to, 0 to 1. 0 turns alertness off.

The highest panic can climb to, 0 to 1. 0 turns panic off.

How fast fear rises, per second, while the AI has been hurt by one of its targets, believes in more than two actors (allies count), faces a Critical threat, or has a Fear Input above 0. Fear Input also multiplies this rate.

How fast fear falls, per second, once the AI believes in no one. While anyone remains, fear holds.

How fast aggression rises, per second, while the AI is hurt, tracks a target, or faces a high threat.

How fast aggression falls, per second, once no target is left and nothing has hurt it.

How fast curiosity rises, per second, while the AI only suspects something, or has heard something it can't see.

How fast curiosity falls, per second, once a target is Tracked or Lost, or nothing is left.

How fast alertness rises, per second, while a target is Detected or Tracked.

How fast alertness falls, per second, at all other times.

How fast panic rises, per second, while fear is above 0.8 with more than one target, or Fear Input is above 0.9.

How fast panic falls, per second, once fear is below 0.5 and Fear Input below 0.3.

An emotion must be above this to be the dominant emotion; when none is, the AI is Calm.

Above this, panic is the dominant emotion whatever the others are.

How much the AI's confidence in a target counts toward its threat score. The five weights are added up and the score is kept between 0 and 1.

How much the damage the target has done to this AI counts toward its threat score (100 damage counts in full).

How much it counts toward the threat score that several senses perceive the target at once: the number of active senses is divided by 8 (the most sense slots an AI can have), so a two-sense AI can add at most a quarter of this weight.

How much who the target is counts toward its threat score: Feared 1.5, High Value 1.2, Enemy 1, Unknown 0.5, Neutral 0.3, Friendly and Teammate 0.

A flat bonus to the threat score while the target is Detected or Tracked (sensed right now), nothing otherwise. The approach factor is 1 whenever the AI senses the target in APS 3.0, so this does not measure closing speed.

The threat score (0 to 1) at which a target becomes a Low threat.

The threat score at which a target becomes a Medium threat.

The threat score at which a target becomes a High threat.

The threat score at which a target becomes a Critical threat.

How fast the damage a target did fades from its threat score once the target is Lost, Remembered or Expired, per second. 0 never forgets.

How far, in cm, a share reaches: the squad's radio range.

The least threat level a target needs before this AI shares it with its squad.

Share a target with the squad by itself as soon as it is Detected: one sees you, the pack comes. Off: share it yourself (Share Target with Squad).

Share by itself when a target is Tracked instead. Works only with Auto Share on Detect off.

The combat roles this kind of AI may claim. Empty means any.

Record how each target plays, for the session: crouch, walk and sprint shares, how encounters ended, where it hid. Your logic reads it (Get Player Behavior Model); APS doesn't change behaviour by itself.

How many encounters before the player model's Has Enough Data turns true.

When no posture is reported, a target slower than this (cm/s) counts as crouching for the player model.

A target faster than this (cm/s) counts as sprinting for the player model.

Seconds the AI must keep sensing a target it doesn't know yet before confidence starts to count: a beat for the player to step back. Only until it first suspects that target. 0 is instant.

The confidence at which On AI Telegraph fires: a warning before the AI commits. Keep it between Suspect Threshold and Detect Threshold.

Hear worse while this AI is off the first player's screen (more than about 75° from where the camera points).

How much of its hearing an off-screen AI keeps, 0 to 1.

Whether this AI's Is Location in Never Search Zone answers. Off: it always answers false. APS enforces nothing by itself.

Replicate a compact per-target perception summary to clients so client-side detection meters, HUD and spectator tools work without custom plumbing. Perception itself stays server-authoritative.

Maximum targets included in the replicated summary. Keep small — this is bandwidth on every AI.

Seconds between replicated state refreshes.

Leave out of the replicated summary targets nobody is near (Replication Relevance Range) or that it barely suspects (Replication Min Confidence).

A target whose believed position is further than this (cm) from every player camera is left out. 0 turns the distance test off.

Targets below this confidence are left out of the replicated summary.

Send nothing when nothing changed by more than Replication Confidence Delta.

How much confidence must move before it is worth sending.