Smooth transitions, lip sync speech blend, body expression override mode, SendTextToSelectedAgent

- Replace FInterpConstantTo with FInterpTo (exponential ease-out) in all 4 anim components
- Apply SmoothStep to crossfade alphas for smooth animation transitions
- Add SpeechBlendAlpha to LipSync: fades out mouth curves when not speaking,
  letting FacialExpression emotion curves (including mouth) show through
- Remove LipSync AnimNode grace period zeroing to avoid overwriting emotion curves
- Revert BodyExpression to override mode (additive broken with full-pose anims)
- Default ExcludeBones = neck_01 to prevent Gaze/Posture conflicts
- Fix mid-crossfade animation pop in BodyExpression SwitchToNewAnim
- Add Neutral emotion fallback in Body and Facial expression components
- Add SendTextToSelectedAgent convenience method on InteractionComponent
- Add debug HUD display for BodyExpression component
- Update CoreRedirects for Posture → Gaze rename

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
j.foucher 2026-03-05 11:09:00 +01:00
parent 8bb4371a74
commit 2e96e3c766
13 changed files with 464 additions and 120 deletions

View File

@ -90,7 +90,7 @@ FontDPI=72
; ── Generic classes: ElevenLabs → PS_AI_ConvAgent ──
+ClassRedirects=(OldName="ElevenLabsLipSyncComponent", NewName="PS_AI_ConvAgent_LipSyncComponent")
+ClassRedirects=(OldName="ElevenLabsFacialExpressionComponent", NewName="PS_AI_ConvAgent_FacialExpressionComponent")
+ClassRedirects=(OldName="ElevenLabsPostureComponent", NewName="PS_AI_ConvAgent_PostureComponent")
+ClassRedirects=(OldName="ElevenLabsPostureComponent", NewName="PS_AI_ConvAgent_GazeComponent")
+ClassRedirects=(OldName="ElevenLabsMicrophoneCaptureComponent", NewName="PS_AI_ConvAgent_MicrophoneCaptureComponent")
+ClassRedirects=(OldName="ElevenLabsLipSyncPoseMap", NewName="PS_AI_ConvAgent_LipSyncPoseMap")
+ClassRedirects=(OldName="ElevenLabsEmotionPoseMap", NewName="PS_AI_ConvAgent_EmotionPoseMap")
@ -98,7 +98,7 @@ FontDPI=72
; ── Generic classes: PS_AI_Agent → PS_AI_ConvAgent (intermediate rename) ──
+ClassRedirects=(OldName="PS_AI_Agent_LipSyncComponent", NewName="PS_AI_ConvAgent_LipSyncComponent")
+ClassRedirects=(OldName="PS_AI_Agent_FacialExpressionComponent", NewName="PS_AI_ConvAgent_FacialExpressionComponent")
+ClassRedirects=(OldName="PS_AI_Agent_PostureComponent", NewName="PS_AI_ConvAgent_PostureComponent")
+ClassRedirects=(OldName="PS_AI_Agent_PostureComponent", NewName="PS_AI_ConvAgent_GazeComponent")
+ClassRedirects=(OldName="PS_AI_Agent_MicrophoneCaptureComponent", NewName="PS_AI_ConvAgent_MicrophoneCaptureComponent")
+ClassRedirects=(OldName="PS_AI_Agent_LipSyncPoseMap", NewName="PS_AI_ConvAgent_LipSyncPoseMap")
+ClassRedirects=(OldName="PS_AI_Agent_EmotionPoseMap", NewName="PS_AI_ConvAgent_EmotionPoseMap")
@ -114,18 +114,23 @@ FontDPI=72
; ── AnimNode structs ──
+StructRedirects=(OldName="AnimNode_ElevenLabsLipSync", NewName="AnimNode_PS_AI_ConvAgent_LipSync")
+StructRedirects=(OldName="AnimNode_ElevenLabsFacialExpression", NewName="AnimNode_PS_AI_ConvAgent_FacialExpression")
+StructRedirects=(OldName="AnimNode_ElevenLabsPosture", NewName="AnimNode_PS_AI_ConvAgent_Posture")
+StructRedirects=(OldName="AnimNode_ElevenLabsPosture", NewName="AnimNode_PS_AI_ConvAgent_Gaze")
+StructRedirects=(OldName="AnimNode_PS_AI_Agent_LipSync", NewName="AnimNode_PS_AI_ConvAgent_LipSync")
+StructRedirects=(OldName="AnimNode_PS_AI_Agent_FacialExpression", NewName="AnimNode_PS_AI_ConvAgent_FacialExpression")
+StructRedirects=(OldName="AnimNode_PS_AI_Agent_Posture", NewName="AnimNode_PS_AI_ConvAgent_Posture")
+StructRedirects=(OldName="AnimNode_PS_AI_Agent_Posture", NewName="AnimNode_PS_AI_ConvAgent_Gaze")
; ── AnimGraphNode classes ──
+ClassRedirects=(OldName="AnimGraphNode_ElevenLabsLipSync", NewName="AnimGraphNode_PS_AI_ConvAgent_LipSync")
+ClassRedirects=(OldName="AnimGraphNode_ElevenLabsFacialExpression", NewName="AnimGraphNode_PS_AI_ConvAgent_FacialExpression")
+ClassRedirects=(OldName="AnimGraphNode_ElevenLabsPosture", NewName="AnimGraphNode_PS_AI_ConvAgent_Posture")
+ClassRedirects=(OldName="AnimGraphNode_ElevenLabsPosture", NewName="AnimGraphNode_PS_AI_ConvAgent_Gaze")
+ClassRedirects=(OldName="AnimGraphNode_PS_AI_Agent_LipSync", NewName="AnimGraphNode_PS_AI_ConvAgent_LipSync")
+ClassRedirects=(OldName="AnimGraphNode_PS_AI_Agent_FacialExpression", NewName="AnimGraphNode_PS_AI_ConvAgent_FacialExpression")
+ClassRedirects=(OldName="AnimGraphNode_PS_AI_Agent_Posture", NewName="AnimGraphNode_PS_AI_ConvAgent_Posture")
+ClassRedirects=(OldName="AnimGraphNode_PS_AI_Agent_Posture", NewName="AnimGraphNode_PS_AI_ConvAgent_Gaze")
; ── Posture → Gaze rename ──
+ClassRedirects=(OldName="PS_AI_ConvAgent_PostureComponent", NewName="PS_AI_ConvAgent_GazeComponent")
+StructRedirects=(OldName="AnimNode_PS_AI_ConvAgent_Posture", NewName="AnimNode_PS_AI_ConvAgent_Gaze")
+ClassRedirects=(OldName="AnimGraphNode_PS_AI_ConvAgent_Posture", NewName="AnimGraphNode_PS_AI_ConvAgent_Gaze")
; ── Factory classes ──
+ClassRedirects=(OldName="ElevenLabsLipSyncPoseMapFactory", NewName="PS_AI_ConvAgent_LipSyncPoseMapFactory")

View File

@ -127,7 +127,10 @@ void FAnimNode_PS_AI_ConvAgent_BodyExpression::Evaluate_AnyThread(FPoseContext&
const bool bHavePrevPose = bCrossfading && PrevPose.GetNumBones() > 0;
// ── Per-bone blend ──────────────────────────────────────────────────────
// ── Per-bone blend (override mode) ──────────────────────────────────────
// Lerp each bone from the upstream pose toward the expression pose,
// weighted by the bone mask and activation alpha.
// ExcludeBones (default: neck_01) prevents conflicts with Gaze/Posture.
for (int32 CompactIdx = 0; CompactIdx < NumBones; ++CompactIdx)
{
if (BoneMask[CompactIdx] <= 0.0f)
@ -136,18 +139,19 @@ void FAnimNode_PS_AI_ConvAgent_BodyExpression::Evaluate_AnyThread(FPoseContext&
const FCompactPoseBoneIndex BoneIdx(CompactIdx);
const float W = BoneMask[CompactIdx] * FinalWeight;
// Compute the expression pose (handle crossfade between prev and active)
FTransform ExpressionPose;
if (bHavePrevPose)
{
// Crossfade between previous and active emotion poses
FTransform BlendedEmotion;
BlendedEmotion.Blend(PrevPose[BoneIdx], ActivePose[BoneIdx], CachedSnapshot.CrossfadeAlpha);
Output.Pose[BoneIdx].BlendWith(BlendedEmotion, W);
ExpressionPose.Blend(PrevPose[BoneIdx], ActivePose[BoneIdx], CachedSnapshot.CrossfadeAlpha);
}
else
{
// Direct blend with active emotion pose
Output.Pose[BoneIdx].BlendWith(ActivePose[BoneIdx], W);
ExpressionPose = ActivePose[BoneIdx];
}
// Override: replace (lerp) upstream toward expression pose
Output.Pose[BoneIdx].BlendWith(ExpressionPose, W);
}
}
@ -169,22 +173,22 @@ void FAnimNode_PS_AI_ConvAgent_BodyExpression::BuildBoneMask(const FBoneContaine
BoneMask.SetNumZeroed(NumBones);
bBoneMaskValid = false;
// Full body mode: all bones get weight 1.0
const FReferenceSkeleton& RefSkel = RequiredBones.GetReferenceSkeleton();
const TArray<FBoneIndexType>& BoneIndices = RequiredBones.GetBoneIndicesArray();
// ── Step 1: Build initial mask (include) ──────────────────────────────
if (!bUpperBodyOnly)
{
// Full body mode: all bones get weight 1.0
for (int32 i = 0; i < NumBones; ++i)
{
BoneMask[i] = 1.0f;
}
bBoneMaskValid = (NumBones > 0);
UE_LOG(LogPS_AI_ConvAgent_BodyExprAnimNode, Log,
TEXT("Bone mask built: FULL BODY (%d bones)."), NumBones);
return;
}
else
{
// Upper body mode: only BlendRootBone descendants
const FReferenceSkeleton& RefSkel = RequiredBones.GetReferenceSkeleton();
const int32 RootMeshIdx = RefSkel.FindBoneIndex(BlendRootBone);
if (RootMeshIdx == INDEX_NONE)
@ -195,9 +199,6 @@ void FAnimNode_PS_AI_ConvAgent_BodyExpression::BuildBoneMask(const FBoneContaine
return;
}
const TArray<FBoneIndexType>& BoneIndices = RequiredBones.GetBoneIndicesArray();
int32 MaskedCount = 0;
for (int32 CompactIdx = 0; CompactIdx < NumBones; ++CompactIdx)
{
const int32 MeshIdx = static_cast<int32>(BoneIndices[CompactIdx]);
@ -210,18 +211,73 @@ void FAnimNode_PS_AI_ConvAgent_BodyExpression::BuildBoneMask(const FBoneContaine
if (Current == RootMeshIdx)
{
BoneMask[CompactIdx] = 1.0f;
++MaskedCount;
break;
}
Current = RefSkel.GetParentIndex(Current);
}
}
}
// ── Step 2: Exclude bone subtrees ────────────────────────────────────
// For each exclude bone, zero out it and all its descendants.
// This prevents conflicts with Gaze/Posture on neck/head bones.
int32 ExcludedCount = 0;
if (ExcludeBones.Num() > 0)
{
// Collect mesh indices for all exclude roots
TArray<int32> ExcludeRootIndices;
for (const FName& BoneName : ExcludeBones)
{
const int32 ExclIdx = RefSkel.FindBoneIndex(BoneName);
if (ExclIdx != INDEX_NONE)
{
ExcludeRootIndices.Add(ExclIdx);
}
else
{
UE_LOG(LogPS_AI_ConvAgent_BodyExprAnimNode, Warning,
TEXT("ExcludeBone '%s' not found in skeleton — ignored."),
*BoneName.ToString());
}
}
// Zero out any bone whose parent chain includes an exclude root
for (int32 CompactIdx = 0; CompactIdx < NumBones; ++CompactIdx)
{
if (BoneMask[CompactIdx] <= 0.0f)
continue; // Already excluded or not included
const int32 MeshIdx = static_cast<int32>(BoneIndices[CompactIdx]);
int32 Current = MeshIdx;
while (Current != INDEX_NONE)
{
if (ExcludeRootIndices.Contains(Current))
{
BoneMask[CompactIdx] = 0.0f;
++ExcludedCount;
break;
}
Current = RefSkel.GetParentIndex(Current);
}
}
}
// ── Count final active bones ─────────────────────────────────────────
int32 MaskedCount = 0;
for (int32 i = 0; i < NumBones; ++i)
{
if (BoneMask[i] > 0.0f)
++MaskedCount;
}
bBoneMaskValid = (MaskedCount > 0);
UE_LOG(LogPS_AI_ConvAgent_BodyExprAnimNode, Log,
TEXT("Bone mask built: %d/%d bones from root '%s'."),
MaskedCount, NumBones, *BlendRootBone.ToString());
TEXT("Bone mask built: %d/%d active bones (root='%s', excluded=%d from %d subtrees)."),
MaskedCount, NumBones, *BlendRootBone.ToString(),
ExcludedCount, ExcludeBones.Num());
}
// ─────────────────────────────────────────────────────────────────────────────

View File

@ -89,17 +89,15 @@ void FAnimNode_PS_AI_ConvAgent_LipSync::Evaluate_AnyThread(FPoseContext& Output)
// ── Inject lip sync curves into the pose output ──────────────────────
//
// IMPORTANT: Always write ALL curves that lip sync has ever touched,
// including at 0.0. If we skip near-zero curves, upstream animation
// values (idle expressions, breathing, etc.) leak through and cause
// visible pops when lip sync curves cross the threshold.
//
// Strategy:
// - While lip sync is producing curves: write them all (including 0s)
// While lip sync is producing curves: write them all (including 0s)
// and track every curve name in KnownCurveNames.
// - After lip sync goes silent: keep writing 0s for a grace period
// (30 frames ≈ 0.5s) so the upstream can blend back in smoothly
// via the component's activation alpha, then release.
//
// When lip sync goes silent (CachedCurves empty): release immediately.
// The component's SpeechBlendAlpha already handles smooth fade-out,
// so by the time CachedCurves becomes empty, values have already
// decayed to near-zero. Releasing immediately allows the upstream
// FacialExpressionComponent's emotion curves (including mouth) to
// flow through without being overwritten by zeros.
if (CachedCurves.Num() > 0)
{
@ -113,7 +111,9 @@ void FAnimNode_PS_AI_ConvAgent_LipSync::Evaluate_AnyThread(FPoseContext& Output)
}
// Zero any known curves NOT in the current frame
// (e.g. a blendshape that was active last frame but decayed away)
// (e.g. a blendshape that was active last frame but decayed away).
// This prevents upstream animation values from leaking through
// during active speech when a curve temporarily goes to zero.
for (const FName& Name : KnownCurveNames)
{
if (!CachedCurves.Contains(Name))
@ -124,24 +124,12 @@ void FAnimNode_PS_AI_ConvAgent_LipSync::Evaluate_AnyThread(FPoseContext& Output)
}
else if (KnownCurveNames.Num() > 0)
{
// Lip sync went silent — keep zeroing known curves for a grace
// period so upstream values don't pop in abruptly.
++FramesSinceLastActive;
if (FramesSinceLastActive < 30)
{
for (const FName& Name : KnownCurveNames)
{
Output.Curve.Set(Name, 0.0f);
}
}
else
{
// Grace period over — release curves, let upstream through
// Lip sync went silent — release immediately.
// The smooth fade-out was handled at the component level
// (SpeechBlendAlpha), so upstream emotion curves can take over.
KnownCurveNames.Reset();
FramesSinceLastActive = 0;
}
}
}
void FAnimNode_PS_AI_ConvAgent_LipSync::GatherDebugData(FNodeDebugData& DebugData)

View File

@ -4,6 +4,7 @@
#include "PS_AI_ConvAgent_ElevenLabsComponent.h"
#include "PS_AI_ConvAgent_BodyPoseMap.h"
#include "Animation/AnimSequence.h"
#include "Engine/Engine.h"
DEFINE_LOG_CATEGORY_STATIC(LogPS_AI_ConvAgent_BodyExpr, Log, All);
@ -119,6 +120,12 @@ const TArray<TObjectPtr<UAnimSequence>>* UPS_AI_ConvAgent_BodyExpressionComponen
if (!BodyPoseMap) return nullptr;
const FPS_AI_ConvAgent_BodyAnimList* AnimList = BodyPoseMap->BodyPoses.Find(Emotion);
// Fallback to Neutral if the requested emotion has no entry in the data asset
if (!AnimList && Emotion != EPS_AI_ConvAgent_Emotion::Neutral)
{
AnimList = BodyPoseMap->BodyPoses.Find(EPS_AI_ConvAgent_Emotion::Neutral);
}
if (!AnimList) return nullptr;
if (!bSpeaking)
@ -178,15 +185,38 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::SwitchToNewAnim(UAnimSequence* Ne
if (!bForce && NewAnim == ActiveAnim) return;
if (!NewAnim) return;
// Current active becomes previous for crossfade
if (CrossfadeAlpha < 1.0f && PrevAnim)
{
// Mid-crossfade: a crossfade is already in progress.
// DON'T reset CrossfadeAlpha — just swap the target animation.
// This preserves PrevAnim's contribution and avoids a visual pop.
//
// Before: Blend(PrevAnim, OldActive, alpha) e.g. 70% Prev + 30% Active
// After: Blend(PrevAnim, NewAnim, alpha) e.g. 70% Prev + 30% New
//
// The crossfade continues naturally — New fades in, Prev fades out.
// Pop is only 30% * (New@0 - OldActive@t) instead of 70% * (Prev - Active).
ActiveAnim = NewAnim;
ActivePlaybackTime = 0.0f;
// CrossfadeAlpha stays where it is — continuity
if (bDebug && DebugVerbosity >= 1)
{
UE_LOG(LogPS_AI_ConvAgent_BodyExpr, Log,
TEXT("Body anim switch (MID-CROSSFADE α=%.2f): target → %s (%s, %s)"),
CrossfadeAlpha, *NewAnim->GetName(),
bIsSpeaking ? TEXT("speaking") : TEXT("idle"),
*UEnum::GetValueAsString(ActiveEmotion));
}
}
else
{
// No crossfade in progress — normal switch with full crossfade
PrevAnim = ActiveAnim;
PrevPlaybackTime = ActivePlaybackTime;
// New anim starts from the beginning
ActiveAnim = NewAnim;
ActivePlaybackTime = 0.0f;
// Begin crossfade
CrossfadeAlpha = 0.0f;
if (bDebug && DebugVerbosity >= 1)
@ -198,6 +228,7 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::SwitchToNewAnim(UAnimSequence* Ne
bIsSpeaking ? TEXT("speaking") : TEXT("idle"),
*UEnum::GetValueAsString(ActiveEmotion));
}
}
}
void UPS_AI_ConvAgent_BodyExpressionComponent::PickAndSwitchAnim()
@ -221,6 +252,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnConversationConnected(
{
bActive = true;
bIsSpeaking = false;
LastEventName = TEXT("Connected");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
// Start with an idle anim
PickAndSwitchAnim();
@ -237,6 +270,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnConversationDisconnected(
{
bActive = false;
bIsSpeaking = false;
LastEventName = TEXT("Disconnected");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
if (bDebug)
{
@ -252,6 +287,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnConversationDisconnected(
void UPS_AI_ConvAgent_BodyExpressionComponent::OnSpeakingStarted()
{
bIsSpeaking = true;
LastEventName = TEXT("SpeakStart");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
// Crossfade from idle anim to a speaking anim
PickAndSwitchAnim();
@ -266,6 +303,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnSpeakingStarted()
void UPS_AI_ConvAgent_BodyExpressionComponent::OnSpeakingStopped()
{
bIsSpeaking = false;
LastEventName = TEXT("SpeakStop");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
// Crossfade from speaking anim to an idle anim
PickAndSwitchAnim();
@ -280,6 +319,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnSpeakingStopped()
void UPS_AI_ConvAgent_BodyExpressionComponent::OnInterrupted()
{
bIsSpeaking = false;
LastEventName = TEXT("Interrupted");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
// Crossfade to idle anim
PickAndSwitchAnim();
@ -303,6 +344,9 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::OnEmotionChanged(
ActiveEmotion = Emotion;
ActiveEmotionIntensity = Intensity;
LastEventName = FString::Printf(TEXT("Emotion:%s"),
*UEnum::GetDisplayValueAsText(Emotion).ToString());
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
// Pick a new anim from the appropriate list for the new emotion
PickAndSwitchAnim();
@ -358,9 +402,12 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::TickComponent(
const float TargetAlpha = bActive ? 1.0f : 0.0f;
if (!FMath::IsNearlyEqual(CurrentActiveAlpha, TargetAlpha, 0.001f))
{
const float BlendSpeed = 1.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpConstantTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, BlendSpeed);
// Exponential ease-out: fast start, gradual approach to target.
// Factor of 3 compensates for FInterpTo's exponential decay
// reaching ~95% in ActivationBlendDuration seconds.
const float InterpSpeed = 3.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, InterpSpeed);
}
else
{
@ -416,6 +463,8 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::TickComponent(
if (NewAnim)
{
SwitchToNewAnim(NewAnim, true);
LastEventName = TEXT("AutoCycle");
LastEventWorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
if (bDebug && DebugVerbosity >= 2)
{
@ -449,8 +498,80 @@ void UPS_AI_ConvAgent_BodyExpressionComponent::TickComponent(
CurrentSnapshot.PrevAnim = PrevAnim;
CurrentSnapshot.ActiveTime = ActivePlaybackTime;
CurrentSnapshot.PrevTime = PrevPlaybackTime;
CurrentSnapshot.CrossfadeAlpha = CrossfadeAlpha;
// Apply SmoothStep for ease-in-out crossfade (raw alpha is linear)
CurrentSnapshot.CrossfadeAlpha = FMath::SmoothStep(0.0f, 1.0f, CrossfadeAlpha);
CurrentSnapshot.ActivationAlpha = CurrentActiveAlpha;
CurrentSnapshot.BlendWeight = BlendWeight;
}
// ── On-screen debug HUD ───────────────────────────────────────────────
if (bDebug && DebugVerbosity >= 1)
{
DrawDebugHUD();
}
}
// ─────────────────────────────────────────────────────────────────────────────
// On-screen debug display
// ─────────────────────────────────────────────────────────────────────────────
void UPS_AI_ConvAgent_BodyExpressionComponent::DrawDebugHUD() const
{
if (!GEngine) return;
const float WorldTime = GetWorld() ? GetWorld()->GetTimeSeconds() : 0.0f;
const float EventAge = WorldTime - LastEventWorldTime;
// Active anim name
FString ActiveName = ActiveAnim ? ActiveAnim->GetName() : TEXT("(none)");
FString PrevName = PrevAnim ? PrevAnim->GetName() : TEXT("---");
// Smoothed crossfade for display
const float SmoothedCrossfade = FMath::SmoothStep(0.0f, 1.0f, CrossfadeAlpha);
// State label
FString StateStr;
if (!bActive)
StateStr = TEXT("INACTIVE");
else if (bIsSpeaking)
StateStr = TEXT("SPEAKING");
else
StateStr = TEXT("IDLE");
// Event label with age
FString EventStr = LastEventName.IsEmpty()
? TEXT("---")
: FString::Printf(TEXT("%s (%.1fs ago)"), *LastEventName, EventAge);
// Use key offset to avoid colliding with other debug messages
// Keys 2000-2010 reserved for BodyExpression
const int32 BaseKey = 2000;
const float DisplayTime = 0.0f; // Refresh every frame
const FColor MainColor = FColor::Cyan;
const FColor WarnColor = FColor::Yellow;
GEngine->AddOnScreenDebugMessage(BaseKey, DisplayTime, MainColor,
FString::Printf(TEXT("=== BODY EXPR: %s ==="), *StateStr));
GEngine->AddOnScreenDebugMessage(BaseKey + 1, DisplayTime, MainColor,
FString::Printf(TEXT(" ActivationAlpha: %.3f (target: %s)"),
CurrentActiveAlpha, bActive ? TEXT("1") : TEXT("0")));
GEngine->AddOnScreenDebugMessage(BaseKey + 2, DisplayTime, MainColor,
FString::Printf(TEXT(" Active: %s t=%.2f"), *ActiveName, ActivePlaybackTime));
GEngine->AddOnScreenDebugMessage(BaseKey + 3, DisplayTime,
CrossfadeAlpha < 1.0f ? WarnColor : MainColor,
FString::Printf(TEXT(" Crossfade: %.3f (smooth: %.3f) Prev: %s"),
CrossfadeAlpha, SmoothedCrossfade, *PrevName));
GEngine->AddOnScreenDebugMessage(BaseKey + 4, DisplayTime, MainColor,
FString::Printf(TEXT(" Emotion: %s (%s) Weight: %.2f"),
*UEnum::GetDisplayValueAsText(ActiveEmotion).ToString(),
*UEnum::GetDisplayValueAsText(ActiveEmotionIntensity).ToString(),
BlendWeight));
GEngine->AddOnScreenDebugMessage(BaseKey + 5, DisplayTime,
EventAge < 1.0f ? FColor::Green : MainColor,
FString::Printf(TEXT(" LastEvent: %s"), *EventStr));
}

View File

@ -136,6 +136,12 @@ UAnimSequence* UPS_AI_ConvAgent_FacialExpressionComponent::FindAnimForEmotion(
if (!EmotionPoseMap) return nullptr;
const FPS_AI_ConvAgent_EmotionPoseSet* PoseSet = EmotionPoseMap->EmotionPoses.Find(Emotion);
// Fallback to Neutral if the requested emotion has no entry in the data asset
if (!PoseSet && Emotion != EPS_AI_ConvAgent_Emotion::Neutral)
{
PoseSet = EmotionPoseMap->EmotionPoses.Find(EPS_AI_ConvAgent_Emotion::Neutral);
}
if (!PoseSet) return nullptr;
// Direct match
@ -282,9 +288,10 @@ void UPS_AI_ConvAgent_FacialExpressionComponent::TickComponent(
const float TargetAlpha = bActive ? 1.0f : 0.0f;
if (!FMath::IsNearlyEqual(CurrentActiveAlpha, TargetAlpha, 0.001f))
{
const float BlendSpeed = 1.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpConstantTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, BlendSpeed);
// Exponential ease-out: fast start, gradual approach to target.
const float InterpSpeed = 3.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, InterpSpeed);
}
else
{
@ -356,13 +363,16 @@ void UPS_AI_ConvAgent_FacialExpressionComponent::TickComponent(
for (const auto& P : ActiveCurves) AllCurves.Add(P.Key);
for (const auto& P : PrevCurves) AllCurves.Add(P.Key);
// Apply SmoothStep for ease-in-out crossfade (raw alpha is linear)
const float SmoothedCrossfade = FMath::SmoothStep(0.0f, 1.0f, CrossfadeAlpha);
for (const FName& CurveName : AllCurves)
{
const float PrevVal = PrevCurves.Contains(CurveName)
? PrevCurves[CurveName] : 0.0f;
const float ActiveVal = ActiveCurves.Contains(CurveName)
? ActiveCurves[CurveName] : 0.0f;
const float Blended = FMath::Lerp(PrevVal, ActiveVal, CrossfadeAlpha);
const float Blended = FMath::Lerp(PrevVal, ActiveVal, SmoothedCrossfade);
// Always include the curve even at 0 — the AnimNode needs
// to see it to block upstream values from popping through.

View File

@ -3,7 +3,9 @@
#include "PS_AI_ConvAgent_GazeComponent.h"
#include "PS_AI_ConvAgent_ElevenLabsComponent.h"
#include "Components/SkeletalMeshComponent.h"
#include "Camera/CameraComponent.h"
#include "GameFramework/Actor.h"
#include "GameFramework/Pawn.h"
#include "Math/UnrealMathUtility.h"
#include "DrawDebugHelpers.h"
@ -38,7 +40,8 @@ static const FName TargetHeadBone(TEXT("head"));
* bAutoTargetEyes = true:
* 1. Try eye bones (FACIAL_L_Eye / FACIAL_R_Eye midpoint) on the target's Face mesh.
* 2. Fallback to "head" bone on any skeletal mesh.
* 3. Fallback to TargetActor origin + (0, 0, FallbackEyeHeight).
* 3. Fallback to CameraComponent location (first-person pawn).
* 4. Fallback to TargetActor origin + (0, 0, FallbackEyeHeight) for non-camera actors.
*
* bAutoTargetEyes = false:
* Always returns TargetActor origin + TargetOffset.
@ -76,7 +79,15 @@ static FVector ResolveTargetPosition(const AActor* Target, bool bAutoEyes,
}
}
// No skeleton — use FallbackEyeHeight
// Fallback: CameraComponent — the canonical eye position for
// first-person pawns and any actor with an active camera.
if (const UCameraComponent* Cam =
const_cast<AActor*>(Target)->FindComponentByClass<UCameraComponent>())
{
return Cam->GetComponentLocation();
}
// Final fallback: actor origin + height offset
return Target->GetActorLocation() + FVector(0.0f, 0.0f, FallbackHeight);
}
@ -335,9 +346,10 @@ void UPS_AI_ConvAgent_GazeComponent::TickComponent(
const float TargetAlpha = bActive ? 1.0f : 0.0f;
if (!FMath::IsNearlyEqual(CurrentActiveAlpha, TargetAlpha, 0.001f))
{
const float BlendSpeed = 1.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpConstantTo(
CurrentActiveAlpha, TargetAlpha, SafeDeltaTime, BlendSpeed);
// Exponential ease-out: fast start, gradual approach to target.
const float InterpSpeed = 3.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpTo(
CurrentActiveAlpha, TargetAlpha, SafeDeltaTime, InterpSpeed);
}
else
{

View File

@ -476,6 +476,37 @@ void UPS_AI_ConvAgent_InteractionComponent::StartConversationWithSelectedAgent()
}
}
void UPS_AI_ConvAgent_InteractionComponent::SendTextToSelectedAgent(const FString& Text)
{
UPS_AI_ConvAgent_ElevenLabsComponent* Agent = SelectedAgent.Get();
if (!Agent)
{
if (bDebug)
{
UE_LOG(LogPS_AI_ConvAgent_Select, Warning,
TEXT("SendTextToSelectedAgent: no agent selected."));
}
return;
}
// Route through relay on clients (can't call Server RPCs on NPC actors).
if (GetOwnerRole() == ROLE_Authority || (GetWorld() && GetWorld()->GetNetMode() == NM_Standalone))
{
Agent->SendTextMessage(Text);
}
else
{
ServerRelaySendText(Agent->GetOwner(), Text);
}
if (bDebug)
{
UE_LOG(LogPS_AI_ConvAgent_Select, Log,
TEXT("SendTextToSelectedAgent: \"%s\" → %s"),
*Text, Agent->GetOwner() ? *Agent->GetOwner()->GetName() : TEXT("(null)"));
}
}
// ─────────────────────────────────────────────────────────────────────────────
// Gaze helpers
// ─────────────────────────────────────────────────────────────────────────────

View File

@ -702,9 +702,10 @@ void UPS_AI_ConvAgent_LipSyncComponent::TickComponent(float DeltaTime, ELevelTic
const float TargetAlpha = bActive ? 1.0f : 0.0f;
if (!FMath::IsNearlyEqual(CurrentActiveAlpha, TargetAlpha, 0.001f))
{
const float BlendSpeed = 1.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpConstantTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, BlendSpeed);
// Exponential ease-out: fast start, gradual approach to target.
const float InterpSpeed = 3.0f / FMath::Max(ActivationBlendDuration, 0.01f);
CurrentActiveAlpha = FMath::FInterpTo(
CurrentActiveAlpha, TargetAlpha, DeltaTime, InterpSpeed);
}
else
{
@ -712,6 +713,24 @@ void UPS_AI_ConvAgent_LipSyncComponent::TickComponent(float DeltaTime, ELevelTic
}
}
// ── Smooth speech blend ────────────────────────────────────────────────
// Fades lip sync curves in/out when speech starts/stops.
// When not speaking, lip sync releases all mouth curves so emotion
// facial expression (from FacialExpressionComponent) controls the mouth.
{
const float TargetSpeech = bIsSpeaking ? 1.0f : 0.0f;
if (!FMath::IsNearlyEqual(SpeechBlendAlpha, TargetSpeech, 0.001f))
{
const float SpeechSpeed = 3.0f / FMath::Max(SpeechBlendDuration, 0.01f);
SpeechBlendAlpha = FMath::FInterpTo(
SpeechBlendAlpha, TargetSpeech, DeltaTime, SpeechSpeed);
}
else
{
SpeechBlendAlpha = TargetSpeech;
}
}
// ── Lazy binding: in packaged builds, BeginPlay may run before the ────────
// ElevenLabsComponent is fully initialized. Retry discovery until bound.
if (!AgentComponent.IsValid())
@ -858,6 +877,12 @@ void UPS_AI_ConvAgent_LipSyncComponent::TickComponent(float DeltaTime, ELevelTic
{
bVisemeTimelineActive = false;
// Fallback: mark as not speaking when timeline ends
if (bIsSpeaking && AudioEnvelopeValue < 0.01f)
{
bIsSpeaking = false;
}
if (AccumulatedText.Len() > 0 && bTextVisemesApplied && bFullTextReceived)
{
AccumulatedText.Reset();
@ -964,6 +989,14 @@ void UPS_AI_ConvAgent_LipSyncComponent::TickComponent(float DeltaTime, ELevelTic
TargetVisemes.FindOrAdd(FName("sil")) = 1.0f;
PlaybackTimer = 0.0f;
// Fallback: if the queue has been dry and envelope is near zero,
// mark as not speaking. OnAgentStopped should have already done this,
// but this handles edge cases where the event doesn't fire.
if (bIsSpeaking && AudioEnvelopeValue < 0.01f)
{
bIsSpeaking = false;
}
if (AccumulatedText.Len() > 0 && bTextVisemesApplied && bFullTextReceived)
{
AccumulatedText.Reset();
@ -1052,6 +1085,24 @@ void UPS_AI_ConvAgent_LipSyncComponent::TickComponent(float DeltaTime, ELevelTic
}
}
// ── Apply speech blend alpha ─────────────────────────────────────────
// When the agent is NOT speaking, fade out all lip sync curves so the
// FacialExpressionComponent's emotion curves (including mouth expressions)
// can flow through unobstructed. Without this, lip sync zeros on mouth
// curves (via the AnimNode → mh_arkit_mapping_pose) would overwrite the
// emotion mouth curves even during silence.
if (SpeechBlendAlpha < 0.001f)
{
CurrentBlendshapes.Reset();
}
else if (SpeechBlendAlpha < 0.999f)
{
for (auto& Pair : CurrentBlendshapes)
{
Pair.Value *= SpeechBlendAlpha;
}
}
// Auto-apply morph targets if a target mesh is set
if (TargetMesh)
{
@ -1090,6 +1141,10 @@ void UPS_AI_ConvAgent_LipSyncComponent::OnAgentStopped()
VisemeTimeline.Reset();
VisemeTimelineCursor = 0.0f;
TotalActiveFramesSeen = 0;
// Speech ended — SpeechBlendAlpha will smoothly fade out lip sync curves,
// allowing FacialExpression emotion curves to take over the mouth.
bIsSpeaking = false;
}
void UPS_AI_ConvAgent_LipSyncComponent::ResetToNeutral()
@ -1113,6 +1168,11 @@ void UPS_AI_ConvAgent_LipSyncComponent::ResetToNeutral()
VisemeTimelineCursor = 0.0f;
TotalActiveFramesSeen = 0;
// Speech interrupted — snap speech blend to 0 immediately
// so emotion curves take over the mouth right away.
bIsSpeaking = false;
SpeechBlendAlpha = 0.0f;
// Snap all visemes to silence immediately (no smoothing delay)
for (const FName& Name : VisemeNames)
{
@ -1136,6 +1196,10 @@ void UPS_AI_ConvAgent_LipSyncComponent::OnAudioChunkReceived(const TArray<uint8>
{
if (!SpectrumAnalyzer) return;
// Mark as speaking — audio is arriving from the agent.
// SpeechBlendAlpha will smoothly fade in lip sync curves.
bIsSpeaking = true;
// Convert int16 PCM to float32 [-1, 1]
const int16* Samples = reinterpret_cast<const int16*>(PCMData.GetData());
const int32 NumSamples = PCMData.Num() / sizeof(int16);

View File

@ -16,10 +16,13 @@ class UAnimSequence;
* AnimSequences from the PS_AI_ConvAgent_BodyExpressionComponent and blends
* them per-bone onto the upstream pose (idle, locomotion).
*
* Two modes:
* - bUpperBodyOnly = true (default): only bones at and below BlendRootBone
* are blended; lower body passes through from the upstream pose.
* - bUpperBodyOnly = false: the emotion pose is applied to the entire skeleton.
* The node uses Override mode: it replaces (lerps) the upstream pose toward
* the expression pose. Use ExcludeBones (default: neck_01) to prevent
* conflicts with Gaze/Posture on neck/head bones.
*
* Region modes:
* - bUpperBodyOnly = true (default): only bones at and below BlendRootBone.
* - bUpperBodyOnly = false: full skeleton.
*
* Graph layout:
* [Upstream body anims (idle, locomotion)] [PS AI ConvAgent Body Expression] [Output]
@ -50,6 +53,13 @@ struct PS_AI_CONVAGENT_API FAnimNode_PS_AI_ConvAgent_BodyExpression : public FAn
ToolTip = "Root bone for upper body blend.\nAll descendants of this bone are blended with the emotion pose.\nDefault: spine_02 (arms, spine, neck, head)."))
FName BlendRootBone = FName(TEXT("spine_02"));
/** Bones to EXCLUDE from body expression blending (each bone and its entire subtree).
* Prevents conflicts with Gaze/Posture on neck/head bones.
* Default: neck_01 (excludes neck + head, leaving them to Gaze/Posture). */
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "Settings",
meta = (ToolTip = "Exclude these bones and their subtrees from blending.\nDefault: neck_01 prevents conflicts with Gaze/Posture on neck/head."))
TArray<FName> ExcludeBones = { FName(TEXT("neck_01")) };
// ── FAnimNode_Base interface ──────────────────────────────────────────────
virtual void Initialize_AnyThread(const FAnimationInitializeContext& Context) override;

View File

@ -232,4 +232,15 @@ private:
/** Cached reference to the agent component on the same Actor. */
TWeakObjectPtr<UPS_AI_ConvAgent_ElevenLabsComponent> AgentComponent;
// ── Debug event tracking ────────────────────────────────────────────────
/** Last event name (for on-screen debug display). */
FString LastEventName;
/** World time when the last event fired. */
float LastEventWorldTime = 0.0f;
/** Draw on-screen debug info (called from TickComponent when bDebug). */
void DrawDebugHUD() const;
};

View File

@ -95,17 +95,18 @@ public:
/** Automatically aim at the target's eye bones (MetaHuman FACIAL_L_Eye / FACIAL_R_Eye).
* When enabled, TargetOffset is ignored and the agent looks at the midpoint
* between the target pawn's eye bones.
* Fallback chain: eye bones head bone ActorOrigin + (0,0,FallbackEyeHeight). */
* Fallback chain: eye bones head bone PawnViewLocation FallbackEyeHeight. */
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "PS AI ConvAgent|Gaze",
meta = (ToolTip = "Auto-target the pawn's eye bones for eye contact.\nFallback: eye bones > head bone > FallbackEyeHeight."))
meta = (ToolTip = "Auto-target the pawn's eye bones for eye contact.\nFallback: eye bones > head bone > PawnViewLocation > FallbackEyeHeight."))
bool bAutoTargetEyes = true;
/** Height offset (cm) from the target actor's origin when no eye/head bones are found.
* Used as fallback when bAutoTargetEyes is true but the target has no skeleton
* (e.g. first-person pawn, simple actor). 160 eye height for a standing human. */
/** Height offset (cm) from the target actor's origin when no eye/head bones are found
* AND the target is not a Pawn. For Pawns without skeleton, GetPawnViewLocation()
* is used instead (accounts for BaseEyeHeight automatically).
* Only applies to non-Pawn actors (props, triggers, etc.). */
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "PS AI ConvAgent|Gaze",
meta = (EditCondition = "bAutoTargetEyes", ClampMin = "0",
ToolTip = "Height offset (cm) when no eye/head bones exist on the target.\n160 = standing human eye level.\nOnly used as last-resort fallback."))
ToolTip = "Height offset (cm) for non-Pawn targets without skeleton.\nPawns use BaseEyeHeight automatically.\nOnly used as last-resort fallback for non-Pawn actors."))
float FallbackEyeHeight = 160.0f;
/** Offset from the target actor's origin to aim at.

View File

@ -184,6 +184,13 @@ public:
UFUNCTION(BlueprintCallable, Category = "PS AI ConvAgent|Interaction")
void StartConversationWithSelectedAgent();
/** Send a text message to the currently selected agent.
* The agent responds with audio and text, just as if it heard the player speak.
* Handles network relay automatically (works on both server and client).
* Does nothing if no agent is selected or not connected. */
UFUNCTION(BlueprintCallable, Category = "PS AI ConvAgent|Interaction")
void SendTextToSelectedAgent(const FString& Text);
/** Clear the current selection. Automatic selection resumes next tick. */
UFUNCTION(BlueprintCallable, Category = "PS AI ConvAgent|Interaction")
void ClearSelection();

View File

@ -89,6 +89,17 @@ public:
ToolTip = "Smoothing speed for viseme transitions.\n35 = smooth and soft, 50 = balanced, 65 = sharp and responsive."))
float SmoothingSpeed = 50.0f;
// ── Speech Blend ────────────────────────────────────────────────────────
/** How long (seconds) to blend lip sync curves in/out when speech starts/stops.
* When the agent is NOT speaking, lip sync releases all mouth curves so the
* FacialExpressionComponent's emotion curves (including mouth) can show through.
* When the agent starts speaking, lip sync smoothly takes over mouth curves. */
UPROPERTY(EditAnywhere, BlueprintReadWrite, Category = "PS AI ConvAgent|LipSync",
meta = (ClampMin = "0.05", ClampMax = "1.0",
ToolTip = "Blend duration when speech starts/stops.\nDuring silence, emotion facial curves control the mouth.\nDuring speech, lip sync takes over."))
float SpeechBlendDuration = 0.15f;
// ── Emotion Expression Blend ─────────────────────────────────────────────
/** How much facial emotion (from PS_AI_ConvAgent_FacialExpressionComponent) bleeds through
@ -163,6 +174,15 @@ public:
UFUNCTION(BlueprintCallable, Category = "PS AI ConvAgent|LipSync")
TMap<FName, float> GetCurrentBlendshapes() const { return CurrentBlendshapes; }
/** True when the agent is currently producing speech audio.
* When false, lip sync releases mouth curves to let emotion curves through. */
UFUNCTION(BlueprintPure, Category = "PS AI ConvAgent|LipSync")
bool IsSpeaking() const { return bIsSpeaking; }
/** Current speech blend alpha (0 = silent/emotion mouth, 1 = lip sync mouth).
* Smooth transition between silence and speech states. */
float GetSpeechBlendAlpha() const { return SpeechBlendAlpha; }
// ── UActorComponent overrides ─────────────────────────────────────────────
virtual void BeginPlay() override;
virtual void EndPlay(const EEndPlayReason::Type EndPlayReason) override;
@ -258,6 +278,14 @@ private:
// Current blend alpha (0 = fully inactive/passthrough, 1 = fully active).
float CurrentActiveAlpha = 1.0f;
// True when the agent is currently producing speech audio.
// Set true in OnAudioChunkReceived, false in OnAgentStopped/OnAgentInterrupted.
bool bIsSpeaking = false;
// Smooth blend alpha for speech state: 0 = not speaking (emotion controls mouth),
// 1 = speaking (lip sync controls mouth). Interpolated each tick.
float SpeechBlendAlpha = 0.0f;
// MetaHuman mode: Face mesh has no morph targets, use animation curves instead.
// Set automatically in BeginPlay when TargetMesh has 0 morph targets.
bool bUseCurveMode = false;