Generative Motion & Gestures ​

In AIRI, your 3D avatar does not simply repeat the same static idle loop. Through the Motion Engine, avatars can gesture while speaking, react emotionally to your words, perform procedural dances, and even synthesize brand-new full-body skeletal animations on demand using neural diffusion models.

1. How Motion Works in AIRI ​

AIRI uses a dual-engine motion architecture:

[ LLM Turn / User Command ]
              │
              ├──► Built-in Presets (11+ VRMA animations: waving, thinking, jumping, dancing)
              │
              ├──► Kinetic ACT Tokens (<|ACT:motion="excited_wave"|>)
              │
              └──► Generative Text-to-Motion (FlowMDM Neural Diffusion via WebGPU)

2. ACT Motion & Emotion Tokens ​

When chatting, your character's LLM can embed kinetic performance markers:

• <|ACT:motion="wave"|>: Triggers a friendly greeting wave.
• <|ACT:motion="think",duration="4"|>: Puts hand to chin thoughtfully during deep answers for 4 seconds.
• <|ACT:emotion="happy",intensity="0.8",duration="3"|>: Triggers a warm smile held for 3 seconds before smoothly returning to neutral.
• <|ACT:emotion="happy",motion="wave",duration="3"|>: Combines expression and motion cues in a single tag.
• <|ACT:emotion="neutral"|> or duration="0": Immediately releases any active expression hold.
• <|DELAY:2|>: Inserts a 2-second pause in speech.
• <|ACT:vfx="fire",duration="4"|>: Manifests elemental visual auras during emotional peaks.

These tokens are automatically intercepted by the parser and dispatched to the avatar renderer, blending smoothly with the current speech audio.

3. Generative Text-to-Motion (FlowMDM) ​

AIRI includes experimental support for FlowMDM (Local WebGPU Neural Motion Diffusion):

• How It Works: Given a text description (e.g., "Take two steps back and do a dramatic bow" or "Celebrate with a victory spin"), the local WebGPU worker synthesizes a 3D skeletal animation track and compiles it into standard VRMA bytecode directly inside your browser or desktop app.
• Zero Cloud Rendering: Runs on your local GPU with zero external cloud dependencies.

Enabling FlowMDM: ​

1. Open Settings → Modules → Text to Motion.
2. Select FlowMDM (Local WebGPU).
3. Click "Download Model Weights" (once downloaded, the weights are cached locally).
4. Test custom motion prompts in the Motion Playground.

4. Resting Postures & Random Cycles ​

In Settings → Character → Vessel → Customize Model:

• Default Idle: Set your character's base resting posture.
• Random Motion Cycles: Enable subtle random shifts (weight transfers, head tilts, hair adjustments) every 15–30 seconds to make the avatar feel alive when idle.