AI Development
Aether — Real-Time AI Decision Engine
A real-time AI inference layer that drives NPC behaviour, dynamic difficulty and narrative branching from a single model graph running inside Unreal Engine.
Challenge
A narrative game needed NPCs that could adapt to player strategy in real time, but behaviour was locked to scripted trees that broke the moment players deviated from the expected path.
Objectives
- Replace scripted decision trees with a live inference layer
- Let NPCs adapt to player behaviour within a session
- Keep inference latency inside a single frame budget
- Expose tuning to designers without touching engine code
Technical Approach
A behaviour model graph was built in Unity Sentis and mirrored in Unreal through a lightweight message bridge, so NPC perception, goals and memory feed a shared inference context. Designers author goals and weights as data assets; the runtime ranks and selects actions each tick. Telemetry logs every decision for offline tuning.
Development
- Model graph authoring and Sentis inference runtime
- Perception-to-context mapping for NPC state
- Data-driven goal and weight authoring
- Frame-budgeted inference scheduling
- Decision telemetry and replay tooling
- Designer-facing tuning dashboard
Key Features
- 01Real-time adaptive NPC behaviour
- 02Session-learned player strategy recognition
- 03Designer-authored goal weights
- 04Decision replay for offline analysis
Technology Stack
Results
- NPCs adapted to player strategy within a single session
- Inference held inside the 16ms frame budget
- Designers tuned behaviour without engine changes
Services Applied
Have a project in mind?
Tell us what you're building and we'll help you determine the right technical approach.