1. Purpose & Key Conclusions
Test Purpose
We benchmarked StateTree vs BehaviorTree runtime performance in UE 5.7 to evaluate whether StateTree is feasible as a BehaviorTree replacement in large-scale AI simulations (100-400 NPCs).
Key Conclusions
1. ~2x Framework Overhead: At unified TickInterval (0.5s), StateTree’s total execution duration is approximately 2x that of BehaviorTree, even with fewer execution counts.
2. High Fixed Overhead: Unreal Insights profiling shows StateTree component’s own Excl time is significantly higher than its internal task execution time, indicating substantial framework overhead independent of actual AI logic.
3. Default Tick Difference: Under default settings, StateTree ticks every frame while BehaviorTree does not, resulting in up to 18x higher CPU consumption at 400 NPCs.
2. Test Environment
- CPU: Intel Core Ultra 7 265K @ 3.90 GHz
- RAM: 32 GB
- GPU: NVIDIA GeForce RTX 5060 (8 GB)
- Engine: UE 5.7 (Packaged Build)
Methodology: Both BT and ST implement identical AI logic (movement, perception, behavioral responses). Data collected via Unreal Insights after 30s warm-up.
3. Supporting Data
3.1 Default Tick Strategy Difference
At 400 NPCs without animation/mesh, StateTree executes every frame while BehaviorTree has variable per-frame execution:
Figure 1: BehaviorTree Single Frame Execution (400 NPCs)
[Image Removed]
Figure 2: StateTree Single Frame Execution (400 NPCs)
[Image Removed] Performance Comparison (400 NPCs, No Animation, Default Settings):
- Execution Count: BT 271,073 vs ST 15,563,600 → 57x difference
- Total Duration: BT 2,721 ms vs ST 49,803 ms → 18x difference
- CPU Time %: BT 0.91% vs ST 16.63% → 18x difference
3.2 Unified TickInterval = 0.5s (400 NPCs)
With Animation (2min test):
- BT: 101,483 executions, 869 ms total
- ST: 98,854 executions, 1,577 ms total
- Ratio: ST is 1.8x slower
Without Animation (2min test):
- BT: 132,365 executions, 1,178 ms total
- ST: 106,872 executions, 2,432 ms total
- Ratio: ST is 2.1x slower (even with fewer executions)
3.3 Framework Overhead Analysis
Figure 3: BehaviorTree Task Execution Duration (No Animation, 2min)
[Image Removed]
Figure 4: StateTree Task Execution Duration (No Animation, 2min)
[Image Removed]
StateTree’s Excl time (framework overhead) is significantly higher than its internal task callees, while BehaviorTree shows lower component overhead relative to task execution.
4. Questions
Q1: Is the ~2x framework overhead expected for StateTree’s architecture? What contributes to this overhead?
Q2: Under what scenarios would StateTree outperform BehaviorTree?
Q3: What optimization strategies are recommended to reduce StateTree overhead for large-scale AI (400+ NPCs)?
Q4: In Insights, StateTree’s Excl time far exceeds internal task time. What internal processes cause this?
Q5: Adjusting ForegroundWorker via console crashes release builds. Is this known? Is there a safe way to configure worker threads at runtime?
[Attachment Removed]
