Multi-agent Traffic Simulation on SUMO urban network California (1.5 x 2 mile closed-loop)
5.84Average Speed (m/s)MotionLM
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| MotionLMType=Passive predictor (imitation learning agent)2026.05 | 5.84 | 0.168 | 0.0561 | 0.0127 | 0.0062 | 1.762 | |
| WayformerType=Passive predictor (imitation learning agent)2026.05 | 5.46 | 0.169 | 0.048 | 0.0108 | 0.0048 | 0.612 | |
| STEER under Gigaflow ArchitectureExecution=Rule-based2026.05 | 5.45 | 0.312 | 0.05 | 0.0076 | 0.0033 | — | |
| Proposed Hierarchical ArchitectureLow-level model=Command-Conditioned Wayformer, High-level module=Stackelberg-style MARL2026.05 | 5.45 | 0.169 | 0.0472 | 0.0047 | 0.0024 | 0.527 |