Micromanagement on AVACraft 3m scenario
81Win RateGPT-4o
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-4oInput Mode=VLM-Optimized, Steps=02025.03 | 81 | |
| GPT-4-TurboInput Mode=VLM-Optimized, Steps=02025.03 | 79 | |
| GPT-4o-miniInput Mode=VLM-Optimized, Steps=02025.03 | 76 | |
| Qwen-VL-PlusInput Mode=VLM-Optimized, Steps=02025.03 | 75 | |
| Qwen3-VL-30BInput Mode=VLM-Optimized, Steps=02025.03 | 50 | |
| Qwen3-VL-8BInput Mode=VLM-Optimized, Steps=02025.03 | 40 | |
| QMIXInput Mode=Vision Only, Steps=5M, Backbone=Swin-Tiny2025.03 | 27.1 | |
| MAPPOInput Mode=Vision+Text†, Steps=5M, Backbone=Swin-Tiny2025.03 | 19.3 | |
| IPPOInput Mode=Vision Only, Steps=5M, Backbone=Swin-Tiny2025.03 | 18.2 | |
| IPPOInput Mode=Vision+Text†, Steps=5M, Backbone=Swin-Tiny2025.03 | 16.6 | |
| QTRANInput Mode=Vision Only, Steps=5M, Backbone=Swin-Tiny2025.03 | 2 | |
| IQLInput Mode=Vision Only, Steps=5M, Backbone=Swin-Tiny2025.03 | 0 | |
| VDNInput Mode=Vision Only, Steps=5M, Backbone=Swin-Tiny2025.03 | 0 |