Trajectory Planning on ROADWork
12.49ADE (pixel)OneVL
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| OneVLModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=Multimodal Auxiliary Supervision, Interpretability=Vision + Language2026.04 | 12.49 | 28.8 | 4.71 | |
| AR CoT+AnswerModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=AR-based Baseline, Interpretability=Language2026.04 | 13.18 | 29.98 | 10.74 | |
| COCONUTModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=Latent CoT Baseline2026.04 | 15.44 | 38.6 | 6.06 | |
| AR AnswerModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=AR-based Baseline2026.04 | 15.98 | 40.29 | 4.74 | |
| CODIModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=Latent CoT Baseline2026.04 | 16.45 | 44.28 | 6.73 | |
| SIM-CoTModel Size=4B, Base Model=Qwen3-VL, Reasoning Strategy=Latent CoT Baseline, Interpretability=Language2026.04 | 16.49 | 44.32 | 6.19 | |
| YNetReasoning Strategy=Previous State-of-the-Art2026.04 | 22.68 | 80.78 | — |