Autonomous Driving Planning on NAVSIM Navhard v2 (Stage 1)
1.426ADE@4s (m)ReCogDrive + CLAP
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| ReCogDrive + CLAPBackbone=ReCogDrive, Number of Cameras=3, Prompt Strategy=CLAP2026.05 | 1.426 | — | — | — | — | — | — | — | |
| DriveVLA-W0 + CLAPBackbone=DriveVLA-W0, Number of Cameras=1, Prompt Strategy=CLAP2026.05 | 1.804 | — | — | — | — | — | — | — | |
| ReCogDrive + Unconstrained Soft PromptBackbone=ReCogDrive, Number of Cameras=3, Prompt Strategy=Unconstrained Soft Prompt2026.05 | 1.835 | — | — | — | — | — | — | — | |
| DriveVLA-W0 + Unconstrained Soft PromptBackbone=DriveVLA-W0, Number of Cameras=1, Prompt Strategy=Unconstrained Soft Prompt2026.05 | 1.895 | — | — | — | — | — | — | — | |
| DriveVLA-W0Backbone=DriveVLA-W0, Number of Cameras=12026.05 | 2.116 | — | — | — | — | — | — | — | |
| ReCogDriveBackbone=ReCogDrive, Number of Cameras=32026.05 | 2.125 | — | — | — | — | — | — | — | |
| ReCogDrive + Explicit NoticeBackbone=ReCogDrive, Number of Cameras=3, Prompt Strategy=Explicit Notice2026.05 | 2.137 | — | — | — | — | — | — | — | |
| Alpamayo-R1.5 + CLAPBackbone=Alpamayo-R1.5, Number of Cameras=4, Prompt Strategy=CLAP2026.05 | 2.281 | — | — | — | — | — | — | — | |
| Alpamayo-R1.5 + Unconstrained Soft PromptBackbone=Alpamayo-R1.5, Number of Cameras=4, Prompt Strategy=Unconstrained Soft Prompt2026.05 | 2.385 | — | — | — | — | — | — | — | |
| SFT Checkpointtraining=Supervised Fine-Tuning2026.02 | 2.539 | 0.953 | 0.789 | 0.978 | 0.993 | 0.836 | 0.976 | 0.651 | |
| DriveVLA-W0 + Explicit NoticeBackbone=DriveVLA-W0, Number of Cameras=1, Prompt Strategy=Explicit Notice2026.05 | 2.581 | — | — | — | — | — | — | — | |
| Alpamayo-R1.5Backbone=Alpamayo-R1.5, Number of Cameras=42026.05 | 2.998 | — | — | — | — | — | — | — | |
| Alpamayo-R1.5 + Explicit NoticeBackbone=Alpamayo-R1.5, Number of Cameras=4, Prompt Strategy=Explicit Notice2026.05 | 3.13 | — | — | — | — | — | — | — | |
| Constrained GRPO, ScRewobjective=Constrained RL, scalarization=Scalarized Rewards (ScRew)2026.02 | 3.344 | 0.955 | 0.887 | 0.987 | 0.993 | 0.868 | 0.976 | 0.743 | |
| Constrained GRPO, ScAdvobjective=Constrained RL, scalarization=Scalarized Advantages (ScAdv)2026.02 | 3.361 | 0.957 | 0.928 | 0.989 | 0.998 | 0.796 | 0.977 | 0.774 | |
| GRPO, ScAdv, Informed Weightsscalarization=Scalarized Advantages (ScAdv), weighting=Informed Weights2026.02 | 3.495 | 0.932 | 0.765 | 0.971 | 0.993 | 0.89 | 0.976 | 0.623 | |
| GRPO with EPDMSreward_function=EPDMS (multiplicative)2026.02 | 6.9 | 0.897 | 0.8 | 0.947 | 0.99 | 0.944 | 0.978 | 0.612 | |
| GRPO, ScAdv, Equal Weightsscalarization=Scalarized Advantages (ScAdv), weighting=Equal Weights2026.02 | 7.172 | 0.908 | 0.661 | 0.9 | 0.989 | 0.954 | 0.973 | 0.49 | |
| GRPO, ScRew, Informed Weightsscalarization=Scalarized Rewards (ScRew), weighting=Informed Weights2026.02 | 8.509 | 0.887 | 0.643 | 0.887 | 0.987 | 0.967 | 0.978 | 0.463 | |
| GRPO, ScRew, Equal Weightsscalarization=Scalarized Rewards (ScRew), weighting=Equal Weights2026.02 | 9.05 | 0.891 | 0.625 | 0.876 | 0.989 | 0.968 | 0.976 | 0.445 | |
| PDM-Closed (privileged)type=privileged rule-based planner2026.02 | — | 0.944 | 0.988 | 1 | 0.995 | 1 | 0.877 | — |