First-hand information for everyone
Researchers have introduced NegWAM, a safety‑post training method that teaches autonomous driving models to stay away from dangerous trajectories.
The core idea is to generate “negative” trajectories that look almost like expert driving but carry hidden safety risks. These trajectories are used as repulsive anchors during training. A scene‑conditioned rider token encodes how a trajectory fits the surroundings. The additional network parts are discarded when the model runs, so real‑time speed is unchanged.
On the NAVSIM v2 benchmark, NegWAM scores 40.4 EPDMS on the hard set, a 4.1 % relative gain over the best scorer‑free baseline, and 93.3 EPDMS on the overall test set, surpassing the next best method by 1.1 points. The improvement is achieved with almost the same inference cost: latency rises by only 6.5 ms (about 0.75 %) and frames per second falls by 0.01.
| Methods | Sensor | Stage |
| NC ↑ |
| DAC ↑ |
| DDC ↑ |
| TLC ↑ |
| EP ↑ |
| TTC ↑ |
| LK ↑ |
| HC ↑ |
| EC ↑ |
| EPDMS ↑ |
| S1 | 94.4 | 98.8 | 100 | 99.5 | 100 | 93.5 | 99.3 | 87.7 | 36.0 |
| PDM-Closed | - | S2 | 90.5 | 90.6 | 95.4 | 98.4 | 100 | 86.6 | 74.2 | 91.9 | 29.7 | 56.6 |
| E2E-based Methods |
| S1 | 96.0 | 79.7 | 97.4 | 99.5 | 81.3 | 93.1 | 90.8 | 96.8 | 73.8 |
| DiffusionDrive [41] | MV | S2 | 82.1 | 72.2 | 88.5 | 98.7 | 85.1 | 78.8 | 49.2 | 89.3 | 71.2 | 24.2 |
| S1 | 96.6 | 80.5 | 96.3 | 99.3 | 82.3 | 94.9 | 91.5 | 97.7 | 67.8 |
| GuideFlow [42] | MV | S2 | 87.3 | 76.7 | 88.8 | 99.2 | 84.3 | 85.1 | 49.7 | 93.1 | 44.2 | 27.1 |
| VLA-based Methods |
| S1 | 96.8 | 83.3 | 99.0 | 99.6 | 84.6 | 95.3 | 96.4 | 97.6 | 78.2 |
| DriveVLA-W0 [43] | SV | S2 | 76.8 | 64.3 | 79.9 | 98.3 | 89.2 | 75.0 | 46.8 | 95.8 | 53.1 | 24.4 |
| S1 | 98.4 | 94.3 | 97.8 | 99.9 | 85.7 | 97.2 | 94.2 | 97.6 | 72.1 |
| SpanVLA [44] | MV | S2 | 86.9 | 84.3 | 87.1 | 98.2 | 85.5 | 82.7 | 62.3 | 96.8 | 67.4 | 40.1 |
| World-Model-based Methods |
| S1 | 97.1 | 88.7 | 99.2 | 99.6 | 84.2 | 97.1 | 96.2 | 97.8 | 69.8 |
| DriveLAW [5] | SV | S2 | 82.3 | 67.9 | 83.7 | 98.1 | 84.7 | 78.5 | 46.7 | 96.3 | 60.6 | 30.4 |
| S1 | 96.3 | 87.8 | 98.2 | 99.6 | 83.1 | 96.9 | 94.9 | 97.6 | 76.9 |
| DriveFuture† [45] | MV | S2 | 82.3 | 78.8 | 88.1 | 98.4 | 83.6 | 79.6 | 47.6 | 97.0 | 75.9 | 34.6 |
| S1 | 97.3 | 90.7 | 99.4 | 100 | 83.3 | 97.3 | 97.3 | 97.6 | 60.9 |
| EponaV2 [6] | SV | S2 | 83.6 | 78.0 | 88.0 | 98.9 | 86.0 | 80.3 | 50.1 | 96.1 | 52.0 | 36.1 |
| S1 | 96.8 | 88.7 | 99.1 | 99.6 | 95.9 | 95.6 | 95.6 | 97.8 | 80.4 |
| DriveWAM [7] | SV | S2 | 79.5 | 75.8 | 87.8 | 98.1 | 96.5 | 75.7 | 48.7 | 95.8 | 73.7 | 37.7 |
| S1 | 96.6 | 93.1 | 98.8 | 99.6 | 96.9 | 95.8 | 96.9 | 97.8 | 78.7 |
| BeyondDrive [17] | SV | S2 | 81.2 | 77.7 | 85.6 | 98.2 | 97.0 | 77.7 | 47.8 | 95.4 | 72.2 | 38.8 |
| S1 | 96.6 | 94.4 | 98.8 | 99.6 | 95.8 | 96.0 | 96.4 | 97.8 | 79.6 |
| Ours | SV | S2 | 79.2 | 78.0 | 86.4 | 98.1 | 96.3 | 75.3 | 48.4 | 95.7 | 73.7 | 40.4 |
The method also generalises without extra training: a zero‑shot closed‑loop simulation on the HUGSIM environment shows route completion of 90.3 % and HD‑Score of 81.7, higher than the leading model's 76.6 % and 33.9.
These results demonstrate that NegWAM can strengthen safety in challenging driving scenarios while maintaining performance and deployment efficiency. However, the gains are reported only for the specific benchmark and zero‑shot transfer. The ability of the approach to handle the variety and unpredictability of real‑world driving remains to be verified.