Dear Simon Gerstenecker,
First of all, thank you very much for creating such a challenging benchmark. I believe it will have a significant impact on the autonomous driving community.
To investigate the robustness of our BridgeDrive, we conducted a thorough evaluation on your benchmark. Here are the results:
| Method |
Bench2Drive |
Fail2Drive In-Distribution |
Fail2Drive Generalization |
| DS ↑ |
DS ↑ |
SR(%) ↑ |
HM ↑ |
DS ↑ |
SR(%) ↑ |
HM ↑ |
| TCP |
59.9 |
24.7 |
39.1 |
30.3 |
24.5 (-0.8%) |
31.4 (-19.7%) |
27.5 (-9.1%) |
| UniAD |
45.8 |
47.5 |
36.3 |
41.2 |
44.0 (-7.4%) |
27.6 (-24.0%) |
33.9 (-17.6%) |
| Orion |
77.8 |
53.0 |
52.0 |
52.5 |
51.2 (-3.4%) |
46.0 (-11.5%) |
48.5 (-7.7%) |
| HiP-AD |
86.8 |
74.1 |
70.7 |
72.4 |
67.1 (-9.4%) |
56.7 (-19.8%) |
61.5 (-15.1%) |
| SimLingo |
85.1 |
82.6 |
79.3 |
80.9 |
71.7 (-13.2%) |
55.0 (-30.6%) |
62.2 (-23.1%) |
| TFv5 |
84.2 |
83.3 |
78.5 |
80.8 |
75.4 (-9.5%) |
61.1 (-22.2%) |
67.5 (-16.5%) |
| TFv6 |
95.2 |
90.2 |
93.3 |
91.7 |
79.5 (-11.9%) |
70.7 (-24.2%) |
74.8 (-18.4%) |
| BridgeDrive (Ours) |
96.3 |
91.6 |
95.0 |
93.3 |
81.9 (-10.5%) |
75.0 (-21.1%) |
78.3 (-16.0%) |
For completeness, I could also send you the evaluation JSON files via email.
I believe including these results in your paper would greatly enhance the comprehensiveness of your work, especially since none of the compared methods are diffusion-based—BridgeDrive would help fill that gap.
I hope you find this message helpful, and I look forward to a fruitful discussion. Cheers!
Dear Simon Gerstenecker,
First of all, thank you very much for creating such a challenging benchmark. I believe it will have a significant impact on the autonomous driving community.
To investigate the robustness of our BridgeDrive, we conducted a thorough evaluation on your benchmark. Here are the results:
For completeness, I could also send you the evaluation JSON files via email.
I believe including these results in your paper would greatly enhance the comprehensiveness of your work, especially since none of the compared methods are diffusion-based—BridgeDrive would help fill that gap.
I hope you find this message helpful, and I look forward to a fruitful discussion. Cheers!