search

Thinklab-SJTU / Bench2Drive

[NeurIPS 2024 Datasets and Benchmarks Track] Closed-Loop E2E-AD Benchmark Enhanced by World Model RL Expert
Other
1.3k stars 85 forks source link

ThinkTwice & DriveAdapter Reproduction #4

Closed ryhnhao closed 5 months ago

ryhnhao commented 5 months ago

Hi, Great and meaningful work on closed-loop benchmarking!

Now (Jun. 11th, 2024), the repo seems to have 4 methods released (AD-MLP & TCP & UniAD & VAD :) I wonder when ThinkTwice & DriveAdapter will be released. Do you have a schedule?

Thanks a lot, and look forward to your reply!

jiaxiaosong1002 commented 5 months ago

@ryhnhao Hi, since ThinkTwice & DriveAdapter requires expert feature at every layer, we do not intend to open source Think2Drive related code recently. Thanks

ryhnhao commented 5 months ago

@jiaxiaosong1002 Hi, thanks for your in-time reply.

I also notice that the paper claimed that TCP also employs expert feature. (As "As a result, methods (TCP/ThinkTwice/DriveAdapter) with expert feature distillation outperforms those without (VAD/UniAD) by a large margin.")

I wonder what's the difference between the expert feature of ThinkTwice&DriveAdapter and that of TCP? And why they are chosen different? To the best of your knowledge, are there other expert feature alternatives available for carla-v2 (apart from your great work Think2Drive)

Thanks a lot, and look forward to your reply!

And BTW, good luck to your papers :)

jiaxiaosong1002 commented 5 months ago

@ryhnhao Thanks!

We try our best to strictly follow the official implementation of each baseline. In TCP, it only uses 1D expert features as it does not involve any BEV transformation. We release the 1D expert feature in the Bench2Drive dataset (expert_assessment subfolder). In ThinkTwice and DriveAdapter, they use every layer of expert feature including both 1D feature and BEV feature.

I am not aware of any other expert with feature. However, there is a good open-sourced rule-based expert (https://github.com/autonomousvision/carla_garage/tree/leaderboard_2) if you need an expert.

ryhnhao commented 5 months ago

@jiaxiaosong1002 Thank you very much. I don't have further issues.