Bugs in the AD-MLP code

[PointsCoder]

I noticed a bug in this code.

In the data_nuscene.pkl you provided, you are actually using the 3rd and the 7th ground truth waypoint coordinates as the acceleration and velocity value in Canbus.

This will cause severe label leakage and make your results much stronger.

I am wondering if there is a mistake in the release or you obtained your results with this data.

[E2E-AD]

We have reviewed the code and data again and have updated the revised training process at this link and on the GitHub website: https://pan.baidu.com/s/1IjN2pUUB35k2ScOitVsuGA?pwd=t4br. Thank you for your feedback.

We replied and addressed the relevant issues around September 20th, and subsequently provided updates on GitHub here.

By re-examining the data used for training, we have arrived at conclusions that differ slightly from what was reported in the previous technical report. Overall, under open-loop evaluation, the L2 error remains superior, while the collision rate has decreased slightly and is still slightly lower than that of VAD. We believe that the fundamental conclusion remains unchanged, which is that superior open-loop evaluation performance can be achieved on the NuScenes dataset with just past motion information and a simple model. This may introduce some challenges in measuring and comparing different methods in this evaluation.

[xinshuoweng]

Thanks for your effort to double-check on this! I noticed that the released model and data pkl files (in May 2023) are still the old ones that have the GT leakage. Would you mind updating the released model files and data pkl file to reflect the newly trained model without GT leakage? Appreciate it!