Clarification on nuScenes implementation details for reproduction

[pirotyyy]

Hi, thank you for the excellent work.

I understand that you may be busy, so please feel free to answer only when you have time.
I am trying to reproduce / port the SeerDrive nuScenes setting based on SSR, and I would really appreciate some clarification on the nuScenes-specific implementation details.

1. BEV resolution

   • The paper says the nuScenes BEV feature is 100x100, while future BEV map prediction is downsampled to 8x8.
   • Is the current semantic map predicted from 8x8 features, or from the full 100x100 BEV?
   • Does the planner consume the 8x8 future BEV feature, the 100x100 current BEV feature, or both?

2. Semantic head and semantic loss

   • Is the semantic head architecture for nuScenes the same as the released NAVSIM implementation?
   • Do you use CE or focal loss for semantic maps on nuScenes?

3. Object detection

   • Are object detection losses / heads also used in the nuScenes experiments?

4. Trajectory anchors and multimodality

   • The paper says nuScenes uses 6 trajectory modes. Are these anchors generated by k-means on the nuScenes training set?
   • Are anchors command-conditioned, or shared across commands?
   • Do the anchors include yaw / heading, or only xy trajectories?

5. Reward heads and inference-time mode selection

   • Are imitation reward and simulation reward both used for nuScenes?
   • During inference, how is the final trajectory mode selected?
     • imitation reward only?
     • simulation reward only?
     • weighted fusion of imitation reward and simulation reward?
     • WoTE or another filtering method?
   • If weighted fusion is used, could you share the weights?

6. Ego status inputs

   • Does the nuScenes model use ego status features such as velocity / acceleration / ego_lcf_feat?
   • Or is the planner trained only from sensor BEV features and trajectory anchors?

These details would be very helpful for reproducing the nuScenes results and adapting the model to SSR-style nuScenes data.

If it is easier for you, I would also be happy to check the nuScenes implementation by myself if you can share the relevant code or config privately. My email is: hhiroki1423@gmail.com .

Thank you again for your work and for any guidance you can provide.