Inquiry about pre-trained AutoVLA weights

[ytzhao]

Hi ucla-mobility team,

First of all, congratulations on AutoVLA being accepted by NeurIPS 2025! This is truly excellent work, and the integration of chain-of-thought reasoning with physical action tokenization is very inspiring.

I'm planning to build upon AutoVLA as a baseline for my research. I noticed from the README that the AutoVLA checkpoints are planned for release in March 2026 ("2026/03: AutoVLA checkpoints"). I would be grateful if you could kindly share:

1. Checkpoint release timeline: Is the release still on track for March 2026? If there are any blockers, an updated timeline would be greatly appreciated.
2. Pre-trained weights: Would it be possible to share the pre-trained weights (both SFT and RFT versions, if available)? Having access to the checkpoints would significantly help the community build upon your excellent work.
3. Evaluation guide: Once the weights are available, could you provide guidance on how to run evaluation? Specifically:
   • Which datasets are required (nuPlan / nuScenes / Waymo)?
   • Are the evaluation scripts complete for reproducing the paper results?
   • Is there a quick-start example for inference with pre-trained weights?

I completely understand if there are technical or administrative delays. Thank you for open-sourcing the training code already, and please let me know if there's any way the community can help with the release process.

Thank you for your time and excellent contribution!

Best regards,
ytzhao