about model checkpoint files: pytorch_model.bin and rl_checkpoint.pt

[LKlkun]

Hello @AIGeeksGroup team,

First, thank you for releasing Nav-R1 and making the code and models available — this is a very exciting project.

I have two points I’d like some clarification on:

In the model release, there is a file named pytorch_model.bin. Could you please confirm what exactly this file corresponds to (e.g., is it the supervised fine-tuned model, or the base model before RL, etc.)?

The README mentions that for RL training you’d resume from the checkpoint runs/navr1_rl/rl_checkpoint.pt. However I cannot find this file in the repository or linked artifacts. Could you please point me to where this rl_checkpoint.pt is available, or let me know if there are plans to release it in the near future?

Thank you for your time and for the great work. Looking forward to your response.

[Believeht029]

Thank you for the kind words and for trying Nav-R1!

• pytorch_model.bin: This file corresponds to the supervised fine-tuned (SFT) Nav-R1 weights in standard PyTorch/Hugging Face format. It is the model before the RL (GRPO) stage.
• rl_checkpoint.pt: We don’t currently include this checkpoint in the repository artifacts. We plan to release the RL checkpoint and will update the README with a download link once it’s available. In the meantime, you can produce it locally by running the RL stage; the training pipeline saves it as runs/navr1_rl/rl_checkpoint.pt (see train_pipeline.py). You can also resume any embodied task training from that path after generating it.

Please let us know if you need a short-term mirror or exact commands to reproduce the RL checkpoint.

[LKlkun]

Hello @Believeht029,

Thank you so much for your detailed explanation — that clarifies everything perfectly.

Nav-R1 has been very helpful for my current research, and I truly appreciate the effort your team has put into making it available.

If possible, could you please share a short-term mirror or temporary access link for the RL checkpoint that achieves performance close to your reported SOTA results?
It would be incredibly helpful for reproducing experiments and benchmarking while waiting for the official release.

Thanks again for your time and for sharing such valuable work with the community.

Best regards