Clarification / release request for navigation task

[lqh52]

Hi V-JEPA team,

Thanks for the great paper and release.

I’m especially interested in the navigation results in:

• Section 3.4: Navigation Planning
• Figure 9
• Table 7: Open Loop Navigation Planning

Could you clarify or release the exact setup used for these results?

The main points that seem unclear are:

1. Frozen or finetuned encoder?
   In first paragraph of section 3 (Results), the paper text says V-JEPA 2.1 is used as a frozen encoder, but the Table 7 caption says “we finetune V-JEPA 2.1 on robot navigation datasets”.

2. Exact world model recipe
   Section 3.4 says you train a CDiT on top of V-JEPA 2.1, predict clean representations instead of noise, and use DDIM.
   Could you share the exact config / architecture changes relative to NWM?

3. Reproducibility
   If possible, could you release the config / checkpoint / eval code for the navigation experiment?

This part of the paper is very interesting, and having the exact setup would make reproduction much easier.

Thanks again.