Accuracy of only fine-tuned models for one epoch based on oracle VLN trajectories.

[MuHeDing]

Hello,author
This is a very interesting and inspiring work, and I want to know the experimental details of this paper: "Training is divided into two stages. First, we fine-tune the model for one epoch based solely on the oracle VLN trajectory. Then, we use the model to collect DAgger trajectory and continue training for one epoch using both VLN and general multimodal data."

Question1: the acccuracy of only fine-tuned models for one epoch based on oracle VLN trajectories, i.e., first-stage accuracy.

Question2: oracle VLN trajectory datasets include R2R , R2R-EnvDrop, ScaleVLN dataset and RxR?

[kellyiss]

Hi, thanks for attention.
Q1: Please refer to the first row of Table 3.
Q2: To save training costs, ScaleVLN is only used in the second stage, and we did not collect DAGGER data on ScaleVLN.