Request for DAgger Data Collection Code

[Lelouchrx]

Hi, thank you for making this excellent work and releasing the codebase.
We are currently trying to reproduce the VLN results reported in the paper and have found the repository very helpful. However, we were unable to locate the code used for DAgger data collection and dataset generation during training.
Would it be possible to share the DAgger data collection pipeline (or provide some guidance on how the training data was generated)? Access to this part of the implementation would greatly help us reproduce the reported results more faithfully and better understand the training procedure described in the paper.
We completely understand if there are constraints on releasing certain components. In that case, any additional details about the data collection process, hyperparameters, or implementation specifics would also be greatly appreciated.
Thank you again for your valuable contribution and for maintaining this project.

[waynechu1021]

Hi, thank you for making this excellent work and releasing the codebase. We are currently trying to reproduce the VLN results reported in the paper and have found the repository very helpful. However, we were unable to locate the code used for DAgger data collection and dataset generation during training. Would it be possible to share the DAgger data collection pipeline (or provide some guidance on how the training data was generated)? Access to this part of the implementation would greatly help us reproduce the reported results more faithfully and better understand the training procedure described in the paper. We completely understand if there are constraints on releasing certain components. In that case, any additional details about the data collection process, hyperparameters, or implementation specifics would also be greatly appreciated. Thank you again for your valuable contribution and for maintaining this project.

Thank you for your attention. For the DAgger part, we use the same one as in the StreamVLN repository, which you can find here. You only need to replace the inference process with ours, and then process this newly generated data according to the data processing part in the README.