Question about the number of synthetic language demonstrations about e.t.

[alexpashevich]

Hi,

Thank you for your interest in our work. There are two different things: generating novel trajectories and replaying existing trajectories. In the first case, we randomly sample task parameters. In the second case, we take existing task parameters. For synthetic language demonstrations (e.g. generating novel trajectories), we use this script. The script spawns multiple threads and randomly samples task parameters. The script is almost identical to the original script provided by ALFRED authors, you can read more about the data generation here. Talking about replaying existing trajectories, we do it using this script in order to record an LMDB dataset used later for training. Feel free to also check Appendix A.2 of our paper where we compare agents trained on synthetic dataset of different sizes.

Hope this helps!