For EQA testing, how do you decide where to start the agent?

[danielgordon10]

This can have a dramatic effect on the accuracy of the agent. Do you pick a random start location, and if so, do you record that in the dataset so it can be replicated?

[abhshkdz]

The set of shortest paths are available here: https://github.com/facebookresearch/EmbodiedQA#download-and-preprocess-the-dataset, and for evaluation, we spawn the agent 10, 30, 50 actions away from the target along these shortest paths: https://github.com/facebookresearch/EmbodiedQA/blob/master/training/train_nav.py#L102

[danielgordon10]

Got it. Thanks.

[danielgordon10]

Would it be possible for you to provide starting coordinates for each question for reproducibility?