Rotations

[msinto93]

In the paper, you say that "augmenting the demonstration data with different rotations is critical because the proposed two-stage network is sensitive to directional bias."

As there is nothing in the code that performs rotations I assume you do this beforehand externally on the dataset and not 'live' during training.

My question is regarding how you rotate the trajectories along-with the environmental grids. As we have an even-numbered grid size (80x80) surely rotating the trajectories will result in the rotated trajectories having a different start state from the [40,40] state? Is this a problem? Is it important that all future trajectories start from [40,40]? If so is there a way to rotate the trajectories to ensure this start state remains constant?

[yfzhang]

Yes, the rotation is done as part of the preprocessing.
The trajectory will always start from the [40,40] state. To make this possible, local map/environment will be cropped accordingly to make sure the trajectory starts from the center. Then trajectory and the grid environment will be rotated together.