octo process traj_transforms

[Lucky-Light-Sun]

Hello, recently I was reviewing the Octo data processing code and noticed an issue in the traj_transformer, specifically in the chunk part. Could you please check if there might be a minor problem in the code?

As shown in the figure, the relative_goal_timestep may be t - (windowsize - 1) + w + h? t means timestep here, w means the history window position, h means the horizon action postion.

Or maybe I misunderstood. Could you help explain it?

relative_goal_timestep = goal_timestep[:, None, None] - (
        t - (window_size + 1) + w + h       # xyg noted: 这里有点像是 t - (window_size - 1) + w + h ???
    )

https://github.com/octo-models/octo/blob/main/octo/data/traj_transforms.py#L84-L87