inconsistency about the initialization distribution of noise

hi authors,

i see in the training part, `x0 = torch.randn(x_traj.shape, device=device)` the `x0` noise is drawn from normal gaussian,
but in inference part, the `noise` is from `noise = torch.rand(1, pred_horizon, action_dim).to(device)` which is uniform distribution.

is this inconsistency a bug ?
any considerations under the hood ?

thanks.