Why is nominal trajectory not used in iLQR control law?

[sklaw]

It seems the following line computes new action without any use of nominal trajectory:

gps/python/gps/algorithm/traj_opt/traj_opt_lqr_python.py

Line 257 in 82fa6cc

traj_distr.K[t, :, :].dot(mu[t, idx_x]) + traj_distr.k[t, :]

But in the following papers, nominal trajectory seems to be essential in the whole iLQR algorithm:
equation (8b) in this paper
equation (3) in this paper

[dujinyu]

@sklaw hello, I am confused by the linear-gaussian policy u = K * x + k, if we use iLQR, why not the cotrol u = \bar{u} + k + K * (x - \bar{x})

[sklaw]

@sklaw hello, I am confused by the linear-gaussian policy u = K * x + k, if we use iLQR, why not the cotrol u = \bar{u} + k + K * (x - \bar{x})

Hi. I'm also confused by this part, too. It seems the author of this code assumed the nominal trajectory is 0... But if so the whole trajectory optimization would be really weird