In (LQRStep.lqr_forward) https://github.com/locuslab/mpc.pytorch/blob/master/mpc/lqr_step.py, we have the following:
if t < T-1:
if isinstance(true_dynamics, mpc.LinDx):
F, f = true_dynamics.F, true_dynamics.f
new_xtp1 = util.bmv(F[t], new_xut)
if f is not None and f.nelement() > 0:
new_xtp1 += f[t]
else:
new_xtp1 = true_dynamics(
Variable(new_xt), Variable(new_ut)).data
new_x.append(new_xtp1)
dx.append(new_xtp1 - x[t+1])(line 218 onwards)
This effectively removes the last state from the trajectory. If there is a cost associated with it, however, doesn't that make the gradient computation incorrect? We won't have dx (or the final x) in the backward function of the same class.
Note that further down the same function, the cost is in fact updated:
if isinstance(true_cost, mpc.QuadCost):
C, c = true_cost.C, true_cost.c
obj = 0.5*util.bquad(new_xut, C[t]) + util.bdot(new_xut, c[t])
else:
obj = true_cost(new_xut)which means the terminal state contributes to the cost, but is dropped from the list of states used elsewhere.
In (
LQRStep.lqr_forward) https://github.com/locuslab/mpc.pytorch/blob/master/mpc/lqr_step.py, we have the following:(line 218 onwards)
This effectively removes the last state from the trajectory. If there is a cost associated with it, however, doesn't that make the gradient computation incorrect? We won't have dx (or the final x) in the backward function of the same class.
Note that further down the same function, the cost is in fact updated:
which means the terminal state contributes to the cost, but is dropped from the list of states used elsewhere.