Closed hjsuh94 closed 3 years ago
One solution might be to accept a u_bound_relative
parameter that becomes additive on state.
I'll give it a stab when I PR the contact stuff.
After trying out many examples, the following four bounds will cover everything:
x_bound_abs
: this helps when there are actual bounds in the system (joint limits, ground, etc.)x_bound_rel
: this helps to establish a trust region around current nominal trajectory.u_bound_abs
: this helps when the system is force / velocity controlled.u_bound_rel
: this helps when the system is position controlled. In addition, quasistatic systems put a cost on du
instead of the absolute u
, which might also be problematic.
For position controlled systems, it could be the case that
ubound
depends on a nominal point onx_trj
, but the current code structure only allows uniform bounds.