Terminal constraints in python

User Questions
1

Hi there,

Thanks for the great tool that you’re developing.
I’m trying to slightly modify your pendulum example by transferring the terminal state from the cost function to the constraints.
I tried removing the cost on the final state and adding the following lines according to the pdf specifying the problem formulation.

# terminal constraints
ocp.constraints.Jbx_e  = np.eye(nx)
# ocp.constraints.Jsbx_e = 1e-2*np.eye(nx)
ocp.constraints.ubx_e  = np.array([0.0, np.pi, 0.0, 0.0])
ocp.constraints.lbx_e  = np.array([0.0, np.pi, 0.0, 0.0])

Unfortunately, it looks like the solver does not take this constraint into account since it converges to a “flat” solution (always 0 on the states and on the control).
Moreover, after these lines, if I ask for ocp.constraints.Jbx_e, I get the following error :
AttributeError: ‘AcadosOcpConstraints’ object has no attribute ‘_AcadosOcpConstraints__Jbx_e’
I guess this comes from the fact that in AcadosOcp.py, the declaration of self.__Jbx_e is commented…
I tried uncommenting it, resulting in no error, but the result of the solver is still the same !

Do you have any idea how to solve this ?

Thanks,

Fb

2

Hi @byfra,

welcome to the Forum and thanks for the post!

This is an error in the Python documentation.
The matrix Jbx_e and all similar matrices (Jbx, Jbu,…) are not actually part of the Python interface (yet).
Instead you have to specify a vector idxbx_e, which is a zero based index vector that defines which components should be bounded.
This means, instead of

ocp.constraints.Jbx_e = np.eye(nx)

you actually have to write

ocp.constraints.idxbx_e = np.array(range(nx))

I will try to implement the setters Jbx etc, which will then internally convert the matrix into the zero based vector mentioned above.

Thanks for pointing this out!

Best,
Jonathan

Edit:
Actually the conversion seems to be partly there already, especially for Jbx_e, so I will try to have another look tomorrow…

3

Thanks for the quick reply, I confirm that the conversion works.
I missed the following setting :
ocp.dims.nbx_e = nx
Now, it works as expected !
Thanks,
Fb

4

Great that this worked.
Setting/checking the dimensions automatically in python is also on the TODO list :sweat_smile: