openocl / openocl

Open Optimal Control Library for Matlab. Trajectory Optimization and non-linear Model Predictive Control (MPC) toolbox.

Open Optimal Control Library

Visit the website openocl.github.io to download the toolbox, and have a look at the API Docs and the examples.

Get started

doc ocl
ocl.examples.cartpole
open ocl.examples.cartpole

Contribute

If you would like to participate in the project, you are very welcome! Contribution guidelines

Some Models

• robotics: https://github.com/jkoendev/openocl_models (very experimental)
• airborne wind energy: https://openawe.github.io/ (outdated, very experimental)

Citing and publications

If you use OpenOCL, please cite us and consider giving us a star on this repository.

Legal notice

Please see license information in the LICENSE file in the github project.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 642682.

Use without warranty.

info [at] openocl.org

openocl's Issues

Assign Var state trajectory with N times nx matrix

Var complains about not matching Var size. Workaround is to assign a column vector.

Restructure bound handling

The internal handling of bounds is quiet complicated and can be simplified.
Maybe bounds are not needed at all for the user. User can just implement constraints. Solver will internally convert constraints to bounds..

Support Compilation of CasadiFunctions

Scaling

add scaling by max and min value
or mean and variance

Initial guess, state trajectory

Would be nice to be able to initialize the state trajectory and automatically with it the collocation states

subsref slicing indexing

Would be a nice feature for the Var class

e.g. solution['state']['p']

Support Parameter

Need to support 2 types of parameters:
plain, just to make quick recompilation
as variable, to optimize for this parameter

are constraints on variable parameters needed? could result in static optimization

Add tests for the core functionalities of the toolbox

Support JIT

Support independent variables other than time

Add testing and benchmark funcitonality

GetInitialGuess needs to be called before calling the solver

Sometime you don't want to call GetInitialGuess if you already have an initial guess from previous runs

Assign multi dimensional trajectories to Var

states = guess.get('state').flat;
guess.get('integratorVars').get('state').set(states);

doesn't work and gives no error

Simulation callback should be optional

also simulationCallbackSetup

discretizationpoints -> number of control intervals

Code cleanup I

• move getOptions to bottom

- make inhertiance of casadi handlers clear, casadiCollocation, casadiOCPHandler

Integrator in collocation

Let collocation interval function return integration constraints etc.
Could implement different integrators, implicit or explicit. not only collocation integrator.
Still solve the implicit integrators within the NLP, and not by calling and external integrator.

Could unify multishooting and collocation.

return values of the simulator

Implement sol.get('state').get('v',1)

currently wrong behaviour

Simulator overrides already set parameters with symbolic variables

Time optimal control

Time variables are somehow special. Only end time in the end, or a time variable for each shooting interval?
Is it possible to unify time variable with parameters?

Constraints on the first state in the trajecotry

Similar to terminalContraints

Pass timed control vector to simulator

Implement mathematical operators for Var

Var + Var
Var + 5
Var * Var etc.

Setting 3x1 MX on 1x1 Var

K>> parameters

parameters =

MX(p)
K>> parameters.size()

ans =

 3     1

K>> self.parameters

ans =

Var with properties:

       id: []
  subVars: [1x1 Var]
thisValue: []
 thisSize: [1 1]
 compiled: 1
   varIds: [1 java.util.HashMap]

K>> self.parameters.set(parameters)

Implement missing operators for casadi variables

CasadiVar < casadi.SX
implement missing operators:

• sum(CasadiVar) -> instead of sum_2

Wrap toolbox in a matlab package

Implement model blocks with inputs and outputs

similar to Simulink blocks, could make modelling very modular, and flexible. E.g. replace tether model without touching aerodynamic model.
Might increase complexity, and might not be intuitive for the user. Possibly slower performance.

Improve NLP construction time (casadi)

Use either checkpointing or mapaccum

Finish casadi workshop code generation and merge branch to master

Unify parameter handling

Var doesn't spit out warnings

I don't see the warnings of Var in matlab 2016b

Path constraints on initial state

The first state is not subject to path constraints because the path constraints are dependent on state algebraic variables etc. associated with the collocation states in the middle and end of the integration interval.

Measure construction time and solving time

OCP methods should be optional

No need to implement empty methods if they are not needed.
e.g. discrete Cost, pathConstraints..

Test slice casadi vars to pure symbolics

casadi concats all variables and then slices at the moment. Should split the vars instead.
jacobian(c,p) not possible because p is not atomic.

Don't use Var internally

Var is convenient for the user to manage initialGuess and solution, but it should not be used internally, e.g. for collocation variables..

detecting parameters need to take changes in bounds into account

when solving a problem again with different set of parameters

Add soft constraints feature

Relax constraints to soft constraints

Add plotting functionality for cost terms and constraints

Scaling and detect parameters false by default

Use states instead of state everywhere

Scaling should be part of NLP not NLPsolver

Named output equations for systems

addOutput('Drag', 0.5*CX*rho*v^2*A)

Plot large values in jacobian in every iteration

Add support for metis

Return trajectory when indexing or slicing

e.g. solution.get('p',1)
if there are multiple 3x1 p, get the first element from each of them

Make NLP solver not only for OCPs

Either we need to remove NLP solving functionality and turn NLPsolver to an OCP solver or add general NLP functionality, static optimization

Include examples of gianni and add tests

Time dependent model

either support them or not, right now it's not very consistent