Skip to content
forked from pyomeca/bioptim

An optimization framework that links CasADi, ipopt and biorbd for Optimal Control Problem

License

Notifications You must be signed in to change notification settings

EveCharbie/bioptim

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

logo

Bioptim is an optimal control program (OCP) framework for biomechanics. It is based on the efficient biorbd biomechanics library and benefits from the powerful algorithmic diff provided by CasADi. It interfaces the robust Ipopt and the fast Acados solvers to suit all your needs for solving OCP in biomechanics.

Status

Type Status
License License
Continuous integration Build status
Code coverage codecov
DOI DOI

The current status of bioptim on conda-forge is

Name Downloads Version Platforms MyBinder
Conda Recipe Conda Downloads Conda Version Conda Platforms Binder

Try bioptim

Anyone can play with bioptim with a working (but slightly limited in terms of graphics) MyBinder by clicking the following badge

Binder

As a tour guide that uses this binder, you can watch the bioptim workshop that we gave at the CMBBE conference on September 2021 by following this link: https://youtu.be/z7fhKoW1y60

Table of Contents

Testing bioptim

How to install

Defining our optimal control problems

A first practical example
A more in depth look at the `bioptim` API
Examples
Performance
Troubleshooting

Citing

How to install

The preferred way to install for the lay user is using anaconda. Another way, more designed for the core programmers, is from the sources. While it is theoretically possible to use bioptim from Windows, it is highly discouraged since it will require manually compiling all the dependencies. A great alternative for Windows users is Ubuntu on Windows supporting Linux.

Installing from Anaconda (For Windows, Linux, and Mac)

The easiest way to install bioptim is to download the binaries from Anaconda repositories. The project is hosted on the conda-forge channel (https://anaconda.org/conda-forge/bioptim).

After having appropriately installed an anaconda client [my suggestion would be Miniconda (https://conda.io/miniconda.html)] and loaded the desired environment to install bioptim in, just type the following command:

conda install -c conda-forge bioptim

This will download and install all the dependencies and install bioptim. And that is it! You can already enjoy bioptiming!

Installing from the sources (For Linux, Mac, and Windows)

Installing from the sources is as easy as installing from Anaconda, with the difference that you will be required to download and install the dependencies by hand (see the section below).

Dependencies

bioptim relies on several libraries. The most obvious one is the biorbd suite (including indeed biorbd and bioviz), but extra libraries are required. Due to the different dependencies, it would be tedious to show how to install them all here. The user is therefore invited to read the relevant documentation.

Here is a list of all direct dependencies (meaning that some dependencies may require other libraries themselves):
Python | numpy | scipy | packaging | setuptools | matplotlib | pandas | pyomeca | CasADi | rbdl-casadi compiled with the CasADi backend | tinyxml | biorbd | vtk | PyQt | bioviz | graphviz | Ipopt | Acados | pyqtgraph | pygmo (only for inverse optimal control)
and optionally: The linear solvers from the HSL Mathematical Software Library with install instructions here.

Linux - Installing dependencies with conda

All these (except for ̀Acados and the HSL lib) can easily be installed using (assuming the anaconda3 environment is loaded if needed) the pip3 command or the Anaconda's following command:

conda install biorbd bioviz python-graphviz -cconda-forge

Since there is no Anaconda nor pip3 package of Acados, a convenient installer is provided with bioptim. The installer can be found and run at [ROOT_BIOPTIM]/external/acados_install_linux.sh. However, the installer requires an Anaconda3 environment. If you have an Anaconda3 environment loaded, the installer should find itself where to install it. If you want to install it elsewhere, you can provide the script with a first argument which is the $CONDA_PREFIX. The second argument that can be passed to the script is the $BLASFEO_TARGET. If you don't know what it is, it is probably better to keep the default. Please note that depending on your computer architecture, Acados may or may not work correctly.

Mac - Installing dependencies with conda

Equivalently for MacOSX:

conda install casadi 'rbdl' 'biorbd' 'bioviz' python-graphviz -cconda-forge

Since there is no Anaconda nor pip3 package of Acados, a convenient installer is provided with bioptim. The Acados installation script is [ROOT_BIOPTIM]/external/acados_install_mac.sh. However, the installer requires an Anaconda3 environment. If you have an Anaconda3 environment loaded, the installer should find itself where to install it. If you want to install it elsewhere, you can provide the script with a first argument, the $CONDA_PREFIX. The second argument that can be passed to the script is the $BLASFEO_TARGET. If you don't know what it is, it is probably better to keep the default. Please note that depending on your computer architecture, Acados may or may not work correctly.

Windows - Installing dependencies with conda

Equivalently for Windows:

conda install casadi 'rbdl' 'biorbd' 'bioviz' python-graphviz -cconda-forge

There is no Anaconda nor pip3 package of Acados. To use the Acados solver on Windows, one must compile it themselves.

The case of HSL solvers

HSL is a collection of state-of-the-art packages for large-scale scientific computation. Among its best-known packages are those for the solution of sparse linear systems (ma27, ma57, etc.), compatible with ̀Ipopt. HSL packages are available at no cost for academic research and teaching. Once you obtain the HSL dynamic library (precompiled libhsl.so for Linux, to be compiled libhsl.dylib for MacOSX, libhsl.dll for Windows), you just have to place it in your Anaconda3 environment into the lib/ folder. You can now use all the options of bioptim, including the HSL linear solvers with Ipopt. We recommend using ma57 as a default linear solver by calling as such:

solver = Solver.IPOPT()
solver.set_linear_solver("ma57")
ocp.solve(solver)

Installation complete

Once bioptim is downloaded, navigate to the root folder and (assuming your conda environment is loaded if needed), you can type the following command:

python setup.py install

Assuming everything went well, that is it! You can already enjoy bioptimizing!

Defining our optimal control problems

Here we will detail our implementation of optimal control problems and some definitions. The mathematical transcription of the OCP is as follows: The optimization variables are the states (x = variables that represent the state of the system at each node and that are subject to continuity constraints), controls (u = decision variables defined at each node that drive the system), algebraic states (s = optimization variables that are defined at each node but that are not subject to the built-in continuity constraints), and parameters (p = optimization variables defined once per phase). The state continuity constraints implementation may vary depending on the transcription of the problem (implicit vs explicit, direct multiple shooting vs direct collocations).

The cost function can include Mayer terms (function evaluated at one node, the default is the last node) and Lagrange terms (functions integrated over the duration of the phase). The Lagrange terms are computed by default as EulerForward Integrals:

L = 0
for i in range(n_shooting):
  L += weight * sum((evaluated_cost[:, i] - target_cost[:, i])**2 * dt)

Where weight is by default 1 and target_cost is by default 0. For more advanced approximations, see QuadratureRule section. They can be used to evaluate more accurately the Lagrange terms of the cost function. The optimization variables can be subject to equality and/or inequality constraints.

A first practical example

The easiest way to learn bioptim is to dive into it. So let us do that and build our first optimal control program together. Please note that this tutorial is designed to recreate the examples/getting_started/pendulum.py file where a pendulum is asked to start in a downward position and end, balanced, in an upward position while only being able to move sideways actively.

The import

We will not spend time explaining the import since every one of them will be explained in detail later, and it is pretty straightforward anyway.

from bioptim import (
  BiorbdModel,
  OptimalControlProgram,
  DynamicsFcn,
  Dynamics,
  BoundsList,
  InitialGuessList,
  ObjectiveFcn,
  Objective,
)

Building the ocp

First of all, let us load a bioMod file using biorbd:

bio_model = BiorbdModel("pendulum.bioMod")

It is convenient since it will provide interesting functions such as the number of degrees of freedom (bio_model.nb_q). Please note that a pendulum.bioMod copy is available at the end of the Getting started section. In brief, the pendulum consists of two degrees of freedom (sideways movement and rotation), with the center of mass near the head.

The dynamics of the pendulum, as for many biomechanical dynamics, is driven by the generalized forces. Generalized forces are forces or moments directly applied to the degrees of freedom as if virtual motors were driven them. In bioptim, this dynamic is called torque driven. In a torque driven dynamics, the states are the positions (also called generalized coordinates, q) and the velocities (also called the generalized velocities, qdot), whereas the controls are the joint torques (also called generalized forces, tau). Let us define such dynamics:

dynamics = Dynamics(DynamicsFcn.TORQUE_DRIVEN)

The pendulum is required to start in a downward position (0 rad) and to finish in an upward position (3.14 rad) with no velocity at the start and end nodes. To define that, it would be nice first to define boundary constraints on the position (q) and velocities (qdot) that match those in the bioMod file and to apply them at the very beginning, the very end, and all the intermediate nodes as well. In this case, the state with index 0 is translation y, and index 1 refers to rotation about x. Finally, the index 2 and 3 are the velocity of translation y and rotation about x,respectively.

bounds_from_ranges uses the ranges from a biorbd model and returns a structure with the minimal and maximal bounds for all the degrees of freedom and velocities on three columns corresponding to the starting, intermediate, and final nodes, respectively. How convenient!

x_bounds = BoundsList()
x_bounds["q"] = bio_model.bounds_from_ranges("q")
x_bounds["qdot"] = bio_model.bounds_from_ranges("qdot")

The first dimension of x_bounds is the degrees of freedom (q) and their velocities (qdot) that match those in the bioMod file. The time is discretized in nodes which is the second dimension declared in x_bounds. If you have more than one phase, we would have x_bound[phase][q and qdot, nodes] In the first place, we want the first and last column (which is equivalent to nodes 0 and -1) to be 0, i.e., the translations and rotations to be null for both the position and so the velocities.

x_bounds["q"][:, [0, -1]] = 0
x_bounds["qdot"][:, [0, -1]] = 0

Finally, override once again the final node for the rotation so it is upside down.

x_bounds["q"][1, -1] = 3.14

At that point, you may want to have a look at the x_bounds["q"].min and x_bounds["q"].max matrices to convince yourself that the initial and final positions are prescribed and that all the intermediate points are free up to certain minimal and maximal values.

Up to that point, nothing prevents the solver from simply using the virtual motor of the rotation to rotate the pendulum upward (like clock hands) to get to the upside-down rotation. What makes this example interesting is that we can prevent this by defining minimal and maximal bounds on the control (the maximal forces that these motors have)

u_bounds = BoundsList()
u_bounds["tau"] = [-100, 0], [100, 0]

Like this, the sideways force ranges from -100 Newton to 100 Newton, but the rotation force ranges from 0 N/m to 0 N/m. Again, u_bounds is defined for the first, the intermediate, and the final nodes, but this time, we do not want to specify anything particular for the first and final nodes, so we can leave them as is.

If you wondering where are defined q, qdot and tau, it is in the configuration of DynamicsFcn.TORQUE_DRIVEN. If you define a custom dynamics, then the variable's name should match those you define yourself.

Who says optimization says cost function. Even though, it is possible to define an OCP without objective, it is not so much recommended, and let us face it... much less fun! So the pendulum's goal (or the cost function) is to perform its task while using the minimum forces possible. Therefore, an objective function that minimizes the generalized forces is defined:

objective_functions = Objective(ObjectiveFcn.Lagrange.MINIMIZE_TORQUE)

At that point, it is possible to solve the program. Still, helping the solver is usually a good idea, so let us give ̀Ipopt a starting point to investigate. The initial guess that we can provide is those for the states (x_init, here q and qdot) and for the controls (u_init, here tau). So let us define both of them quickly

x_init = InitialGuessList()
x_init["q"] = [0, 0]
x_init["qdot"] = [0, 0]

u_init = InitialGuessList()
u_init["tau"] = [0, 0]

Please note that initial guess is optional. The default value if a value is not provided is zero.

On the same train of thought, if we want to help the solver even more, we can also define a variable scaling for the states (x_scaling, here q and qdot) and for the controls (u_scaling, here tau). *Note that the scaling should be declared in the order in which the variables appear. We encourage you to choose a variable scaling the same order of magnitude to the expected optimal values.

x_scaling = VariableScalingList()
x_scaling.add("q", scaling=[1, 3])
x_scaling.add("qdot", scaling=[85, 85])
   
u_scaling = VariableScalingList()
u_scaling.add("tau", scaling=[900, 1])

We now have everything to create the ocp! For that, we have to decide how much time the pendulum has to get up there (phase_time) and how many shooting points are defined for the multishoot (n_shooting). Thereafter, you have to send everything to the OptimalControlProgram class and let bioptim prepare everything for you. For simplicity's sake, I copied all the pieces of code previously visited in the building of the ocp section here:

ocp = OptimalControlProgram(
        bio_model,
        dynamics,
        n_shooting=25,
        phase_time=3,
        x_bounds=x_bounds,
        u_bounds=u_bounds,
        x_init=x_init,
        u_init=u_init,
        objective_functions=objective_functions,
    )

Checking the ocp

Now you can check if the ocp is well-defined for the initial values. This checking will help see if your constraints and objectives are okay. To visualize it, you can use

ocp.check_conditioning()

This call will print two different plots!

The first shows the Jacobian matrix of constraints and the norm of each Hessian matrix of constraints. There is one matrix for each phase. The first half of the plot can be used to verify if some constraints are redundant. It simply compares the rank of the Jacobian with the number of constraints for each phase. The second half of the plot can be used to verify if the equality constraints are linear.

The second plot window shows the hessian of the objective for each phase. It calculates if the problem can be convex by checking if the matrix is positive semi-definite. It also calculates the condition number for each phase thanks to the eigenvalues.

If everything is okay, let us solve the ocp !

Solving the ocp

It is now time to see Ipopt in action! To solve the ocp, you simply have to call the solve() method of the ocp class

solver = Solver.IPOPT(show_online_optim=True)
sol = ocp.solve(solver)

If you feel fancy, you can even activate the online optimization graphs! However, for such an easy problem, Ipopt will not leave you the time to appreciate the real-time updates of the graph... For a more complicated problem, you may also wish to visualize the objectives and constraints during the optimization (useful when debugging, because who codes the right thing the first time). You can do it by calling

ocp.add_plot_penalty(CostType.OBJECTIVES)
ocp.add_plot_penalty(CostType.CONSTRAINTS)

or alternatively asks for both at once using

ocp.add_plot_penalty(CostType.ALL)

That's it!

Show the results

If you want to look at the animated data, bioptim has an interface to bioviz designed to visualize bioMod files. For that, simply call the animate() method of the solution:

sol.animate()

If you did not fancy the online graphs but would enjoy them anyway, you can call the method graphs():

sol.graphs()

If you are interested in the results of individual objective functions and constraints, you can print them using the print_cost() or access them using the detailed_cost_values():

# sol.detailed_cost  # Invoke this for adding the details of the objectives to sol for later manipulations
sol.print_cost()  # For printing their values in the console

And that is all! You have completed your first optimal control program with bioptim!

Solving using multi-start

Due to the gradient descent methods, we can affirm that the optimal solution is a local minimum. However, it is impossible to know if a global minimum was found. For highly non-linear problems, there might exist a wide range of local optima. Solving the same problem with different initial guesses can be helpful to find the best local minimum or to compare the different optimal kinematics. It is possible to multi-start the problem by creating a multi-start object with MultiStart() and running it with its method run(). An example of how to use multi-start is given in examples/getting_started/multi-start.py.

Solving stochastic optimal control problems (SOCP)

It is possible to solve SOCP (also called optimal feedback control problem) using the class StochasticOptimalControlProgram. You just have to add the type of SOCP that you want to solve using SocpType.TRAPEZOIDAL_EXPLICIT(motor_noise_magnitude, sensory_noise_magnitude), SocpType.TRAPEZOIDAL_IMPLICIT(motor_noise_magnitude, sensory_noise_magnitude), or SocpType.COLLOCATION(motor_noise_magnitude, sensory_noise_magnitude). Our implementation of SOCP is based on Van Wouwe 2022 (https://doi.org/10.1371/journal.pcbi.1009338). In the examples folder examples/stochastic_optimal_control, you will find arm_reaching_muscle_driven.py which is our implementation of the arm reaching task (6 muscles) described in the above-mentioned article. Our implementation of the integration of the covariance matrix with a collocation scheme is based on Gillis 2013 (https://ieeexplore.ieee.org/abstract/document/6761121). You will also find our implementation of the example of Gillis 2013 in the same folder (obstacle_avoidance_collocations.py).

We recommend the user to use the SocpType.COLLOCATION implementation if a great level of dynamics consistency is needed, or SocpType.TRAPEZOIDAL_IMPLICIT with a Cholesky decomposition of the covariance matrix for a faster resolution.

The complete example files

If you did not completely follow (or were too lazy to!) you will find the complete files described in the Getting started section here. You will find that the file is a bit different from the example/getting_started/pendulum.py, but it is merely different on the surface.

The pendulum.py file

import biorbd_casadi as biorbd
from bioptim import (
    BiorbdModel,
    OptimalControlProgram,
    DynamicsFcn,
    Dynamics,
    BoundsList,    
    InitialGuessList,
    ObjectiveFcn,
    Objective,
)

bio_model = BiorbdModel("pendulum.bioMod")
dynamics = Dynamics(DynamicsFcn.TORQUE_DRIVEN)

# Bounds are optional (default -inf -> inf)
x_bounds = BoundsList()
x_bounds["q"] = bio_model.bounds_from_ranges("q")
x_bounds["q"][:, [0, -1]] = 0
x_bounds["q"][1, -1] = 3.14
x_bounds["dot"] = bio_model.bounds_from_ranges("qdot")
x_bounds["qdot"][:, [0, -1]] = 0

u_bounds = BoundsList()
u_bounds["tau"] = [-100, 0], [100, 0]

objective_functions = Objective(ObjectiveFcn.Lagrange.MINIMIZE_TORQUE)

# Initial guess is optional (default = 0)
x_init = InitialGuessList()
x_init["q"] = [0, 0]
x_init["qdot"] = [0, 0]
u_init = InitialGuessList()
u_init = [0, 0]

ocp = OptimalControlProgram(
        bio_model,
        dynamics,
        n_shooting=25,
        phase_time=3,
        x_bounds=x_bounds,
        u_bounds=u_bounds,
        x_init=x_init,
        u_init=u_init,
        objective_functions=objective_functions,
    )
    
sol = ocp.solve(show_online_optim=True)
sol.print_cost()
sol.animate()

The pendulum.bioMod file

Here is a simple pendulum that can be interpreted by biorbd. For more information on how to build a bioMod file, one can read the doc of biorbd.

version 4

// Seg1
segment Seg1
    translations	y
    rotations	x
    ranges  -1 5
            -2*pi 2*pi
    mass 1
    inertia
        1 0 0
        0 1 0
        0 0 0.1
    com 0.1 0.1 -1
    mesh 0.0   0.0   0.0
    mesh 0.0  -0.0  -0.9
    mesh 0.0   0.0   0.0
    mesh 0.0   0.2  -0.9
    mesh 0.0   0.0   0.0
    mesh 0.2   0.2  -0.9
    mesh 0.0   0.0   0.0
    mesh 0.2   0.0  -0.9
    mesh 0.0   0.0   0.0
    mesh 0.0  -0.0  -1.1
    mesh 0.0   0.2  -1.1
    mesh 0.0   0.2  -0.9
    mesh 0.0  -0.0  -0.9
    mesh 0.0  -0.0  -1.1
    mesh 0.2  -0.0  -1.1
    mesh 0.2   0.2  -1.1
    mesh 0.0   0.2  -1.1
    mesh 0.2   0.2  -1.1
    mesh 0.2   0.2  -0.9
    mesh 0.0   0.2  -0.9
    mesh 0.2   0.2  -0.9
    mesh 0.2  -0.0  -0.9
    mesh 0.0  -0.0  -0.9
    mesh 0.2  -0.0  -0.9
    mesh 0.2  -0.0  -1.1
endsegment

    // Marker 1
    marker marker_1
        parent Seg1
        position 0 0 0
    endmarker

    // Marker 2
    marker marker_2
        parent Seg1
        position 0.1 0.1 -1
    endmarker

A more in-depth look at the bioptim API

In this section, we will have an in-depth look at all the classes one can use to interact with the bioptim API. All the classes covered here can be imported using the command:

from bioptim import ClassName

The OCP

An optimal control program is an optimization that uses control variables to drive some state variables. Bioptim includes two types of transcription methods: the direct collocation and the direct multiple shooting. To summarize, it defines a large optimization problem by discretizing the control and the state variables into a predetermined number of intervals, the beginning of the interval being the shooting points. By defining strict continuity/collocation constraints, it can ensure proper dynamics of the system (i.e. state continuity). The OCP are the solved using gradient descending algorithms until a local minimum is found.

Class: OptimalControlProgram

This is the main class that holds an ocp. Most of the attributes and methods are for internal use; therefore the API user should not care much about them. Once an OptimalControlProgram is constructed, it is usually ready to be solved.

The full signature of the OptimalControlProgram can be scary at first, but should become clear soon. Here it is:

OptimalControlProgram(
    bio_model: [list, BioModel],
    dynamics: [Dynamics, DynamicsList],
    n_shooting: [int, list],
    phase_time: [float, list], 
    x_bounds: BoundsList,
    u_bounds: BoundsList,
    x_init: InitialGuessList
    u_init: InitialGuessList,
    objective_functions: [Objective, ObjectiveList],
    constraints: [Constraint, ConstraintList],
    parameters: ParameterList,
    ode_solver: OdeSolver,
    control_type: [ControlType, list],
    all_generalized_mapping: BiMapping,
    q_mapping: BiMapping,
    qdot_mapping: BiMapping,
    tau_mapping: BiMapping,
    plot_mappings: Mapping,
    phase_transitions: PhaseTransitionList,
    n_threads: int,
    use_sx: bool,
)

Of these, only the first four are mandatory.
bio_model is the model loaded with classes such as BiorbdModel, MultiBiorbdModel, or a custom class. In the case of a multiphase optimization, one model per phase should be passed in a list.
dynamics is the system's dynamics during each phase (see The dynamics section).
n_shooting is the number of shooting points of the direct multiple shooting (method) for each phase.
phase_time is the final time of each phase. If the time is free, this is the initial guess.
x_bounds is the minimal and maximal value the states can have (see The bounds section) .
u_bounds is the minimal and maximal value the controls can have (see The bounds section).
x_init is the initial guess for the states variables (see The initial conditions section).
u_init is the initial guess for the controls variables (see The initial conditions section).
x_scaling is the scaling applied to the states variables (see The variable scaling section).
xdot_scaling is the scaling applied to the state derivative variables (see The variable scaling section).
u_scaling is the scaling applied to the controls variables (see The variable scaling section).
objective_functions is the objective function set of the ocp (see The objective functions section).
constraints is the constraint set of the ocp (see The constraints section).
parameters is the parameter set of the ocp (see The parameters section). It is a list (one element for each phase) of np.ndarray of shape (6, i, n), where the 6 components are [Mx, My, Mz, Fx, Fy, Fz], for the ith force platform (defined by the externalforceindex) for each node n.
ode_solver is the ode solver used to solve the dynamic equations.
control_type is the type of discretization of the controls (usually CONSTANT) (see ControlType section).
all_generalized_mapping is used to reduce the number of degrees of freedom by linking them (see The mappings section). This one applies the same mapping to the generalized coordinates (q), velocities (qdot), and forces (tau). q_mapping the mapping applied to q.
qdot_mapping the mapping applied to q_dot.
tau_mapping the mapping applied to tau.
plot_mappings is to force some plots to be linked together.
n_threads is to solve the optimization using multiple threads. This number is the number of threads to use.
use_sx is if the CasADi graph should be constructed in SX. SX will tend to solve much faster than MX graphs, however they necessitate a huge amount of RAM.

Please note that a common ocp will usually define only these parameters:

ocp = OptimalControlProgram(
    bio_model: [list, BioModel],
    dynamics: [Dynamics, DynamicsList],
    n_shooting: [int, list],
    phase_time: [float, list],
    x_init: InitialGuessList
    u_init: InitialGuessList, 
    x_bounds: BoundsList,
    u_bounds: BoundsList,
    objective_functions: [Objective, ObjectiveList],
    constraints: [Constraint, ConstraintList],
    n_threads: int,
)

The main methods one will be interested in are:

ocp.update_objectives()
ocp.update_constraints()
ocp.update_parameters()
ocp.update_bounds()
ocp.update_initial_guess()

These allow to modify the ocp after being defined. It is advantageous when solving the ocp for the first time, then adjusting some parameters and reoptimizing afterward.

Moreover, the method

solution = ocp.solve(Solver)

is called to solve the ocp (the solution structure is discussed later). The Solver class can be used to select the nonlinear solver to solve the ocp:

  • IPOPT
  • ACADOS
  • SQP method

Note that options can be passed to the solver parameter. One can refer to their respective solver's documentation to know which options exist. The show_online_optim parameter can be set to True so the graphs nicely update during the optimization with the default values. One can also directly declare online_optim as an OnlineOptim parameter to customize the behavior of the plotter. Note that show_online_optim and online_optim are mutually exclusive. Please also note that OnlineOptim.MULTIPROCESS is not available on Windows and only none of them are available on Macos. To see how to run the server on Windows, please refer to the getting_started/pendulum.py example. It is expected to slow down the optimization a bit. show_options can be also passed as a dict to the plotter to customize the plotter's behavior. If online_optim is set to SERVER, then a server must be started manually by instantiating an PlottingServer class (see ressources/plotting_server.py). The following keys are additional options when using OnlineOptim.SERVER and OnlineOptim.MULTIPROCESS_SERVER:

  • host: the host to use (default is localhost)
  • port: the port to use (default is 5030)

Finally, one can save and load previously optimized values by using

ocp.save(solution, file_path)
ocp, solution = OptimalControlProgram.load(file_path)

IMPORTANT NOTICE: Please note that saved solution depends on the bioptim version used to create the .bo file, and retro-compatibility is NOT enforced. In other words, an optimized solution from a previous version will probably NOT load on a newer bioptim version. To save the solution in a way independent of the version of bioptim, one may use the stand_alone flag to True.

Finally, the add_plot(name, update_function) method can create new dynamics plots. The name is simply the name of the figure. If one with the same name already exists, the axes are merged. The update_function is a function handler with signature: update_function(states: np.ndarray, constrols: np.ndarray: parameters: np.ndarray) -> np.ndarray. It is expected to return a np.ndarray((n, 1)), where n is the number of elements to plot. The axes_idx parameter can be added to parse the data in a more exotic manner. For instance, on a three-axes figure, if one wanted to plot the first value on the third axes and the second value on the first axes and nothing on the second, the axes_idx=[2, 0] would do the trick. The interested user can have a look at the examples/getting_started/custom_plotting.py example.

Class: NonLinearProgram

The NonLinearProgram is, by essence, the phase of an ocp. The user is expected not to change anything from this class but can retrieve valuable information from it.

One main use of nlp is to get a reference to the bio_model for the current phase: nlp.model. Another essential value stored in nlp is the shape of the states and controls: nlp.shape, which is a dictionary where the keys are the names of the elements (for instance, q for the generalized coordinates)

It would be tedious, and probably not much useful, to list all the elements of nlp here.
The interested user is invited to look at the docstrings for this class to get a detailed overview of it.

Class: VariationalOptimalControlProgram

The VariationalOptimalControlProgram class inherits from OptimalControlProgram and is used to solve optimal control problems using the variational approach. A variational integrator does the integration. The formulation being completely different from the other approaches, it needed its own class. The parameters are the same as in OptimalControlProgram apart from the following changes:

  • bio_model must be a VariationalBiorbdModel
  • The phases have not been implemented yet; hence, only final_time must be specified, and it must be a float.
  • There are no velocities in the variational approach, so you must only specify the q_init and not the q_bounds instead of x_init and x_bounds.
  • You can specify an initial guess for the velocities at the first node and the last node using qdot_init and qdot_bounds and the keys must be "qdot_start" and "qdot_end". These velocities are implemented as parameters of the OCP, you can access them with sol.parameters["qdot_start"] and sol.parameters["qdot_end"] at the end of the optimization.

Class: PlottingServer

If one wants to use the OnlineOptim.SERVER plotter, one can instantiate this class to start a server. This is not mandatory as if as_multiprocess is set to True in the show_options dict [default behavior], this server is started automatically. The advantage of starting the server manually is that one can plot online graphs on a remote machine. An example of such a server is provided in resources/plotting_server.py.

The model

Bioptim is designed to work with any model, as long as it inherits from the class bioptim.Model. Models built with biorbd are already compatible with bioptim. They can be used as is or modified to add new features.

Class: BiorbdModel

The BiorbdModel class implements a BioModel of the biorbd dynamics library. Some methods may not be interfaced yet; it is accessible through:

bio_model = BiorbdModel("path/to/model.bioMod")
bio_model.marker_names  # for example returns the marker names
# if the methods is not interfaced, it can be accessed through
bio_model.model.markerNames()

Class: MultiBiorbdModel

The MultiBiorbdModel class implements BioModel of multiple models of biorbd dynamics library. Some methods may not be interfaced yet; it is accessible through:

bio_model = MultiBiorbdModel(("path/to/model.bioMod", "path/to/other/model.bioMod"))

Class: HolonomicBiorbdModel

The HolonomicBiorbdModel class implements a BioModel of the biorbd dynamics library. Since the class inherits from BiorbdModel, all the methods of BiorbdModel are available. You can define the degrees of freedom (DoF) that are independent (that define the movement) and the ones that are dependent (that are defined by the independent DoF and the holonomic constraint(s)). You can add some holonomic constraints to the model. For this, you can use one of the functions of HolonomicConstraintFcn or add a custom one. You can refer to the examples in bioptim/examples/holonomic_constraints to see how to use it. Some methods may not be interfaced yet; it is accessible through:

bio_model = HolonomicBiorbdModel("path/to/model.bioMod")
holonomic_constraints = HolonomicConstraintsList()
holonomic_constraints.add("holonomic_constraints", HolonomicConstraintsFcn.function, **kwargs)
bio_model.set_holonomic_configuration(holonomic_constraints, independent_joint_index, dependent_joint_index)

Two dynamics are implemented in the differential algebraic equations handling constraints at the acceleration level in constrained_forward_dynamics(...). Moreover, the other was inspired by Robotran, which uses index reduction methods to satisfy the constraints: partitioned_forward_dynamics(...)

Class VariationalBiorbdModel

The VariationalBiorbdModel class implements a BioModel of the biorbd dynamics library. It is used in Discrete Mechanic and Optimal Control (DMOC) and Discrete Mechanics and Optimal Control in Constrained Systems (DMOCC). Since the class inherits from HolonomicBiorbdModel, all the HolonomicBiorbdModel and BiorbdModel methods are available. This class is used in VariationalOptimalControlProgram. You can refer to the examples in bioptim/examples/discrete_mechanics_and_optimal_control to see how to use it. Some methods may not be interfaced yet; it is accessible through:

bio_model = VariationalBiorbdModel("path/to/model.bioMod")
holonomic_constraints = HolonomicConstraintsList()
holonomic_constraints.add("holonomic_constraints", HolonomicConstraintsFcn.function, **kwargs)
bio_model.set_holonomic_configuration(holonomic_constraints)
VariationalOptimalControlProgram(bio_model, ...)

Class: CustomModel

The BioModel class is the base class for BiorbdModel and any custom models. The methods are abstracted and must be implemented in the child class, or at least raise a NotImplementedError if they are not implemented. For example:

from bioptim import Model

class MyModel(CustomModel, metaclass=ABCMeta):
    def __init__(self, *args, **kwargs):
        ...

    def name_dof(self):
        return ["dof1", "dof2", "dof3"]

    def marker_names(self):
        raise NotImplementedError

see the example examples/custom_model/ for more details.

The dynamics

By essence, an optimal control program (ocp) links two types of variables: the states (x) and the controls (u). Conceptually, the controls are the driving inputs of the system, which participate in changing the system states. In the case of biomechanics, the states (x) are usually the generalized coordinates (q) and velocities (qdot), i.e., the pose of the musculoskeletal model and the joint velocities. On the other hand, the controls (u) can be the generalized forces, i.e., the joint torques, but can also be the muscle excitations, for instance. States and controls are linked through Ordinary differential equations: dx/dt = f(x, u, p), where p can be additional parameters that act on the system but are not time-dependent.

The following section investigates how to instruct bioptim of the dynamic equations the system should follow.

Class: Dynamics

This class is the main class to define a dynamics. It, therefore, contains all the information necessary to configure (i.e., determining which variables are states or controls) and perform the dynamics. When constructing an OptimalControlProgram(), Dynamics is the expected class for the dynamics parameter.

The user can minimally define a Dynamics as: dyn = Dynamics(DynamicsFcn). The DynamicsFcn is the one presented in the corresponding section below.

The options

The full signature of Dynamics is as follows:

Dynamics(dynamics_type, configure: Callable, dynamic_function: Callable, phase: int)

The dynamics_type is the selected DynamicsFcn. It automatically defines both configure and dynamic_function. If a function is sent instead, this function is interpreted as configure and the DynamicsFcn is assumed to be DynamicsFcn.CUSTOM If one is interested in changing the behavior of a particular DynamicsFcn, they can refer to the Custom dynamics functions right below.

The phase is the index of the phase the dynamics applies to. The add() method of DynamicsList usually takes care of this, but it can be useful when declaring the dynamics out of order.

Custom dynamic functions

If an advanced user wants to define their own dynamic function, they can define the configuration and/or the dynamics.

The configuration is what tells bioptim which variables are states and which are control. The user is expected to provide a function handler with the following signature: custom_configure(ocp: OptimalControlProgram, nlp: NonLinearProgram). In this function, the user is expected to call the relevant ConfigureProblem class methods:

  • configure_q(nlp, as_states: bool, as_controls: bool)
  • configure_qdot(nlp, as_states: bool, as_controls: bool)
  • configure_q_qdot(nlp, as_states: bool, as_controls: bool)
  • configure_tau(nlp, as_states: bool, as_controls: bool)
  • configure_residual_tau(nlp, as_states: bool, as_controls: bool)
  • configure_muscles(nlp, as_states: bool, as_controls: bool) where as_states add the variable to the states vector and as_controls to the controls vector. Please note that this is not necessarily mutually exclusive. Finally, the user is expected to configure the dynamic by calling ConfigureProblem.configure_dynamics_function(ocp, nlp, custom_dynamics)

Defining the dynamic function must be done when one provides a custom configuration, but it can also be defined by providing a function handler to the dynamic_function parameter for Dynamics. The signature of this custom dynamic function is as follows: custom_dynamic(states: MX, controls: MX, parameters: MX, nlp: NonLinearProgram. This function is expected to return a tuple[MX] of the derivative of the states. Some methods defined in the class DynamicsFunctions can be useful, but will not be covered here since it is initially designed for internal use. Please note that MX type is a CasADi type. Anyone who wants to define custom dynamics should be at least familiar with this type beforehand.

Class: DynamicsList

A DynamicsList is simply a list of Dynamics. The add() method can be called exactly as if one was calling the Dynamics constructor. If the add() method is used more than one, the phase parameter is automatically incremented.

So a minimal use is as follows:

dyn_list = DynamicsList()
dyn_list.add(DynamicsFcn)

Class: DynamicsFcn

The DynamicsFcn class is the configuration and declaration of all the already available dynamics in bioptim. Since this is an Enum, it is possible to use tab key on the keyboard to dynamically list them all, depending on the capabilities of your IDE.

Please note that one can change the dynamic function associated to any of the configuration by providing a custom dynamics_function. For more information on this, please refer to the Dynamics and DynamicsList section right before.

TORQUE_DRIVEN

The torque driven defines the states (x) as q and qdot and the controls (u) as tau. The derivative of q is trivially qdot. The derivative of qdot is given by the biorbd function: qddot = bio_model.ForwardDynamics(q, qdot, tau). If external forces are provided, they are added to the ForwardDynamics function. Possible options:

  • with_contact = True: The derivative of qdot is given by the biorbd function that includes non-acceleration contact point defined in the bioMod: qddot = bio_model.ForwardDynamicsConstraintsDirect(q, qdot, tau).
  • with_passive_torque = True: The passive torque is taken into account in the tau.
  • with_ligament = True: The tau generated by the ligament is taken into account in the tau.

TORQUE_DERIVATIVE_DRIVEN

The torque derivative driven defines the states (x) as q, qdot, tau and the controls (u) as taudot. The derivative of q is trivially qdot. The derivative of qdot is given by the biorbd function: qddot = bio_model.ForwardDynamics(q, qdot, tau). The derivative of tau is trivially taudot. If external forces are provided, they are added to the ForwardDynamics function. Possible options:

  • with_contact = True: The derivative of qdot is given by the biorbd function that includes non-acceleration contact point defined in the bioMod: qddot = bio_model.ForwardDynamicsConstraintsDirect(q, qdot, tau).
  • with_passive_torque = True: The passive torque is taken into account in the tau.
  • with_ligament = True: The tau generated by the ligament is taken into account in the tau.

TORQUE_ACTIVATIONS_DRIVEN

The torque driven defines the states (x) as q and qdot and the controls (u) as the level of activation of tau. The derivative of q is trivially qdot. The actual tau is computed from the activation by the biorbd function: tau = bio_model.torque(torque_act, q, qdot). Then, the derivative of qdot is given by the biorbd function: qddot = bio_model.ForwardDynamics(q, qdot, tau).

Please note, this dynamics is expected to be very slow to converge, if it ever does. One is therefore encourage using TORQUE_DRIVEN instead, and to add the TORQUE_MAX_FROM_ACTUATORS constraint. This has been shown to be more efficient and allows defining minimum torque. Possible options:

  • with_contact = True: The derivative of qdot is given by the biorbd function that includes non-acceleration contact point defined in the bioMod: qddot = bio_model.ForwardDynamicsConstraintsDirect(q, qdot, tau).
  • with_passive_torque = True: The passive torque is taken into account in the tau.
  • with_ligament = True: The tau generated by the ligament is taken into account in the tau.
  • with_residual_torque = True: The residual torque is taken into account in the tau.

JOINTS_ACCELERATION_DRIVEN

The joints acceleration driven defines the states (x) as q and qdot and the controls (u) as qddot_joints. The derivative of q is trivially qdot. The joints' acceleration qddot_joints is the acceleration of the actual joints of the biorb_model without its root's joints. The model's root's joints acceleration qddot_root are computed by the biorbd function: qddot_root = boirbd_model.ForwardDynamicsFreeFloatingBase(q, qdot, qddot_joints). The derivative of qdot is the vertical stack of qddot_root and qddot_joints.

This dynamic is suitable for bodies in free fall.

MUSCLE_DRIVEN

The torque driven defines the states (x) as q and qdot and the controls (u) as the muscle activations. The derivative of q is trivially qdot. Possible options: The actual tau is computed from the muscle activation converted in muscle forces and thereafter converted to tau by the biorbd function: bio_model.muscularJointTorque(muscles_states, q, qdot). The derivative of qdot is given by the biorbd function: qddot = bio_model.ForwardDynamics(q, qdot, tau).

  • with_contact = True: The actual tau is computed from the sum of tau to the a converted in muscle forces and thereafter converted to tau by the biorbd function: bio_model.muscularJointTorque(a, q, qdot). The derivative of qdot is given by the biorbd function that includes non-acceleration contact point defined in the bioMod: qddot = bio_model.ForwardDynamics(q, qdot, tau).
  • with_passive_torque = True: The passive torque is taken into account in the tau.
  • with_ligament = True: The tau generated by the ligament is taken into account in the tau.
  • with_residual_torque = True: The torque driven defines the states (x) as q and qdot and the controls (u) as the tau and the muscle activations (a). The actual tau is computed from the sum of tau to the muscle activation converted in muscle forces and thereafter converted to tau by the biorbd function: bio_model.muscularJointTorque(a, q, qdot).
  • with_excitations = True: The torque driven defines the states (x) as q, qdot and muscle activations (a) and the controls (u) as the tau and the EMG. The derivative of a is computed by the biorbd function: adot = model.activationDot(emg, a)

HOLOMOMIC_TORQUE_DRIVEN

This dynamics have been implemented to be used with HolonomicBiorbdModel. It is a torque driven only applied on the independent degrees of freedom.

CUSTOM

This leaves the user to define both the configuration (what are the states and controls) and to define the dynamic function. CUSTOM should not be called by the user, but the user should pass the configure_function directly. You can have a look at Dynamics and DynamicsList sections for more information about how to configure and define custom dynamics.

The bounds

The bounds provide a class that has minimal and maximal values for a variable. It is, for instance, useful for the inequality constraints that limit the maximal and minimal values of the states (x) and the controls (u) . In that sense, it is what is expected by the OptimalControlProgram for its u_bounds and x_bounds parameters. It can however be used for much more. If not provided for one variable, then it is -infinity to +infinity for that particular variable.

Class: BoundsList

The BoundsList class is the main class to define bounds. The constructor can be called by sending two boundary matrices (min, max) as such: bounds["name"] = min_bounds, max_bounds. Or by providing a previously declared bounds: bounds.add("name", another_bounds). The add nomenclature can also be used with the min and max, but must be specified as such: bounds.add("name", min_bound=min_bounds, max_bound=max_bounds). The min_bounds and max_bounds matrices must have dimensions that fit the chosen InterpolationType, the default type being InterpolationType.CONSTANT_WITH_FIRST_AND_LAST_DIFFERENT, which is 3 columns.

Please note that to change any option, you must use the .add nomenclature

The full signature of BoundsList.add is as follows:

BoundsList.add("name", bounds, min_bounds, max_bound, interpolation_type, phase)

The first parameters are presented before. The phase is the index of the phase the bounds apply to. If you add twice the same element on the same phase, the first is then overrided.

If the interpolation type is CUSTOM, then the bounds are function handlers of signature:

custom_bound(current_shooting_point: int, n_elements: int, n_shooting: int)

where current_shooting_point is the current point to return, n_elements is the number of expected lines and n_shooting is the number of total shooting point (that is if current_shooting_point == n_shooting, this is the end of the phase)

The main methods the user will be interested in is the min property that returns the minimal bounds and the max property that returns the maximal bounds. Unless it is a custom function, min and max are numpy.ndarray and can be directly modified to change the boundaries. It is also possible to change min and max simultaneously by directly slicing the bounds as if it was a numpy.array, effectively defining an equality constraint: for instance bounds["name"][:, 0] = 0. Please note that if more than one phase is present in the bounds, then you must specify on which phase it should apply like so: bounds[phase_index]["name"]...

The initial conditions

The initial conditions the solver should start from, i.e., initial values of the states (x) and the controls (u). In that sense, it is what is expected by the OptimalControlProgram for its u_init and x_init parameters. If not specified for one variable, then it is set to zero for that particular variable.

Class InitialGuessList

The InitialGuessList class is the main class to define initial guesses. The .add can be called by sending one initial guess matrix (init) as such: init["name"] = init. The init matrix must have the dimensions that fits the chosen InterpolationType, the default type being InterpolationType.CONSTANT, which is 1 column.

The full signature of InitialGuessList.add is as follows:

InitialGuessList.add("name", initial_guess, interpolation_type, phase)

The first parameters are presented before. The phase is the index of the phase the initial guess applies to.

If the interpolation type is CUSTOM, then the InitialGuess is a function handler of signature:

custom_init(current_shooting_point: int, n_elements: int, n_shooting: int)

where current_shooting_point is the current point to return, n_elements is the number of expected lines and n_shooting is the number of total shooting point (that is if current_shooting_point == n_shooting, this is the end of the phase)

The main methods the user will be interested in is the init property that returns the initial guess. Unless it is a custom function, init is a numpy.ndarray and can be directly modified to change the initial guess.

If someone wants to add noise to the initial guess, you can provide the following:

init = init.add_noise(
    bounds: BoundsList, 
    magnitude: list | int | float | np.ndarray,
    magnitude_type: MagnitudeType, n_shooting: int, 
    bound_push: list | int | float, 
    seed: int
    )

The bounds must contain all the keys defined in the init list. The parameters, except MagnitudeType must be specified for each phase unless you want the same value for every phases.

The variable scaling

The scaling applied to the optimization variables, it is what is expected by the OptimalControlProgram for its x_scaling, xdot_scaling and u_init parameters.

Class VariableScalingList

A VariableScalingList is a list of VariableScaling. The add() method can be called exactly as if one was calling the VariableScaling constructor.

So a minimal use is as follows:

scaling = VariableScalingList()
scaling.add("q", scaling=[1, 1])

The constraints

The constraints are hard penalties of the optimization program. That means the solution won't be considered optimal unless all the constraint set is fully respected. The constraints come in two format: equality and inequality.

Class: Constraint

The Constraint provides a class that prepares a constraint, so it can be added to the constraint set by bioptim. When constructing an OptimalControlProgram(), Constraint is the expected class for the constraint parameter. It is also possible to later change the constraint by calling the method update_constraints(the_constraint) of the OptimalControlProgram

The Constraint class is the main class to define constraints. The constructor can be called with the type of the constraint and the node to apply it to, as such: constraint = Constraint(ConstraintFcn, node=Node.END). By default, the constraint will be an equality constraint equals to 0. To change this behaviour, one can add the parameters min_bound and max_bound to change the bounds to their desired values.

The full signature of Constraint is as follows:

Constraint(ConstraintFcn, node: node, index: list, phase: int, list_index: int, target: np.ndarray **extra_param)

The first parameters are presented before. The list is the list of elements to keep. For instance, if one defines a TRACK_STATE constraint with index=0, then only the first state is tracked. The default value is all the elements. The phase is the index of the phase the constraint should apply to. If it is not sent, phase=0 is assumed. The list_index is the ith element of a list for a particular phase This is usually taken care by the add() method of ConstraintList, but it can be useful when declaring the constraints out of order, or when overriding previously declared constraints using update_constraints. The target is a value subtracted to the constraint value. It is useful to define tracking problems. The dimensions of the target must be of [index, node]

The ConstraintFcn class provides a list of some predefined constraint functions. Since this is an Enum, it is possible to use tab key on the keyboard to dynamically list them all, assuming you IDE allows for it. It is possible however to define a custom constraint by sending a function handler in place of the ConstraintFcn. The signature of this custom function is: custom_function(pn: PenaltyController, **extra_params) The PenaltyController contains all the required information to act on the states and controls at all the nodes defined by node, while **extra_params are all the extra parameters sent to the Constraint constructor. The function is expected to return an MX vector of the constraint to be inside min_bound and max_bound. Please note that MX type is a CasADi type. Anyone who wants to define custom constraint should be at least familiar with this type beforehand.

Class: ConstraintList

A ConstraintList is simply a list of Constraints. The add() method can be called exactly as calling the Constraint constructor. If the add() method is used more than once, the list_index parameter is automatically incremented for the prescribed phase. If no phase is prescribed by the user, the first phase is assumed.

So a minimal use is as follows:

constraint_list = ConstraintList()
constraint_list.add(constraint)

Class: ConstraintFcn

The ConstraintFcn class is the declaration of all the already available constraints in bioptim. Since this is an Enum, it is possible to use the tab key on the keyboard to dynamically list them all, depending on the capabilities of your IDE. The existing contraint functions in alphabetical order:

  • BOUND_STATE — Adds bounds on states. Same aim as bounds["state_name"] = min_bounds, max_bounds but with a different numerical behaviour.
  • BOUND_CONTROL — Adds bounds on controls. Same aim as bounds["control_name"] = min_bounds, max_bounds but with a different numerical behaviour.
  • NON_SLIPPING — Adds a constraint of static friction at contact points constraining for small tangential forces.
    This constraint assumes that the normal forces is positive (that is having an additional TRACK_CONTACT_FORCES with max_bound=np.inf). The extra parameters tangential_component_idx: int, normal_component_idx: int, and static_friction_coefficient: float must be passed to the Constraint constructor.
  • PROPORTIONAL_CONTROL — Links one control to another, such that u[first_dof] - first_dof_intercept = coef * (u[second_dof] - second_dof_intercept). The extra parameters first_dof: int and second_dof: int must be passed to the Constraint constructor.
  • PROPORTIONAL_STATE — Links one state to another, such that x[first_dof] - first_dof_intercept = coef * (x[second_dof] - second_dof_intercept). The extra parameters first_dof: int and second_dof: int must be passed to the Constraint constructor.
  • SUPERIMPOSE_MARKERS — Matches one marker with another one. The extra parameters first_marker_idx: int and second_marker_idx: int informs which markers are to be superimposed.
  • TIME_CONSTRAINT — Adds the time to the optimization variable set. It will leave the time free within the given boundaries.
  • TORQUE_MAX_FROM_ACTUATORS — Adds a constraint of maximal torque to the generalized forces controls such that the maximal tau are computed from the biorbd method bio_model.torque_max(q, qdot). This is an efficient alternative to torque activation dynamics. The extra parameter min_torque can be passed to ensure that the model is never too weak.
  • TRACK_ALL_CONTROLS — Tracks all the control variables toward a target.
  • TRACK_ANGULAR_MOMENTUM — Constraints the angular momentum in the global reference frame toward a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the momentum should be tracked.
  • TRACK_COM_POSITION — Constraints the center of mass toward a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the center of mass should be tracked.
  • TRACK_COM_VELOCITY — Constraints the center of mass velocity toward a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the velocity should be tracked.
  • TRACK_CONTACT_FORCES — Tracks the non-acceleration point reaction forces toward a target.
  • TRACK_LINEAR_MOMENTUM — Constraints the linear momentum toward a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the momentum should be tracked.
  • TRACK_MARKER_WITH_SEGMENT_AXIS — Tracks a marker using a segment, that is aligning an axis toward the marker. The extra parameters marker_idx: int, segment_idx: int, and axis: Axis must be passed to the Constraint constructor
  • TRACK_MARKERS_VELOCITY — Tracks the skin marker velocities toward a target.
  • TRACK_MARKERS — Tracks the skin markers toward a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the markers should be tracked.
  • TRACK_MUSCLES_CONTROL — Tracks the muscles (part of the control variables) toward a target.
  • TRACK_SEGMENT_WITH_CUSTOM_RT —Links a segment with an RT (for instance, an Inertial Measurement Unit). It does so by computing the homogenous transformation between the segment and the RT and then converting this to Euler angles. The extra parameters segment_idx: int and rt_idx: int must be passed to the Constraint constructor.
  • TRACK_STATE — Tracks the state's variable toward a target.
  • TRACK_TORQUE — Tracks the generalized forces (part of the control variables) toward a target.
  • CUSTOM — The user should not directly send CUSTOM, but the user should pass the custom_constraint function directly. You can look at Constraint and ConstraintList sections for more information about how to define custom constraints.

The objective functions

The objective functions are soft penalties of the optimization program. In other words, the solution tries to minimize the value as much as possible but will not complain if the objective remains high. The objective functions come in two formats: Lagrange and Mayer.

The Lagrange objective functions are integrated over the whole phase (actually over the selected nodes, usually Node.ALL). One should note that integration is not given by the dynamics function but by the rectangle approximation over a node.

The Mayer objective functions are values at a single node, usually the Node.LAST.

Class: Objective

The Objective provides a class that prepares an objective function so that it can be added to the objective set by bioptim. When constructing an OptimalControlProgram(), Objective is the expected class for the objective_functions parameter. It is also possible to later change the objective functions by calling the method update_objectives(the_objective_function) of the OptimalControlProgram

The Objective class is the main class to define objectives. The constructor can be called with the type of the objective and the node to apply it to, as such: objective = Objective(ObjectiveFcn, node=Node.END). Please note that ObjectiveFcn should either be a ObjectiveFcn.Lagrange or ObjectiveFcn.Mayer.

The full signature of Objective is as follows:

Objective(ObjectiveFcn, node: Node, index: list, phase: int, list_index: int, quadratic: bool, target: np.ndarray, weight: float, **extra_param)

The first parameters are presented before. The list is the list of elements to keep. When defining a MINIMIZE_STATE objective_function with index=0, only the first state is minimized. The default value is all the elements. The phase is the index of the phase the objective function should apply to. If it is not sent, phase=0 is assumed. The list_index is the ith element of a list for a particular phase This is usually taken care by the add() method of ObjectiveList, but it can be useful when declaring the objectives out of order or when overriding previously declared objectives using update_objectives. quadratic defines if the objective function should be squared. This is particularly useful when minimizing toward 0 instead of minus infinity. The target is a value subtracted from the objective value. It is relevant to define tracking problems. The dimensions of the target must be of [index, node]. Finally, weight is the weighting that should be applied to the objective. The higher the weight is, the more important the objective is compared to the other objective functions.

The ObjectiveFcn class provides a list of some predefined objective functions. Since ObjectiveFcn.Lagrange and ObjectiveFcn.Mayer are Enum, it is possible to use tab key on the keyboard to dynamically list them all, assuming you IDE allows for it. It is possible, however, to define a custom objective function by sending a function handler in place of the ObjectiveFcn. In this case, an additional parameter must be sent to the Objective constructor: the custom_type with either ObjectiveFcn.Lagrange or ObjectiveFcn.Mayer. The signature of the custom function is: custom_function(pn: PenaltyController, **extra_params) The PenaltyController contains all the required information to act on the states and controls at all the nodes defined by node, while **extra_params are all the extra parameters sent to the Objective constructor. The function is expected to return an MX vector of the objective function. Please note that MX type is a CasADi type. Anyone who wants to define custom objective functions should be at least familiar with this type beforehand.

Class: ObjectiveList

An ObjectiveList is a list of Objective. The add() method can be called exactly as calling the Objective constructor. If the add() method is used more than once, the list_index parameter is automatically incremented for the prescribed phase. If no phase is prescribed by the user, the first phase is assumed.

So a minimal use is as follows:

objective_list = ObjectiveList()
objective_list.add(objective)

Class: ObjectiveFcn

Here a list of objective function with its type (Lagrange and/or Mayer) in alphabetical order:

  • MINIMIZE_ALL_CONTROLS (Lagrange) — Minimizes all the control variables toward zero (or a target).
  • MINIMIZE_ANGULAR_MOMENTUM (Lagrange and Mayer) — Minimizes the angular momentum in the global reference frame toward zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the momentum should be minimized.
  • MINIMIZE_COM_ACCELERATION (Lagrange and Mayer) — Minimizes the center of mass acceleration towards zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the acceleration should be minimized.
  • MINIMIZE_COM_POSITION (Lagrange and Mayer) — Minimizes the center of mass position toward zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the center of mass should be minimized.
  • MINIMIZE_COM_VELOCITY (Lagrange and Mayer) — Minimizes the center of mass velocity towards zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the velocity should be minimized.
  • MINIMIZE_CONTACT_FORCES (Lagrange) — Minimizes the non-acceleration points of the reaction forces toward zero (or a target).
  • MINIMIZE_LINEAR_MOMENTUM (Lagrange and Mayer) — Minimizes the linear momentum towards zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be provided to specify the axes along which the momentum should be minimized.
  • MINIMIZE_MARKERS_DISPLACEMENT (Lagrange) — Minimizes the difference between a state at a node and the same state at the next node, effectively minimizing the velocity. The extra parameter coordinates_system_idx can be specified to compute the marker position in that coordinate system. Otherwise, it is computed in the global reference frame.
  • MINIMIZE_MARKERS_VELOCITY or MINIMIZE_MARKERS_ACCELERATION (Lagrange and Mayer) — Minimizes the marker velocities or accelerations toward zero (or a target).
  • MINIMIZE_MARKERS (Lagrange and Mayer) — Minimizes the position of the markers toward zero (or a target). The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the markers should be minimized.
  • MINIMIZE_MUSCLES_CONTROL (Lagrange) — Minimizes the muscles' controls (part of the control variables) toward zero (or a target).
  • MINIMIZE_PREDICTED_COM_HEIGHT (Mayer) — Minimizes the maximal height of the center of mass, predicted from the parabolic equation, assuming vertical axis is Z (2): CoM_dot[2]**2 / (2 * -g) + CoM[2]. To maximize a jump, one can use this function at the end of the push-off phase and declare a weight of -1.
  • MINIMIZE_SOFT_CONTACT_FORCES (Lagrange) — Minimizes the external forces induced by soft contacts (or a target).
  • MINIMIZE_STATE_DERIVATIVE (Lagrange) — Minimizes the difference between a state at a node and the same state at the next node, i.e., minimizes the generalized state derivative.
  • MINIMIZE_STATE (Lagrange and Mayer) — Minimizes the state variable towards zero (or a target).
  • MINIMIZE_TIME (Lagrange and Mayer) — Adds the time to the optimization variable set. It will minimize the time toward minus infinity or a target. If the Mayer term is used, min_bound and max_bound can also be defined.
  • MINIMIZE_TORQUE_DERIVATIVE (Lagrange) — Minimizes the difference between a tau at a node and the same tau at the next node, i.e., minimizes the generalized forces derivative.
  • MINIMIZE_TORQUE (Lagrange) — Minimizes the generalized forces (part of the control variables) toward zero (or a target).
  • PROPORTIONAL_CONTROL (Lagrange) — Minimizes the difference between one control and another, such that u[first_dof] - first_dof_intercept = coef * (u[second_dof] - second_dof_intercept). The extra parameters first_dof: int and second_dof: int must be passed to the Objective constructor.
  • PROPORTIONAL_STATE (Lagrange and Mayer) — Minimizes the difference between one state and another, such that x[first_dof] - first_dof_intercept = coef * (x[second_dof] - second_dof_intercept). The extra parameters first_dof: int and second_dof: int must be passed to the Objective constructor.
  • SUPERIMPOSE_MARKERS (Lagrange and Mayer) — Tracks one marker with another one. The extra parameters first_marker_idx: int and second_marker_idx: int informs which markers are to be superimposed
  • TRACK_ALL_CONTROLS (Lagrange) — Tracks all the control variables toward a target.
  • TRACK_CONTACT_FORCES (Lagrange) — Tracks the non-acceleration points of the reaction forces toward a target.
  • TRACK_MARKER_WITH_SEGMENT_AXIS (Lagrange and Mayer) — Minimizes the distance between a marker and an axis of a segment, that is aligning an axis toward the marker. The extra parameters marker_idx: int, segment_idx: int and axis: Axis must be passed to the Objective constructor
  • TRACK_MARKERS_VELOCITY or TRACK_MARKERS_ACCELERATION (Lagrange and Mayer) — Tracks the marker velocities or accelerations toward a target.
  • TRACK_MARKERS (Lagrange and Mayer) — Tracks the skin markers towards a target. The extra parameter axis_to_track: Axis = (Axis.X, Axis.Y, Axis.Z) can be sent to specify the axes along which the markers should be tracked.
  • TRACK_MUSCLES_CONTROL (Lagrange) — Tracks the muscles' controls (part of the control variables) toward a target.
  • TRACK_SEGMENT_WITH_CUSTOM_RT (Lagrange and Mayer) — Minimizes the distance between a segment and an RT (for instance, an Inertial Measurement Unit). It does so by computing the homogenous transformation between the segment and the RT and then converting this to Euler angles. The extra parameters segment_idx: int and rt_idx: int must be passed to the Objective constructor.
  • TRACK_SOFT_CONTACT_FORCES (Lagrange) — Tracks the external forces induced by soft contacts toward a target.
  • TRACK_STATE (Lagrange and Mayer) — Tracks the state variable toward a target.
  • TRACK_TORQUE (Lagrange — Tracks the generalized forces (part of the control variables) toward a target.
  • CUSTOM (Lagrange and Mayer) — The user should not directly send CUSTOM, but pass the custom_objective function directly. You can look at Objective and ObjectiveList sections for more information about defining custom objective function.

The parameters

Parameters are time-independent variables (e.g., a muscle maximal isometric force, the value of gravity ). that affect the dynamics of the whole system. Due to the variety of parameters, it was impossible to provide predefined parameters but the time. Therefore, all the parameters are custom-made.

Class: ParameterList

The ParameterList provides a class that prepares the parameters, so it can be added to the parameter set to optimize by bioptim. When constructing an OptimalControlProgram(), ParameterList is the expected class for the parameters parameter. It is also possible to later change the parameters by calling the method update_parameters(the_parameter_list) of the OptimalControlProgram

The ParameterList class is the main class to define parameters. Please note that, unlike other lists, Parameter is not accessible. This is for simplicity reasons, as it would complicate the API quite a bit to permit it. Therefore, one should not call the Parameter constructor directly.

Here is the full signature of the add() method of the ParameterList:

ParameterList.add(parameter_name: str, function: Callable, initial_guess: InitialGuess, bounds: Bounds, size: int, phase: int, **extra_parameters)

The parameter_name is the parameter's name (reference for the output data as well). The function is the function that modifies the biorbd model, it will be called just prior to applying the dynamics. The signature of the custom function is: custom_function(BioModel, MX, **extra_parameters), where BiorbdModel is the model to apply the parameter to, the MX is the value the parameter will take, and the **extra_parameters are those sent to the add() method. This function is expected to modify the bio_model, and not return anything. Please note that MX type is a CasADi type. Anyone who wants to define custom parameters should be at least familiar with this type beforehand. The initial_guess is the initial value of the parameter. The bounds are the maximal and minimal values of the parameter. The size is the number of elements of this parameter. If an objective function is provided, the return of the objective function should match the size. The phase that the parameter applies to. Even though a parameter is time-independent, one biorbd_model is loaded per phase. Since parameters are associated to a specific bio_model, one must define a parameter per phase.

The multinode constraints

Bioptim can declare multiphase optimisation programs. The goal of a multiphase ocp is usually to handle changing dynamics. The user must understand that each phase is, therefore, a full ocp by itself, with constraints that links the end of which with the beginning of the following.

Class: BinodeConstraintList

The BinodeConstraintList provides a class that prepares the binode constraints. When constructing an OptimalControlProgram(), BinodeConstraintList is the expected class for the binode_constraints parameter.

The BinodeConstraintList class is the main class to define parameters. Please note that, unlike other lists, BinodeConstraint is not accessible since binode constraints do not make sense for single-phase ocp. Therefore, one should not call the PhaseTransition constructor directly.

Here is the full signature of the add() method of the BinodeConstraintList:

BinodeConstraintList.add(BinodeConstraintFcn, phase_first_idx, phase_second_idx, first_node, second_node, **extra_parameters)

The BinodeConstraintFcn is binode constraints function to use. The default is EQUALITY. When declaring a custom transition phase, BinodeConstraintFcn is the function handler to the custom function. The signature of the custom function is: custom_function(binode_constraint:BinodeConstraint, nlp_pre: NonLinearProgram, nlp_post: NonLinearProgram, **extra_parameters), where nlp_pre is the non linear program of the considered phase, nlp_post is the non linear program of the second considered phase, and the **extra_parameters are those sent to the add() method. This function is expected to return the cost of the binode constraint computed in the form of an MX. Please note that MX type is a CasADi type. Anyone who wants to define binode constraints should be at least familiar with this type beforehand. The phase_first_idx is the index of the first phase. The phase_second_idx is the index of the second phase. The first_node is the first node considered. The second_node is the second node considered.

Class: BinodeConstraintFcn

The BinodeConstraintFcn class is the already available binode constraint in bioptim. Since this is an Enum, it is possible to use the tab key on the keyboard to dynamically list them all, depending on the capabilities of your IDE.

  • EQUALITY — The states are equals.
  • COM_EQUALITY — The positions of centers of mass are equals.
  • COM_VELOCITY_EQUALITY — The velocities of centers of mass are equals.
  • CUSTOM — CUSTOM should not be directly sent by the user, but the user should pass the custom_transition function directly. You can look at the BinodeConstraintList section for more information about defining a custom transition function.

The phase transitions

Bioptim can declare multiphase optimisation programs. The goal of a multiphase ocp is usually to handle changing dynamics. The user must understand that each phase is, therefore, a full ocp by itself, with constraints that links the end of which with the beginning of the following. Due to some limitations created by using MX variables, some things can be done, and some cannot during a phase transition.

Class: PhaseTransitionList

The PhaseTransitionList provides a class that prepares the phase transitions. When constructing an OptimalControlProgram(), PhaseTransitionList is the expected class for the phase_transitions parameter.

The PhaseTransitionList class is the main class to define parameters. Please note that, unlike other lists, PhaseTransition is not accessible since phase transition does not make sense for single-phase ocp. Therefore, one should not call the PhaseTransition constructor directly.

Here is the full signature of the add() method of the PhaseTransitionList:

PhaseTransitionList.add(PhaseTransitionFcn, phase_pre_idx, **extra_parameters)

The PhaseTransitionFcn is the transition phase function to use. The default is CONTINUOUS. When declaring a custom transition phase, PhaseTransitionFcn is the function handler to the custom function. The signature of the custom function is: custom_function(transition: PhaseTransition nlp_pre: NonLinearProgram, nlp_post: NonLinearProgram, **extra_parameters), where nlp_pre is the nonlinear program at the end of the phase before the transition, nlp_post is the nonlinear program at the beginning of the phase after the transition, and the **extra_parameters are those sent to the add() method. This function is expected to return the cost of the phase transition computed from the states pre- and post-transition in the form of an MX. Please note that MX type is a CasADi type. Anyone who wants to define phase transitions should be at least familiar with this type beforehand. The phase_pre_idx is the index of the phase before the transition. If the phase_pre_idx is set to the index of the last phase, then this is equivalent to set PhaseTransitionFcn.CYCLIC.

Class: PhaseTransitionFcn

The PhaseTransitionFcn class is the already available phase transitions in bioptim. Since this is an Enum, it is possible to use the tab key on the keyboard to dynamically list them all, depending on the capabilities of your IDE.

  • CONTINUOUS — The states at the end of the phase_pre equals the states at the beginning of the phase_post
  • IMPACT — The impulse function of biorbd: qdot_post = bio_model.qdot_from_impact, q_pre, qdot_pre) is applied to compute the velocities of the joint post impact. These computed states at the end of the phase_pre equals those at the beginning of the phase_post. If a bioMod has more contact points than the model in the previous phase, then the IMPACT transition phase should also be used.
  • CYCLIC — Apply the CONTINUOUS phase transition from the end of the last phase to the beginning the first one, effectively creating a cyclic movement.
  • CUSTOM — the user should not send CUSTOM directly but pass the custom_transition function. You can look at the PhaseTransitionList section for more information about defining a custom transition function.

The results

Bioptim offers different ways to manage and visualize the results from an optimization. This section explores the different methods that can be called to have a look at your data.

Everything related to managing the results can be accessed from the solution class returned from

sol = ocp.solve()

Data manipulation

The Solution structure holds all the optimized values. To get the states variable, control variables, and time, one can invoke each property.

states = sol.states
controls = sol.controls
time = sol.time

If the program was a single-phase problem, then the returned values are dictionaries, otherwise, it is a list of dictionaries of size equal to the number of phases. The keys of the returned dictionaries correspond to the name of the variables. For instance, if generalized coordinates (q) are states, the state dictionary has q as key. In any case, the key all is always there.

# single-phase case
q = sol.states["q"]  # generalized coordinates
q = sol.states["all"]  # all states
# multiple-phase case - states of the first phase
q = sol.states[0]["q"]
q = sol.states[0]["all"]

The values inside the dictionaries are np.ndarray of dimension n_elements x n_shooting, unless the data were previously altered by integrating or interpolating (then the number of columns may differ).

The parameters are very similar but differ because they are always a dictionary (since they do not depend on the phases). Also, the values inside the dictionaries are of dimension n_elements x 1.

Integrate

It is possible to integrate (also called simulate) the states at will by calling the sol.integrate() method. The shooting_type: Shooting parameter allows you to select the type of integration to perform (see the enum Shooting for more detail). The keep_intermediate_points parameter allows us to keep the intermediate shooting points (usually a multiple of n_steps of the Runge-Kutta) or collocation points. If set to false, these points are not stored in the output structure. By definition, setting keep_intermediate_points to True while asking for Shooting.MULTIPLE would return the same structure. This will therefore raise an error if set to False with Shooting.MULTIPLE. The merge_phase: bool parameter requests to merge all the phases into one [True] or not [False]. The continuous: bool parameter can be deceiving. It is mostly for internal purposes.

Here are the tables of the combinations for sol.integrate and shooting_types. As the argument keep_intermediates_points does not significantly affect the implementations, it has been withdrawn from the tables. If implemented, it will be done with keep_intermediates_points=True or False.

Let us begin with shooting_type = Shooting.SINGLE, it re-integrates the ocp as a single phase ocp :

Shooting.SINGLE
OdeSolver
merge_phase
Solution
Integrator
Implemented Comment
DMS True OCP âś…
DMS False OCP âś…
DMS True SCIPY âś…
DMS False SCIPY âś…
COLLOCATION True OCP ❌ COLLOCATION Solvers cannot be used with single shooting
COLLOCATION False OCP ❌ COLLOCATION Solvers cannot be used with single shooting
COLLOCATION True SCIPY âś…
COLLOCATION False SCIPY âś…
Shooting.SINGLE_DISCONTINUOUS_PHASES

Let's pursue with shooting_type = Shooting.SINGLE_DISCONTINUOUS_PHASES, it re-integrates each phase of the ocp as a single phase ocp. Thus, SINGLE and SINGLE_DISCONTINUOUS_PHASES are equivalent if there is only one phase. Here is the table:

OdeSolver
merge_phase
Solution
Integrator
Implemented Comment
DMS True OCP âś…
DMS False OCP âś…
DMS True SCIPY âś…
DMS False SCIPY âś…
COLLOCATION True OCP ❌ COLLOCATION Solvers cannot be used with single shooting
COLLOCATION False OCP ❌ COLLOCATION Solvers cannot be used with single shooting
COLLOCATION True SCIPY âś…
COLLOCATION False SCIPY âś…
Shooting.MULTIPLE

Let us finish with shooting_type = Shooting.MULTIPLE, please note that this cannot be used with keep_intermediates_points=False. Also, the word MULTIPLE refers to direct multiple shooting.

OdeSolver
merge_phase
Solution
Integrator
Implemented Comment
DMS True OCP âś…
DMS False OCP âś…
DMS True SCIPY âś…
DMS False SCIPY âś…
COLLOCATION True OCP ❌ The solution cannot be re-integrated with the ocp solver
COLLOCATION False OCP ❌ The solution cannot be re-integrated with the ocp solver
COLLOCATION True SCIPY âś… This is re-integrated with solve_ivp, as direct multiple shooting problem
COLLOCATION False SCIPY âś… This is re-integrated with solve_ivp, as direct multiple shooting problem

Interpolation

The sol.interpolation(n_frames: [int, tuple]) method returns the states interpolated by changing the number of shooting points. If the program is multiphase, but only a int is sent, then the phases are merged, and the interpolation keeps their respective time ratio consistent. If one does not want to merge the phases, then a tuple with one value per phase can be sent.

Merge phases

Finally, sol.merge_phases() returns a Solution structure with all the phases merged into one.

Please note that, apart from sol.merge_phases(), these data manipulation methods return an incomplete Solution structure. This structure can be used for further analyses but cannot be used for visualization. If one wants to visualize integrated or interpolated data, they must use the corresponding parameters or the visualization method they use.

Data visualization

The first data visualizing method is sol.graphs(). This method will spawn all the graphs associated with the ocp. This is the same method that is called by the online plotter. To add and modify plots, one should use the ocp.add_plot() method. By default, this graphs the states as multiple shootings. If one wants to simulate in single shooting, the option shooting_type=Shooting.SINGLE will do the trick.

A second one is sol.animate(). This method summons one or more bioviz figures (depending on whether phases were merged) and animates the model. Please note that despite bioviz best efforts, plotting a lot of meshing vertices in MX format is slow. So even though it is possible, it is suggested to animate without the bone meshing (by passing the parameter show_meshes=False) To do so, we strongly suggest saving the data and loading them in an environment where bioptim is compiled with the Eigen backend, which will be much more efficient. If n_frames is set, an interpolation is performed. Otherwise, the phases are merged if possible, so a single animation is shown. To prevent phase merging, one can set n_frames=-1.

In order to print the values of the objective functions and constraints, one can use the sol.print_cost() method. If the parameter cost_type=CostType.OBJECTIVE is passed, only the values of each objective functions are printed. The same is true for the constraints with CostType.CONSTRAINTS. Please note that for readability purposes, this method prints the sum by phases for the constraints.

The extra stuff and the Enum

It was hard to categorize the remaining classes and enum. So I present them in bulk in this extra stuff section.

The mappings

The mapping is a way to link things stored in a list. For instance, consider these vectors: a = [0, 0, 0, 10, -9] and b = [10, 9]. Even though they are quite different, they share some common values. It is, therefore, possible to retrieve a from b, and conversely.

This is what the Mapping class does for the rows of numpy arrays. So if one was to declare the following Mapping: b_from_a = Mapping([3, -4]). Then, assuming a is a numpy.ndarray column vector (a = np.array([a]).T), it would be possible to summon b from a like so:

b = b_from_a.map(a)

Note that the -4 opposed the fourth value. Conversely, using the a_from_b = Mapping([None, None, None, 0, -1]) mapping, and assuming b is a numpy.ndarray column vector (b = np.array([b]).T), it would be possible to summon b from a like so:

a = a_from_b.map(b)

Note that the None are replaced by zeros.

The BiMapping is no more no less than a list of two mappings that link two matrices both ways: BiMapping(a_to_b, b_to_a)

The SelectionMapping is a subclass of BiMapping where you only have to precise the size of the first matrix, and the mapping b_to_a to get the second matrix from the first. If some elements depend on others, you can add an argument dependency:SelectionMapping(size(int), b_to_a; tuple[int, int, ...], dependencies :tuple([int, int, bool]))

Enum: Node

The node targets some specific nodes of the ocp or a phase. The accepted values are:

  • START: The first node
  • MID: The middle node
  • INTERMEDIATES: All the nodes but the first and the last one
  • PENULTIMATE: The second to last node of the phase
  • END: The last node
  • ALL: All the nodes
  • TRANSITION: The last node of a phase and the first node of the next phase

Class: OdeSolver

The ordinary differential equation (ode) solver to solve the dynamics of the system. The RK4 and RK8 are the ones with the most options available. IRK may be more robust but slower. CVODES is the one with the least options since it is not in-house implemented.

The accepted values are:

  • For Direct multiple shooting:
    • RK1: Runge-Kutta of the 1st order also known as Forward Euler
    • RK2: Runge-Kutta of the 2nd order also known as Midpoint Euler
    • RK4: Runge-Kutta of the 4th order
    • RK8: Runge-Kutta of the 8th order
    • IRK: Implicit Runge-Kutta (Legendre and Radau, from 0th to 9th order)
    • CVODES: cvodes solver
  • For Direct collocation:
    • COLLOCATION: Legendre and Radau, from 0th to 9th order
    • TRAPEZOIDAL: Trapezoidal rule

Enum: Solver

The nonlinear solver to solve the whole ocp. Each solver has some requirements (for instance, ̀Acados necessitates that the graph is SX). Feel free to test each of them to see which fits your needs best. ̀Ipopt is a robust solver, that may be slow. ̀Acados, on the other hand, is a very fast solver, but is much more sensitive to the relative weightings of the objective functions and the initial guess. It is perfectly designed for MHE and NMPC problems.

The accepted values are:

  • Ě€Ipopt
  • Ě€Acados
  • Ě€SQP

Enum: PhaseDynamics

  • SHARED_DURING_THE_PHASE
  • ONE_PER_NODE

The argument should be set to SHARED_DURING_THE_PHASE if we assume the dynamics are the same within each phase of the ocp problem. This argument increases the speed to mount the problem; it should be considered each time you build an Optimal Control Program. The default value is ONE_PER_NODE, meaning we consider the dynamic equations to be different for each shooting node (e.g., when applying a different external force at each shooting node).

In the case, you want to use this feature you have to specify it when adding the dynamics of each phase.

dynamics = Dynamics(DynamicsFcn.TORQUE_DRIVEN, phase_dynamics=PhaseDynamics.SHARED_DURING_THE_PHASE)

Enum: ControlType

The type the controls are. Typically, the controls for an optimal control program are constant over the shooting intervals. However, one may want to get non-constant values. Bioptim has therefore implemented some other types of controls.

The accepted values are:

  • CONSTANT: The controls remain constant over the interval. The number of control is therefore equals to the number of shooting points.
  • LINEAR_CONTINUOUS: The controls are linearly interpolated over the interval. Since they are continuous, the end of an interval corresponds to the beginning of the next. The number of controls equals to the number of shooting point + 1 control.
  • CONSTANT_WITH_LAST_NODE: The controls remain constant over the interval. The number of control is therefore equals to the number of shooting point + 1 controls. (Note that the last control has no effect on the states so it should be constrained. This type of control is useful when you want the instantaneous slope of the states at the last node)
  • NONE: The problem has no controls. (It is useful if the system is driven by parameters)

Enum: PlotType

When adding a plot, it is possible to change the aspect of it.

The accepted values are: PLOT: Normal plot that links the points. INTEGRATED: Plot that links the points within an interval but is discrete between its end and the beginning of the next interval. STEP: Step plot, constant over an interval. POINT: Point plot.

Enum: OnlineOptim

The type of online plotter to use.

The accepted values are: NONE: No online plotter. DEFAULT: Use the default online plotter depending on the OS (MULTIPROCESS on Linux, MULTIPROCESS_SERVER on Windows and NONE on MacOS). MULTIPROCESS: The online plotter is in a separate process. SERVER: The online plotter is in a separate server. MULTIPROCESS_SERVER: The online plotter using the server automatically setup on a separate process.

Enum: InterpolationType

Defines wow a time-dependent variable is interpolated. It is mainly used for phases time span. Therefore, first and last nodes refer to the first and last nodes of a phase.

The accepted values are:

  • CONSTANT: Requires only one column; all the values are equal during the whole period of time.
  • CONSTANT_WITH_FIRST_AND_LAST_DIFFERENT: Requires three columns. The first and last columns correspond to the first and last nodes, while the middle corresponds to all the other nodes.
  • LINEAR: Requires two columns. It corresponds to the first and last nodes. The middle nodes are linearly interpolated to get their values.
  • EACH_FRAME: Requires as many columns as there are nodes. It is not an interpolation per se, but it allows the user to specify all the nodes individually.
  • ALL_POINTS: Requires as many columns as there are collocation points. It is not an interpolation per se, but it allows the user to specify all the collocation points individually.
  • SPLINE: Requires five columns. It performs a cubic spline to interpolate between the nodes.
  • CUSTOM: User-defined interpolation function.

Enum: MagnitudeType

The type of magnitude you want for the added noise. Either relative to the bounds (0 is no noise, 1 is the value of your bounds), or absolute

The accepted values are:

  • ABSOLUTE: Absolute noise of a chosen magnitude.
  • RELATIVE: Relative noise to the bounds (0 is no noise, 1 is the value of your bounds).

Enum: Shooting

The type of integration to perform

  • SINGLE: It re-integrates the solution as a single-phase optimal control problem
  • SINGLE_DISCONTINUOUS_PHASE: It re-integrates each phase of the solution as a single-phase optimal control problem. The phases are, therefore, not continuous.
  • MULTIPLE: The word MULTIPLE is used as a common terminology to be able to execute DMS and COLLOCATION. It refers to the fact that there are several points per interval, shooting points in DMS and collocation points in COLLOCATION.

Enum: CostType

The type of cost

  • OBJECTIVES: The objective functions
  • CONSTRAINTS: The constraints
  • ALL: All the previously described cost type

Enum: SolutionIntegrator

The type of integrator used to integrate the solution of the optimal control problem

  • OCP: The OCP integrator initially chosen with OdeSolver
  • SCIPY_RK23: The scipy integrator RK23
  • SCIPY_RK45: The scipy integrator RK45
  • SCIPY_DOP853: The scipy integrator DOP853
  • SCIPY_BDF: The scipy integrator BDF
  • SCIPY_LSODA: The scipy integrator LSODA

Enum: QuadratureRule

The type of integration used to integrate the cost function terms of Lagrange:

  • RECTANGLE_LEFT: The integral is approximated by a left rectangle rule (Left Riemann sum).
  • RECTANGLE_RIGHT: The integral is approximated by a right rectangle rule (Right Riemann sum).
  • MIDPOINT: The integral is approximated by a midpoint rectangle rule (Midpoint Riemann sum).
  • APPROXIMATE_TRAPEZOIDAL: The integral is approximated by a trapezoidal rule using the state at the beginning of the next interval.
  • TRAPEZOIDAL: The integral is approximated by a trapezoidal rule using the state at the end of the current interval.

Enum: RigidBodyDynamics

The type of transcription of any dynamics (e.g., rigidbody_dynamics or soft_contact_dynamics):

  • ODE: the dynamics is handled explicitly in the continuity constraint of the ordinary differential equation of the Direct Multiple Shooting approach.
  • DAE_INVERSE_DYNAMICS: it adds an extra control qddot to respect inverse dynamics on nodes; this is a DAE-constrained OCP.
  • DAE_FORWARD_DYNAMICS: it adds an extra control qddot to respect forward dynamics on nodes; this is a DAE-constrained OCP.
  • DAE_INVERSE_DYNAMICS_JERK: it adds an extra control qdddot and an extra state qddot to respect inverse dynamics on nodes; this is a DAE-constrained OCP.
  • DAE_FORWARD_DYNAMICS_JERK: it adds an extra control qdddot and an extra state qddot to respect forward dynamics on nodes; this is a DAE-constrained OCP.

Enum: SoftContactDynamics

The type of transcription of any dynamics (e.g., rigidbody_dynamics or soft_contact_dynamics):

  • ODE: soft contact dynamics is handled explicitly.
  • CONSTRAINT: an extra control fext is added, and it ensures respecting soft contact_dynamics on nodes through a constraint.

Enum: DefectType

  • EXPLICIT: The defect comes from the explicit formulation.
  • IMPLICIT: The defect comes from the implicit formulation.
  • NOT_APPLICABLE: The defect is not applicable.

Examples

In this section, we describe all the examples implemented with bioptim. They are ordered in separate files. Each subsection corresponds to the different files, dealing with different examples and topics. Please note that the examples from the paper (see Citing) can be found in this repo https://github.com/s2mLab/BioptimPaperExamples.

Run examples

A GUI to access the examples can be run to facilitate the testing of bioptim You can run the file __main__.py in the examples folder or execute the following command.

python -m bioptim.examples

Please note that pyqtgraph must be installed to run this GUI.

Getting started

In this subsection, all the examples of the getting_started file are described.

This example is a trivial box sent upward. It is designed to investigate the different bounds defined in bioptim. Therefore, it shows how to define the bounds, i.e., the minimal and maximal values of the state and control variables.

All the types of interpolation are shown: CONSTANT, CONSTANT_WITH_FIRST_AND_LAST_DIFFERENT, LINEAR, EACH_FRAME, SPLINE, and CUSTOM.

When the CUSTOM interpolation is chosen, the functions custom_x_bounds_min and custom_x_bounds_max
provide custom x bounds. The functions custom_u_bounds_min and custom_u_bounds_max provide custom u bounds. In this particular example, linear interpolation is mimicked using these four functions.

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. It is designed to show how to define custom constraints function if the available constraints do not fulfill your need.

This example reproduces the behavior of the SUPERIMPOSE_MARKERS constraint.

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. It is designed to show how to define a custom dynamics function if the provided ones are not sufficient.

This example reproduces the behavior of the DynamicsFcn.TORQUE_DRIVEN using custom dynamics.

The custom_dynamic function is used to provide the derivative of the states. The custom_configure function is used to tell the program which variables are states and controls.

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. It is designed to investigate the different ways to define the initial guesses at each node sent to the solver.

All the types of interpolation are shown: CONSTANT, CONSTANT_WITH_FIRST_AND_LAST_DIFFERENT, LINEAR, EACH_FRAME, SPLINE, and CUSTOM.

When the CUSTOM interpolation is chosen, the custom_init_func function is used to custom the initial guesses of the states and controls. In this particular example, the CUSTOM interpolation mimics linear interpolation.

This example is a trivial box that tries to superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. It is designed to show how one can define its own custom objective function if the provided ones are not sufficient.

This example reproduces the behavior of the Mayer.SUPERIMPOSE_MARKERS objective function.

This example is close to the example of the custom_constraint.py file. We use the custom_func_track_markers to define the objective function. In this example, the CUSTOM objective mimics ObjectiveFcn.SUPERIMPOSE_MARKERS.

This example is a clone of the pendulum.py example with the difference that the model now evolves in an environment where gravity can be modified. The goal of the solver is to find the optimal gravity (target = 8 N/kg) while performing the pendulum balancing task.

It is designed to show how to define parameters.

The my_parameter_function function is used to modify the dynamics. In our case, we want to optimize the gravity. This function is called right before defining the dynamics of the system. The my_target_function function is a penalty function. Both functions define a new parameter, and then a parameter objective function is linked to this new parameter.

This example is a trivial multiphase box that must superimpose different markers at the beginning and end of each phase with one of its corners. It is designed to show how to define CUSTOM phase transition constraints if the provided ones are insufficient.

This example mimics the behavior of the most common PhaseTransitionFcn.CONTINUOUS

The custom_phase_transition function defines the constraint of the transition to apply. This function can be used when adding some phase transitions to the list of phase transitions.

Different phase transitions can be considered. By default, all the phase transitions are continuous. However, if one or more phase transitions are desired to be continuous, it is possible to define and use a function like the custom_phase_transition function or directly use PhaseTransitionFcn.IMPACT. If a phase transition is desired between the last and the first phase, use the dedicated PhaseTransitionFcn.Cyclic.

This example is a trivial example of using the pendulum without any objective. It is designed to show how to create new plots and expand pre-existing ones with new information.

We define the custom_plot_callback function, which returns the value(s) to plot. We use this function as an argument of ocp.add_plot. Let us describe the creation of the plot "My New Extra Plot". custom_plot_callback takes two arguments, x and the array [0, 1, 3], as you can see below :

ocp.add_plot("My New Extra Plot", lambda x, u, p: custom_plot_callback(x, [0, 1, 3]), plot_type=PlotType.PLOT)

We use the plot_type PlotType.PLOT. It is a way to plot the first, second, and fourth states (i.e., q_Seg1_TransY, q_Seg1_RotX and qdot_Seg1_RotX) in a new window entitled "My New Extra Plot". Please note that for further information about the different plot types, you can refer to the section "Enum: PlotType".

#TODO

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. Moreover, the movement must be cyclic, meaning that the states at the end and the beginning are equal. It is designed to provide a comprehensible example of the way to declare a cyclic constraint or objective function.

A phase transition loop constraint is treated as a hard penalty (constraint) if weight is <= 0 [or if no weight is provided], or as a soft penalty (objective) otherwise, as shown in the example below :

phase_transitions = PhaseTransitionList()
if loop_from_constraint:
    phase_transitions.add(PhaseTransitionFcn.CYCLIC, weight=0)
else:
    phase_transitions.add(PhaseTransitionFcn.CYCLIC, weight=10000)

loop_from_constraint is a boolean. It is one of the parameters of the prepare_ocp function of the example. This parameter is a way to determine if the looping cost should be a constraint [True] or an objective [False].

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner on a different marker at the end. While doing so, a force pushes the box upward. The solver must minimize the force to lift the box while reaching the marker in time. It is designed to show how to use external forces. An example of external forces that depends on the state (for example, a spring) can be found at 'examples/torque_driven_ocp/spring_load.py'

Bioptim expects external_forces to be a np.ndarray [6 x n x n_shooting], where the six components are [Mx, My, Mz, Fx, Fy, Fz], expressed at the origin of the global reference frame for each node.

#TODO

This example mimics what a jumper does when maximizing the predicted height of the center of mass at the peak of an aerial phase. It does so with a simplistic two segments model. It is a clone of 'torque_driven_ocp/maximize_predicted_height_CoM.py' using the option MINIMIZE_PREDICTED_COM_HEIGHT. It is different in that the contact forces on the ground have to be downward (meaning that the object is limited to push on the ground, as one would expect when jumping).

Moreover, the lateral forces must respect some NON_SLIPPING constraint (i.e., the ground reaction forces have to remain inside of a cone of friction), as shown in the part of the code defining the constraints:

constraints = ConstraintList()
   constraints.add(
   ConstraintFcn.TRACK_CONTACT_FORCES,
   min_bound=min_bound,
   max_bound=max_bound,
   node=Node.ALL,
   contact_index=1,
   )
constraints.add(
    ConstraintFcn.TRACK_CONTACT_FORCES,
    min_bound=min_bound,
    max_bound=max_bound,
    node=Node.ALL,
    contact_index=2,
    )
constraints.add(
    ConstraintFcn.NON_SLIPPING,
    node=Node.ALL,
    normal_component_idx=(1, 2),
    tangential_component_idx=0,
    static_friction_coefficient=mu,
    )

Let us describe the code above. First, we create a list of constraints. Then, two contact forces are defined with the indexes 1 and 2, respectively. The last step is the implementation of a non-slipping constraint for the two forces defined before.

This example is designed to show how to use min_bound and max_bound values so they define inequality constraints instead of equality constraints, which can be used with any ConstraintFcn.

This example shows how to use the joints' acceleration dynamic to achieve the same goal as the simple pendulum but with a double pendulum for which only the angular acceleration of the second pendulum is controlled.

In fact, examples of mapping can be found at 'examples/symmetrical_torque_driven_ocp/symmetry_by_mapping.py'. and 'examples/getting_started/example_inequality_constraint.py'.

#TODO

#TODO

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and a different marker at the end of each phase. Moreover, a constraint on the rotation is imposed on the cube. It is designed to show how to define a multiphase optimal control program.

In this example, three phases are implemented. The long_optim boolean allows users to choose between solving the precise optimization or the approximate. In the first case, 500 points are considered: n_shooting = (100, 300, 100). Otherwise, 50 points are considered: n_shooting = (20, 30, 20). Three steps are necessary to define the objective functions, the dynamics, the constraints, the path constraints, the initial guesses, and the control path constraints. Each step corresponds to one phase.

Let us take a look at the definition of the constraints:

constraints = ConstraintList()
constraints.add(
    ConstraintFcn.SUPERIMPOSE_MARKERS, node=Node.START, first_marker_idx=0, second_marker_idx=1, phase=0
)
constraints.add(ConstraintFcn.SUPERIMPOSE_MARKERS, node=Node.END, first_marker_idx=0, second_marker_idx=2, phase=0)
constraints.add(ConstraintFcn.SUPERIMPOSE_MARKERS, node=Node.END, first_marker_idx=0, second_marker_idx=1, phase=1)
constraints.add(ConstraintFcn.SUPERIMPOSE_MARKERS, node=Node.END, first_marker_idx=0, second_marker_idx=2, phase=2)

First, we define a list of constraints, and then we add constraints to the list. At the beginning, marker 0 must superimpose marker 1. At the end of the first phase (the first 100 shooting nodes if we solve the precise optimization), marker 0 must superimpose marker 2. Then, at the end of the second phase, marker 0 must superimpose marker 1. At the end of the last step, marker 0 must superimpose marker 2. Please, note that the definition of the markers is implemented in the bioMod file corresponding to the model. Further information about the definition of the markers is available in the biorbd documentation.

#TODO

Examples of time optimization can be found in 'examples/optimal_time_ocp/'.

This example is a clone of getting_started/pendulum.py. It is designed to show how to create and solve a problem and, afterward, save it to the hard drive and reload it. It shows an example of the *.bo method.

Let us take a look at the most important lines of the example. To save the optimal control program and the solution, use ocp.save(sol, "pendulum.bo"). To load the optimal control program and the solution, use ocp_load, sol_load = OptimalControlProgram.load("pendulum.bo"). Then, to show the results, use sol_load.animate().

The first part of this example is a single shooting simulation from initial guesses. It is not an optimal control program. It is merely the simulation of values that is applying the dynamics. The main goal of this kind of simulation is to get a sense of the initial guesses passed to the solver.

The second part of the example is to solve the program and simulate the results from this solution. The main goal of this kind of simulation, especially in single shooting (i.e., not resetting the states at each node) is to validate the dynamics obtained by multiple shooting. If they both are equal, it usually means great confidence can be held in the solution. Another goal would be to reload fast a previously saved optimized solution.

#TODO

The pendulum.py file

This example is another way to present the pendulum example of the 'Getting started' section.

This example is a clone of the pendulum.py example with the difference that the states and controls are constrained instead of bounded. Sometimes the OCP converges faster with constraints than boundaries.

It is designed to show how to use bound_state and bound_control.

Torque-driven OCP

In this section, you will find different examples showing how to implement torque-driven optimal control programs.

This example mimics what a jumper does to maximize the predicted height of the center of mass at the peak of an aerial phase. It does so with a very simple two segments model. It is designed to give a sense of the goal of the different MINIMIZE_COM functions and the use of weight=-1 to maximize instead of minimize.

Let us take a look at the definition of the objective functions used for this example to understand better how to implement that:

objective_functions = ObjectiveList()
if objective_name == "MINIMIZE_PREDICTED_COM_HEIGHT":
    objective_functions.add(ObjectiveFcn.Mayer.MINIMIZE_PREDICTED_COM_HEIGHT, weight=-1)
elif objective_name == "MINIMIZE_COM_POSITION":
    objective_functions.add(ObjectiveFcn.Lagrange.MINIMIZE_COM_POSITION, axis=Axis.Z, weight=-1)
elif objective_name == "MINIMIZE_COM_VELOCITY":
    objective_functions.add(ObjectiveFcn.Lagrange.MINIMIZE_COM_VELOCITY, axis=Axis.Z, weight=-1)

Another interesting point of this example is the definition of the constraints. Thanks to the com_constraints boolean, the user can easily choose to apply or not constraints on the center of mass. Here is the definition of the constraints for our example:

constraints = ConstraintList()
if com_constraints:
    constraints.add(
        ConstraintFcn.TRACK_COM_VELOCITY,
        node=Node.ALL,
        min_bound=np.array([-100, -100, -100]),
        max_bound=np.array([100, 100, 100]),
    )
    constraints.add(
        ConstraintFcn.TRACK_COM_POSITION,
        node=Node.ALL,
        min_bound=np.array([-1, -1, -1]),
        max_bound=np.array([1, 1, 1]),
    )

This example is designed to show how to use min_bound and max_bound values so they define inequality constraints instead of equality constraints, which can be used with any ConstraintFcn. This example is close to the example_inequality_constraint.py file available in 'examples/getting_started/example_inequality_constraint.py'.

This trivial spring example aims to achieve the highest upward velocity. It can, however, only load a spring by pulling downward and then letting it go so it gains velocity. It is designed to show how to use external forces to interact with the body.

This example is close to the custom_dynamics.py file you can find in 'examples/getting_started/custom_dynamics.py'. Indeed, we generate an external force thanks to the custom_dynamic function. Then, we configure the dynamics with the custom_configure function.

This example uses the data from the balanced pendulum example to generate data to track. When it optimizes the program, contrary to the vanilla pendulum, it tracks the values instead of 'knowing' that it is supposed to balance the pendulum. It is designed to show how to track marker and kinematic data.

Note that the final node is not tracked.

In this example, we use both ObjectiveFcn.Lagrange.TRACK_MARKERS and ObjectiveFcn.Lagrange.TRACK_TORQUE objective functions to track data, as shown in the definition of the objective functions used in this example:

objective_functions = ObjectiveList()
objective_functions.add(
    ObjectiveFcn.Lagrange.TRACK_MARKERS, axis_to_track=[Axis.Y, Axis.Z], weight=100, target=markers_ref
)
objective_functions.add(ObjectiveFcn.Lagrange.TRACK_TORQUE, target=tau_ref)

This is a good example of how to load data for tracking tasks and how to plot data. The extra parameter axis_to_track allows users to specify the axes to track the markers (x and y axes in this example). This example is close to the example_save_and_load.py and custom_plotting.py files you can find in the examples/getting_started repository.

This example is a trivial box that must superimpose one of its corners on a marker at the beginning of the movement and superimpose the same corner to a different marker at the end. It is a clone of 'getting_started/custom_constraint.py'

It is designed to show how to use the TORQUE_ACTIVATIONS_DRIVEN, which limits the torque to [-1; 1]. This is useful when the maximal torques are not constant. Please note that such a dynamics may not converge when it is used on a more complicated model. A solution that defines non-constant constraints seems a better idea. An example can be found in the bioptim paper.

Let us take a look at the structure of the code. First, tau_min, tau_max, and tau_init are respectively initialized to -1, 1 and 0 if the integer actuator_type (a parameter of the prepare_ocp function) equals 1. In this case, the dynamics function used is DynamicsFcn.TORQUE_ACTIVATIONS_DRIVEN.

This example uses a representation of a human body by a trunk_leg segment and two arms. It is designed to show how to use a model that has quaternions in their degrees of freedom.

#TODO

#TODO

#TODO

#TODO

#TODO

#TODO

#TODO

The slider.py file

#TODO

#TODO

Muscle-driven OCP

In this folder, you will find four examples of muscle-driven optimal control programs. The two first refer to tracking examples. The two last refer to reaching tasks.

This example is about muscle activation/skin marker or state tracking. Random data are created by generating a random set of muscle activations and then by generating the kinematics associated with these controls. The solution is trivial since no noise is applied to the data. Still, it is a relevant example of how to track data using a musculoskeletal model. In a real situation, muscle activation and kinematics would indeed be acquired using data acquisition devices.

The difference between muscle activation and excitation is that the latter is the derivative of the former.

The generate_data function is used to create random data. First, a random set of muscle activations is generated, as shown below: U = np.random.rand(n_shooting, n_mus).T

Then, the kinematics associated with these data are generated by numerical integration, using scipy.integrate.solve_ivp.

To implement this tracking task, we use the ObjectiveFcn.Lagrange.TRACK_STATE objective function in the case of state tracking, or the ObjectiveFcn.Lagrange.TRACK_MARKERS objective function in the case of marker tracking. We also use the ObjectiveFcn.Lagrange.TRACK_MUSCLES_CONTROL objective function. The user can choose between marker or state tracking thanks to the string kin_data_to_track, which is one of the prepare_ocp function parameters.

This example concerns muscle excitation(EMG)/skin marker or state tracking. Random data are created by generating a random set of EMG and then by generating the kinematics associated with these data. The solution is trivial since no noise is applied to the data. Still, it is a relevant example of how to track data using a musculoskeletal model. The EMG and kinematics would be acquired in the real world using data acquisition devices.

There is no major difference with the previous example. Some dynamic equations link muscle activation and excitation.

The static_arm.py file

This is a basic example of using the biorbd muscle-driven model to perform an optimal reaching task. The arms must reach a marker placed upward in front while minimizing the muscles' activity.

For this reaching task, we use the ObjectiveFcn.Mayer.SUPERIMPOSE_MARKERS objective function. At the end of the movement, marker 0 and marker 5 should superimpose. The weight applied to the SUPERIMPOSE_MARKERS objective function is 1000. Please note that the bigger this number, the greater the model will try to reach the marker.

Please note that using show_meshes=True in the animator may be long due to the creation of a large CasADi graph of the mesh points.

This is a basic example of how to use biorbd model driven by muscle to perform an optimal reaching task with a contact dynamics. The arms must reach a marker placed upward in front while minimizing the muscles' activity.

The only difference with the previous example is that we use the arm26_with_contact.bioMod model and the DynamicsFcn.MUSCLE_ACTIVATIONS_AND_TORQUE_DRIVEN_WITH_CONTACT dynamics function instead of DynamicsFcn.MUSCLE_ACTIVATIONS_AND_TORQUE_DRIVEN.

Please note that using show_meshes=True in the animator may be long due to the creation of a huge CasADi graph of the mesh points.

Muscle driven with contact

All the examples in the folder muscle_driven_with_contact show some dynamics and prepare some OCP for the tests. They are not relevant and will be removed when unitary tests for the dynamics will be implemented.

In this example, we implement inequality constraints on two contact forces. It is designed to show how to use min_bound and max_bound values for the definition of inequality constraints instead of equality constraints, which can be used with any ConstraintFcn.

In this case, the dynamics function used is DynamicsFcn.MUSCLE_ACTIVATIONS_AND_TORQUE_DRIVEN_WITH_CONTACT.

In this example, we implement inequality constraints on two contact forces. It is designed to show how to use min_bound and max_bound values so they define inequality constraints instead of equality constraints, which can be used with any ConstraintFcn.

In this case, the dynamics function used is DynamicsFcn.MUSCLE_EXCITATIONS_AND_TORQUE_DRIVEN_WITH_CONTACT instead of DynamicsFcn.MUSCLE_ACTIVATIONS_AND_TORQUE_DRIVEN_WITH_CONTACT used in the previous example.

In this example, we track both muscle controls and contact forces, as it is defined when adding the two objective functions below, using both ObjectiveFcn.Lagrange.TRACK_MUSCLES_CONTROL and ObjectiveFcn.Lagrange.TRACK_CONTACT_FORCES objective functions.

objective_functions = ObjectiveList()
objective_functions.add(ObjectiveFcn.Lagrange.TRACK_MUSCLES_CONTROL, target=muscle_activations_ref)
objective_functions.add(ObjectiveFcn.Lagrange.TRACK_CONTACT_FORCES, target=contact_forces_ref)

Let us take a look at the structure of this example. First, we load data to track and generate data using the data_to_track.prepare_ocp optimization control program. Then, we track these data using muscle_activation_ref and contact_forces_ref as shown below:

ocp = prepare_ocp(
    biorbd_model_path=model_path,
    phase_time=final_time,
    n_shooting=ns,
    muscle_activations_ref=muscle_activations_ref[:, :-1],
    contact_forces_ref=contact_forces_ref,
)

Optimal time OCP

In this section, you will find four examples showing how to play with time parameters.

This example is a trivial multiphase box that must superimpose different markers at beginning and end of each phase with one of its corners. The time is free for each phase. It is designed to show how to define a multiphase ocp problem with free time.

In this example, the number of phases is 1 or 3. prepare_ocp function takes time_min, time_max and final_time as arguments. There are arrays of length 3 in the case of a 3-phase problem. In the example, these arguments are defined as shown below:

final_time = [2, 5, 4]
time_min = [1, 3, 0.1]
time_max = [2, 4, 0.8]
ns = [20, 30, 20]
ocp = prepare_ocp(final_time=final_time, time_min=time_min, time_max=time_max, n_shooting=ns)

We can make out different time constraints for each phase, as shown in the code below:

constraints.add(ConstraintFcn.TIME_CONSTRAINT, node=Node.END, min_bound=time_min[0], max_bound=time_max[0], phase=0)
if n_phases == 3:
    constraints.add(
        ConstraintFcn.TIME_CONSTRAINT, node=Node.END, min_bound=time_min[1], max_bound=time_max[1], phase=1
    )
    constraints.add(
        ConstraintFcn.TIME_CONSTRAINT, node=Node.END, min_bound=time_min[2], max_bound=time_max[2], phase=2
    )

This is a clone of the example/getting_started/pendulum.py where a pendulum must be balanced. The difference is that the time to perform the task is now free and minimized by the solver, as shown in the definition of the objective function used for this example:

objective_functions = ObjectiveList()
objective_functions.add(ObjectiveFcn.Mayer.MINIMIZE_TIME, weight=weight, min_bound=min_time, max_bound=max_time)

Please note that a weight of -1 will maximize time.

This example shows how to define such an optimal control program with a Mayer criterion (value of final_time).

The difference between Mayer and Lagrange minimization time is that the former can define bounds to the values, while the latter is the most common way to define optimal time.

This is a clone of the example/getting_started/pendulum.py where a pendulum must be balanced. The difference is that the time to perform the task is now free for the solver to change. This example shows how to define such an optimal control program.

In this example, a time constraint is implemented:

constraints = Constraint(ConstraintFcn.TIME_CONSTRAINT, node=Node.END, min_bound=time_min, max_bound=time_max)

Symmetrical torque driven OCP

In this section, you will find an example using symmetry by constraint and another using symmetry by mapping. In both cases, we simulate two rods. We must superimpose a marker on one rod at the beginning and another on the same rod at the end while keeping the degrees of freedom opposed.

The difference between the first example (symmetry_by_mapping) and the second one (symmetry_by_constraint) is that one (mapping) removes the degree of freedom from the solver, while the other (constraints) imposes a proportional constraint (equals to -1), so they are opposed. Please note that even though removing a degree of freedom seems a good idea, it is unclear if it is faster when solving with IPOPT.

This example imposes a proportional constraint (equals to -1) so that the rotation around the x-axis remains opposed for the two rodes during the movement.

Let us take a look at the definition of such a constraint:

constraints.add(ConstraintFcn.PROPORTIONAL_STATE, node=Node.ALL, first_dof=2, second_dof=3, coef=-1)

In this case, a proportional constraint is generated between the third degree of freedom defined in the bioMod file (first_dof=2) and the fourth one (second_dof=3). Looking at the cubeSym.The bioMod file used in this example shows that the dof with index 2 corresponds to the rotation around the x-axis for the first segment Seg1. The dof with index 3 corresponds to the rotation around the x-axis for the second segment Seg2.

This example imposes the symmetry as a mapping by completely removing the degree of freedom from the solver variables but interpreting the numbers properly when computing the dynamics.

A BiMapping is used. The way to understand the mapping is that if one is provided with two vectors, what would be the correspondence between those vectors. For instance, BiMapping([None, 0, 1, 2, -2], [0, 1, 2]) would mean that the first vector (v1) has 3 components, and to create it from the second vector (v2), you would do the following: v1 = [v2[0], v2[1], v2[2]]. Conversely, the second v2 has 5 components and is created from the vector v1 using: v2 = [0, v1[0], v1[1], v1[2], -v1[2]]. For the dynamics, it is assumed that v1 is to be sent to the dynamic functions (the full vector with all the degrees of freedom), while v2 is the one sent to the solver (the one with fewer degrees of freedom).

The BiMapping used is defined as a problem parameter, as shown below:

all_generalized_mapping = BiMapping([0, 1, 2, -2], [0, 1, 2])

Track

In this section, you will find the description of three tracking examples.

This example is a trivial example where a stick must keep a corner of a box in line for the whole duration of the movement. The initial and final positions of the box are dictated; the rest is fully optimized. It is designed to show how to use the tracking function for tracking a marker with a body segment.

In this case, we use the ConstraintFcn.TRACK_MARKER_WITH_SEGMENT_AXIS constraint function, as shown below in the definition of the constraints of the problem:

constraints = ConstraintList()
constraints.add(
ConstraintFcn.TRACK_MARKER_WITH_SEGMENT_AXIS, node=Node.ALL, marker_idx=1, segment_idx=2, axis=Axis.X
)

Here, we minimize the distance between the marker with index 1 and the x-axis of the segment with index 2. We align the axis toward the marker.

This example is a trivial example where a stick must keep its coordinate system of axes aligned with the one from a box during the whole duration of the movement. The initial and final positions of the box are dictated; the rest is fully optimized. It is designed to show how to use the tracking RT function for tracking any RT (for instance, Inertial Measurement Unit [IMU]) with a body segment.

To implement this tracking task, we use the ConstraintFcn.TRACK_SEGMENT_WITH_CUSTOM_RT constraint function, which minimizes the distance between a segment and an RT. The extra parameters segment_idx: int and rt_idx: int must be passed to the Objective constructor.

#TODO

Moving estimation horizon (MHE)

In this section, we perform MHE on the pendulum example.

The mhe.py file

In this example, MHE is applied to a simple pendulum simulation. Data are generated (states, controls, and marker trajectories) to simulate the movement of a pendulum, using scipy.integrate.solve_ivp. These data are used to perform MHE.

In this example, 500 shooting nodes are defined. As the size of the MHE window is 10, 490 iterations are performed to solve the complete problem.

For each iteration, the new marker trajectory is considered so that real-time data acquisition is simulated. For each iteration, the list of objectives is updated, the problem is solved with the new frame added to the window, the oldest frame is discarded with the warm_start_mhe function, and it is saved. The results are plotted to compare estimated data to real data.

#TODO

#TODO

Acados

In this section, you will find three examples to investigate bioptim using acados.

The cube.py file

This is a basic example of a cube that must reach a target at the end of the movement, starting from an initial position while minimizing states and torques. This problem is solved using acados.

The pendulum.py file

This simple yet meaningful optimal control program consists of a pendulum starting downward and ending upward while minimizing the generalized forces. The solver can only move the pendulum sideways.

This simple example is an excellent place to investigate bioptim using acados as it describes the most common dynamics (the joint torque driven). It also defines an objective function and some boundaries and initial guesses.

The static_arm.py file

This basic example shows how to use biorbd model driven by muscle to perform an optimal reaching task. The arm must reach a marker while minimizing the muscles' activity and the states. We solve the problem using both acados and ipopt.

Inverse optimal control

In this section, you will find an example of inverse optimal control with bioptim.

This basic example is a rigid double pendulum that circles a fixed point. The movement is inspired by the motion of gymnasts on the bar apparatus. This example is separated into three parts:

  • The first part is the definition of the problem. The problem is solved with specific weightings.
  • The second part solves the problem with only one objective at a time for the Pareto front.
  • The third part solves the inverse optimal control problem aiming to retrieve the initial weightings. A the end of the example, the markers' trajectories are plotted to show that the movement is the same.

Discrete mechanics and optimal control

#TODO

Fatigue

#TODO

Holonomic constraints

#TODO

SQP method

#TODO

Stochastic optimal control

#TODO

Performance

If you find yourself asking, "Why is bioptim so slow? I thought it was lightning fast!" Then this section may help you improve your code to get better performance.

use_sx

Set use_sx to True in the OptimalControlProgram class to use the SX symbolic variables. These are faster but require more RAM, so ensure you have enough RAM to use this option.

n_threads

Set n_threads to the number of threads you want to use in the OptimalControlProgram class. By default, it is set to 1. It will split the computation of the continuity constraints between threads and speed up the computation. If applicable to your problem, use the next option too.

expand

(For objective and constraint functions) Set the expand argument to True for objective and constraint functions to speed up the computation. It will turn MX symbolic variables into SX symbolic variables, which is faster but requires more RAM.

Troubleshooting

Despite our best efforts to assist you with this long Readme and several examples, you may experience some problems with bioptim. Fortunately, this troubleshooting section will guide you through solving some known issues.

Freezing compute

If your computer freezes before any optimization is performed, it is probably because your problem requires too much RAM. If you are using use_sx and/or expand options, try turning them off. If it does not work, try reducing the number of nodes.

Free variables

Sometimes when working on advanced custom problems, you may have free variables that prevent the solver from being launched. If this occurs, try reloading your model inside of the custom function. We have found this solution to be effective when working with biorbd models.

Non-converging problems

If Ipopt converges to an infeasible solution, ensure the boundaries are sound for the problem's constraints. If the problem still does not converge, try changing the initial guess of the problem.

If the problem takes numerous iterations to solve (much more than expected), check the weights on objective functions and the weight of the actual variables.

If the problem still does not converge, try observing the evolution of the objective function and the constraints through a live plot. It is always good to see how they evolve through the iterations.

Citing

If you use bioptim, we would be grateful if you could cite it as follows: @article{michaud2022bioptim, title={Bioptim, a python framework for musculoskeletal optimal control in biomechanics}, author={Michaud, Benjamin and Bailly, Fran{\c{c}}ois and Charbonneau, Eve and Ceglia, Amedeo and Sanchez, L{'e}a and Begon, Mickael}, journal={IEEE Transactions on Systems, Man, and Cybernetics: Systems}, year={2022}, publisher={IEEE} }

About

An optimization framework that links CasADi, ipopt and biorbd for Optimal Control Problem

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%