API  4.3
For MATLAB, Python, Java, and C++ users
MocoStudy: custom optimal control problems


MocoStudy allows one to solve a generic optimal control problem, where the costs, constraints, and solver settings are fully defined by the user.

Solving a custom optimal control problem with direct collocation requires substantial tuning of the direct collocation method's settings. Use the more specific interfaces MocoInverse and MocoTrack if possible.

For a mathematical description, see Optimal control problem.

In Moco, "goals", such as tracking marker data or effort, can be enforced either as cost terms to minimize or endpoint constraints.

Overview

A MocoStudy contains a MocoProblem and a MocoSolver.

  • MocoProblem: describes the variables, dynamics, constraints, and cost functional of the optimal control problem, using an OpenSim model.
  • MocoSolver: solves the MocoProblem using direct collocation.

When building a MocoStudy programmatically (e.g., in C++), the workflow is as follows:

  1. Build the MocoProblem (set the model, constraints, etc.).
  2. Call MocoStudy::initTropterSolver() or MocoStudy::initCasADiSolver(), which returns a reference to a MocoSolver.
  3. Edit the settings of the MocoSolver.
  4. Call MocoStudy::solve(). This returns the MocoSolution.
  5. (Optional) Postprocess the solution, perhaps using MocoStudy::visualize().

After calling MocoStudy::solve(), you can edit the MocoProblem and/or the MocoSolver and call MocoStudy::solve() again, if you wish.

MocoProblem

A MocoProblem is a list of MocoPhases, but Moco currently supports only single-phase problems. A MocoPhase consists of the following:

  1. An OpenSim model that describes states, controls, dynamics, and kinematic constraints.
  2. MocoGoals: describes the cost functional to minimize or endpoint constraints to enforce.
  3. MocoPathConstraints: describes non-kinematic algebraic constraints to enforce over the entire phase.
  4. MocoParameters: describes Model (time-invariant) properties to optimize.
  5. Bounds on initial and final time.
  6. Bounds (initial, final, and over all time) for state variables, stored in MocoVariableInfo.
  7. Bounds (initial, final, and over all time) for control variables, stored in MocoVariableInfo.

Bounds

Time

By default, the initial time and final time are unconstrained. However, this is rarely what you want, and often causes a problem to behave poorly. Moco issues a warning if you do not set bounds for initial or final time.

Defaults

Setting bounds for state and control variables is an important but often tedious process. To help, Moco sets some variable bounds automatically. See MocoPhase::setStateInfo() and MocoPhase::setControlInfo() for details.

Although Moco supports setting initial and final bounds for controls, such functionality is not too useful: the initial and final values of a control have little influence on the value of the control at nearby times.

Defining the cost functional and endpoint constraints

See Moco: Available goals for more information.

Parameter optimization with MocoParameter

Moco allows you to optimize most model parameters by adding a MocoParameter to your MocoProblem. See MocoParameter for more information.

Prescribed kinematics

You may wish to solve a problem in which you prescribe known kinematics for all degrees of freedom and optimize only auxiliary dynamics or actuator behavior. In this formulation, the generalized coordinate and speed variables are replaced with known quantities. The system still contains auxiliary state variables, control variables, and auxiliary dynamics. If none of the parameter variables affect the multibody system, then the multibody dynamics is reduced to a force balance: applied forces must match the net generalized forces determined by the kinematics (that is, inverse dynamics).

To prescribe motion in the "reduce-DOF" way, add a PositionMotion component to your model. PositionMotion requires that all DOFs are prescribed; Moco does not support prescribing a subset of the degrees of freedom.

See MocoInverse: solving muscle and actuator redundancy and Prescribed kinematics for more information.

Prescribed kinematics with kinematic constraints

Prescribing kinematics is more complicated if the system contains kinematic constraints (e.g., weld constraints, point-on-line constraints, etc.), as such constraints can conflict with the prescribed kinematics (which is also essentially a constraint).

Note
If a model contains kinematic constraints, you must ensure that the prescribed motion obeys the kinematic constraints. PositionMotion does not ensure this for you. MocoInverse attempts to alter your provided kinematics to satisfy the kinematic constraints (using Model::assemble()).

Implicit auxiliary dynamics

OpenSim's Components support auxiliary dynamics, that is, additional dynamics added to the multibody dynamics equations that define a model's computational system (e.g., muscle activation dynamics). Components support dynamics in explicit form (i.e., \( \dot{y} = f(y) \)), but it is often desirable to implement auxiliary dynamics in implicit form (i.e., \( f(y, \dot{y}) = 0 \)). For example, expressing tendon compliance dynamics explicitly requires inverting the force-velocity curve, which introduces singularities. Solving the dynamics implicitly by enforcing muscle-tendon equilibrium (i.e. \( F^T - F^M cos(\alpha) = 0 \)) is more natural and avoids such numerical issues [1].

OpenSim's numerical integrators only support explicit differential equations, but Moco allows Components to express dynamics with implicit differential equations in addition to explicit differential equations. Currently, implicit differential equations are provided for the following pairs of components and state variables:

See the documentation for these components to learn how to set their dynamics mode.

Implicit differential equations require an additional control for the derivative of the state variable; MocoTrajectories associated with a model containing implicit auxiliary dynamics contain columns for these state variable derivatives.

To set the mode for multibody dynamics, see Implicit multibody dynamics mode.

Note
Implicit versus explicit integrators: Differential equations can be implicit and explicit, but so can numerical integrators. This is often a point of confusion. Implicit and explicit numerical integrators describe types of methods for solving differential equations; both implicit and explicit numerical integrators can be used with explicit differential equations, while only implicit numerical integrators can be used with implicit differential equations.

To implement implicit auxiliary differential equations yourself, see exampleCustomImplicitAuxiliaryDynamics.cpp.

MocoSolver

A MocoSolver attempts to use an optimal control method to solve the MocoProblem, and exposes settings you can use to control the transcription to a generic optimization problem and the optimization process.

Implicit multibody dynamics mode

MocoSolvers support two modes for expressing differential equations for multibody dynamics. When using the implicit dynamics mode, the differential equations for multibody dynamics are expressed as implicit differential equations. See the descriptions for the specific transcription schemes for more detail. As with implicit auxiliary dynamics, using implicit multibody dynamics often leads to a more robust problem, especially for systems with small masses [2].

Kinematic constraints

Moco supports enforcing OpenSim kinematic constraints in optimal control problems. When a solver is initialized, any kinematic constraints enabled in the model are automatically detected, and additional information about each constraint is stored. This information includes the number of scalar path constraints (including derivatives if they exist), the scalar constraint kinematic levels (i.e. holonomic, non-holonomic, or acceleration), and the number of Lagrange multipliers that are added to the problem to enforce the constraints.

The solvers allow the user to choose whether to enforce kinematic constraint derivatives or minimize Lagrange multipliers in the problem.

MocoTropterSolver& solver = moco.initTropterSolver();
solver.set_enforce_constraint_derivatives(true);
solver.set_lagrange_multiplier_weight(10);

If constraint derivatives are enforced, velocity correction variables are added to the problem according to the method described in Posa, Kuindersma, and Tedrake, 2016, "Optimization and stabilization of trajectories for constrained dynamical systems."

Note
Enforcing kinematic constraints is only supported with Hermite-Simpson transcription, since this scheme is necessary for implementing the method described in Posa et al. 2016. See Hermite–Simpson transcription for a mathematical description.

It is possible to modify the default bounds on these velocity correction variables:

// Default bounds: [-0.1, 0.1]
solver.set_velocity_correction_bounds({-0.25, 0.25});

If constraint derivatives are not enforced or if the multibody constraint Jacobian is rank-deficient, it is recommended that the Lagrange multiplier minimization term is added to the problem to impose uniqueness for these variables.

Even though OpenSim Models can contain non-holonomic constraints (that is, \( \nu(q, u, p) = 0 \)), Moco solvers support only holonomic constraints; that is, \( \phi(q) = 0 \).

By default, the controls at mesh interval midpoints are constrained by linear interpolation of control mesh endpoint values and is the recommended setting. However, you may disable this behavior in the solver:

MocoCasADiSolver& solver = moco.initCasADiSolver();
solver.set_interpolate_control_midpoints(false);

MocoCasADiSolver versus MocoTropterSolver

Moco provides two direct collocation solvers that transcribe continuous optimal control problems into finite-dimensional NLPs that are passed to well-established derivative-based NLP solvers. MocoCasADiSolver uses the third-party CasADi library, and MocoTropterSolver uses a direct collocation solver developed internally named Tropter. CasADi is an open-source package for algorithmic differentiation and is a bridge to NLP solvers IPOPT, SNOPT, and others.

Compared to MocoTropterSolver, MocoCasADiSolver supports parallelization, implicit auxiliary dynamics, and implicit multibody dynamics with kinematic constraints. However, MocoTropterSolver may be able to solve parameter optimization problems more quickly that MocoCasADiSolver.

CasADi is an ideal library for employing direct collocation, but two limitations led us to create Tropter: CasADi did not initially support finite differences, and CasADi’s open-source license is more restrictive than OpenSim’s. Those distributing Moco as a dependency of closed-source software may prefer distributing Moco without CasADi, as CasADi’s “weak copyleft” GNU Lesser General Public License 3.0 places requirements on how CasADi is redistributed.

Tips for solving a custom problem

  1. Make sure every variable has bounds.
  2. Create a dynamically-consistent initial guess by solving a feasibility problem. That is, solve a MocoStudy without any cost terms.
  3. Actuators should only be as strong as they need to be. Problems tend to converge slower if some actuator controls are very small compared to those of other actuators in the model due to unecessarily large max force settings.
  4. Add reserve actuators to muscle-driven problems (whether inverse, tracking, predicdtive, etc.) to improve convergence and identify muscles that are too weak to drive a desired motion.
  5. If controls display high frequency oscillations, try tightening the solver's convergence and constraint tolerances.

References

[1] Groote, F., Kinney, A., Rao, A., Fregly, B. (2016). Evaluation of Direct Collocation Optimal Control Problem Formulations for Solving the Muscle Redundancy Problem Annals of Biomedical Engineering 44(10), 2922-2936. https://dx.doi.org/10.1007/s10439-016-1591-9

[2] Bogert, A., Blana, D., Heinrich, D. (2011). Implicit methods for efficient musculoskeletal simulation and optimal control Procedia IUTAM 2(), 297-316. https://dx.doi.org/10.1016/j.piutam.2011.04.027