QpSolverCollection — real‑time optimal‑control toolkit

Overview

Fork of isri‑aist/QpSolverCollection with additional solvers, utilities and examples developed during my work at RWTH‑IRT

This fork broadens the original library with Ipopt‑powered nonlinear programming, several MPC and NMPC front‑ends, and a battery of regression tests that guard numerical correctness down to the solver back‑end. Everything – from automatic differentiation to QP factorisation – stays inside C++ templates so you can exploit the full power of modern compilers without run‑time overhead.

What’s new in this fork

• IpoptInterface – full wrapper around COIN‑OR Ipopt for large‑scale, sparse NLPs.
• MpcSolver  ➜ ready‑to‑use linear MPC wrapper.
• SQPSolver  ➜ first‑order SQP with pluggable QP back‑ends.
• Prototype NMPC module built on multiple shooting and the SQP core.
• Extensive unit tests (HS071, constrained Rosenbrock, drone attitude control, quadrotor MPC …)

Key ideas

• Header‑only, no macros – just add the include/ folder to your project.
• AD everywhere – gradients, Jacobians and Hessians are generated on‑the‑fly from the same expressions that define your problem.
• Dense or sparse – flip a template parameter and every matrix adapts; sparsity patterns are cached once.
• Unified NLP façade – write cost_impl, equality_constraints_impl, inequality_constraints_impl; the base class synthesises the 15 routines Ipopt (or you) may ask for.
• Real‑time focus – every allocation is decided at compile time; run‑time heap hits are avoided completely in dense mode.

How it fits together

user code (struct MyOCP : ProblemBase<…>)   // supply cost & constraints
│
├──▶ IpoptInterface<MyOCP>          // Newton‑type, second order (new)
│
├──▶ SQPSolver<MyOCP>               // first‑order SQP
│
├──▶ MpcSolver                      // linear MPC (internally QP)
│
└──▶ NMPC::MultipleShooting         // nonlinear MPC (experimental)

All solvers accept Eigen vectors as initial guesses and expose the final primal/dual solution without copies.

Typical call‑site

MyOCP ocp;
IpoptInterface<MyOCP> nlp;

nlp.lower_bound_x() << …;          // box constraints
nlp.parameters()     << …;          // static params
nlp.solve(/*x₀*/ VectorXd::Zero(MyOCP::VAR_SIZE),
          /*λ₀*/ VectorXd::Zero(MyOCP::DUAL_SIZE));

std::cout << "Optimal cost  = " << nlp.cost()            << '\n'
          << "‖g‖∞          = " << nlp.constr_violation() << '\n';

Switch to SQP by replacing the two middle lines:

SQPSolver<MyOCP> sqp;
sqp.solve(x0, VectorXd::Zero(MyOCP::NE),
          x_min, x_max, g_min, g_max);

Mathematics

Given state $\mathbf x\in\mathbb R^{n}$, control $\mathbf u\in\mathbb R^{m}$ and parameters $\mathbf p$, the toolkit assumes the generic nonlinear program

All partial derivatives are produced automatically.

Need the code? Browse the repository on GitHub – pull requests and issues are very welcome!