After several iterations, and with update only, lower/upper bounds appear swapped

[fp018]

Description

I’m using OsqpEigen to solve a QP repeatedly in a loop. The solver is initialized once and, at each iteration, I update only the constraint matrix A and the bounds (lb, ub).

The solver fails with:

ERROR in validate_data: Lower bound at index 0 is greater than upper bound: 1.0000e+00 > -1.0000e+00
ERROR in osqp_setup: Problem data validation.
[OsqpEigen::Solver::initSolver] Unable to setup the workspace.
[OsqpEigen::Solver::updateLinearConstraintsMatrix] Unable to Initialize the solver.

These numbers match the values I pass to updateBounds, but they appear swapped: the intended lower bound should be -1 and the upper bound 1.
This happens at the second iteration or after several iterations, depending on the problem formulation.

The periodical update is

Eigen::Matrix<c_float, -1, 1> ub_f = ub.cast<c_float>();
Eigen::Matrix<c_float, -1, 1> lb_f = lb.cast<c_float>();

if (!solver.updateLinearConstraintsMatrix(A_s)) return false;
if (!solver.updateBounds(lb_f, ub_f)) return false;

Environment:

OsqpEigen version: v0.10.3

OSQP version: v1.0.0

They are both compiled from source code, and OSQP with -DOSQP_USE_FLOAT=ON. I have used the same approach before with an older version of the library, and it worked fine.

Thanks in advance for any help or suggestions!

[S-Dafarra]

Hi @fp018, thanks for opening the issue!

At a first sight, I am not able to spot anything wrong. The updateBounds code is

osqp-eigen/src/Solver.cpp

Lines 366 to 408 in ae2b631

	bool OsqpEigen::Solver::updateBounds(
	const Eigen::Ref<const Eigen::Matrix<c_float, Eigen::Dynamic, 1>>& lowerBound,
	const Eigen::Ref<const Eigen::Matrix<c_float, Eigen::Dynamic, 1>>& upperBound)
	{
	if (!m_isSolverInitialized)
	{
	debugStream() << "[OsqpEigen::Solver::updateBounds] The solver is not initialized"
	<< std::endl;
	return false;
	}
	// check if the dimension of the upperBound vector is correct
	if (upperBound.rows() != getData()->m)
	{
	debugStream() << "[OsqpEigen::Solver::updateBounds] The size of the upper bound must be "
	"equal to the number of the variables."
	<< std::endl;
	return false;
	}
	// check if the dimension of the lowerBound vector is correct
	if (lowerBound.rows() != getData()->m)
	{
	debugStream() << "[OsqpEigen::Solver::updateBounds] The size of the lower bound must be "
	"equal to the number of the variables."
	<< std::endl;
	return false;
	}
	// update lower and upper constraints
	#ifdef OSQP_EIGEN_OSQP_IS_V1
	if (osqp_update_data_vec(m_solver.get(), nullptr, lowerBound.data(), upperBound.data()))
	{
	#else
	if (osqp_update_bounds(m_workspace.get(), lowerBound.data(), upperBound.data()))
	{
	#endif
	debugStream() << "[OsqpEigen::Solver::updateBounds] Error when the update bounds is called."
	<< std::endl;
	return false;
	}
	return true;
	}

which calls https://osqp.org/docs/interfaces/C.html#_CPPv420osqp_update_data_vecP10OSQPSolverPK9OSQPFloatPK9OSQPFloatPK9OSQPFloat

We are also testing this in

osqp-eigen/tests/MPCUpdateMatricesTest.cpp

Line 292 in ae2b631

REQUIRE(solver.updateBounds(lowerBound, upperBound));

Are you able to provide a "minimum reproducible example" for us to try and reproduce the issue?

[fp018]

Thanks for the quick response!

After some attempts, I traced back the problem to a bad memory allocation.

Right now, I am solving the QP in a function executed periodically with a ROS timer, and if I use Eigen::VectorXd, for lower and upper bounds, the solver fails after some iterations, while with constant-sized Eigen vectors, it seems to work.

[S-Dafarra]

osqp-eigen internally stores a raw pointer to the vector data. I suppose that if the vector changes size, or the pointer to the raw data gets invalid, then there might be issues. Do you have a simple of example leading to a failure?