pettni / smooth_feedback Goto Github PK

• Scaling algorithm is slightly different from osqp paper, should compare the two alternatives
• Adaptive rho is not implemented

To comply with coding standards that prohibit dynamic memory allocation.

QP:

• Split work allocation and solving
• Expose solving with pre-allocated workspace

MPC:

• Split allocation and solving
• Allocate on construction

ASIF:

• Split allocation and solving
• Allocate on construction

• Implement P(I)D
• Update readme
• Add example with filter + (MPC / PID) + ASIF
  • Pub to ROS and create gif for repo
• Use asserts instead of exceptions
• cmake file for lapack
• Eigen 3.4 in CI
• Clean up benchmarks
  • one pipeline for problems, solve using three methods (osqp/sparse/dense)
  • add benchmark pic in readme
• Template MPC on differentiation type
• Don't require autodiff

Problem: Need to differentiate w.r.t. time. Not possible when time is a chrono type

Alternative: At start take in t0 (chrono) and function (double -> X). At call t (chrono) evaluate at (t - t0).

The intel math kernel library (package libmkl-dev) contains both dense and sparse LDLT decompositions.

The dense version is a LAPACK implementation (dsytrf), must figure out how to link against it.

/usr/include/mkl/mkl_lapack.h

The sparse version can be interfaced with through the <Eigen/PardisoSupport> header.

/usr/include/mkl/mkl_pardiso.h

In use:
Dense:

• lapacke
  Sparse:
• custom

Available:

Dense:

• MKL pardiso dsytrs (might be used automatically if MKL is linked..)
• Eigen::LDLT seems like it would not cause problems as long as a zero is not encountered..
  Sparse:
• MKL pardiso LDLT via <Eigen/ParsidoSupport> supports indefinite matrices. This will likely be faster