Page MenuHomec4science
Files F116546038

kiteNMPF.cpp
No OneTemporary
Actions

File Metadata

Created
Sat, Jun 7, 12:29

kiteNMPF.cpp
View Options

#include "kiteNMPF.h"
#include "utility"
#include "pseudospectral/chebyshev.hpp"
using namespace casadi;
const DM KiteNMPF::DEFAULT_LBG = -DM::inf(1);
const DM KiteNMPF::DEFAULT_UBG = DM::inf(1);
const DM KiteNMPF::DEFAULT_LBX = -DM::inf(15);
const DM KiteNMPF::DEFAULT_UBX = DM::inf(15);
const DM KiteNMPF::DEFAULT_LBU = -DM::inf(4);
const DM KiteNMPF::DEFAULT_UBU = DM::inf(4);
KiteNMPF::KiteNMPF(std::shared_ptr<KiteDynamics> _Kite, const Function &_Path) : Kite(std::move(_Kite))
{
PathFunc = _Path;
/** set up default values */
LBX = DEFAULT_LBX;
UBX = DEFAULT_UBX;
LBU = DEFAULT_LBU;
UBU = DEFAULT_UBU;
LBG = DEFAULT_LBG;
UBG = DEFAULT_UBG;
Q = 1e-2 * SX::diag(SX({1e4, 1e4, 5e3}));
R = 1e-4 * SX::diag(SX({1, 1, 1, 1}));
W = 1e-3;
Scale_X = DM::eye(15); invSX = DM::eye(15);
Scale_U = DM::eye(4); invSU = DM::eye(4);
NUM_SHOOTING_INTERVALS = 10;
/** @attention : need sensible reference velocity */
DM vel_ref = 0.05;
this->setReferenceVelocity(vel_ref);
WARM_START = false;
_initialized = false;
}
void KiteNMPF::createNLP()
{
/** get dynamics function and state Jacobian */
SX dynamics = Kite->getSymbolicDynamics();
SX X = Kite->getSymbolicState();
SX U = Kite->getSymbolicControl();
/** state and control dimensionality */
int n = 15;
int m = 4;
/** Order of polynomial interpolation */
int N = NUM_SHOOTING_INTERVALS;
/** define augmented dynamics of path parameter */
SX V = SX::sym("V", 2);
SX Uv = SX::sym("Uv");
SX Av = SX::zeros(2,2); Av(0,1) = 1;
SX Bv = SX::zeros(2,1); Bv(1,0) = 1;
/** parameter dynamics */
SX p_dynamics = SX::mtimes(Av, V) + SX::mtimes(Bv, Uv);
/** augmented system */
SX aug_state = SX::vertcat({X, V});
SX aug_control = SX::vertcat({U, Uv});
SX aug_dynamics = SX::vertcat({dynamics, p_dynamics});
/** evaluate augmented dynamics */
Function aug_dynamo = Function("AUG_DYNAMO", {aug_state, aug_control}, {aug_dynamics});
DynamicsFunc = aug_dynamo;
/** ----------------------------------------------------------------------------------*/
const int num_segments = 1;
const int poly_order = 10;
const int dimx = 15;
const int dimu = 4;
const int dimp = 1;
Chebyshev<SX, poly_order, num_segments, dimx, dimu, dimp> spectral;
SX diff_constr = spectral.CollocateDynamics(DynamicsFunc, 0, 1);
//diff_constr = diff_constr(Slice(0, diff_constr.size1() - dimx));
/** define an integral cost */
SX x = SX::sym("x", dimx);
SX u = SX::sym("u", dimu);
SXVector tmp = PathFunc(SXVector{x[13]});
SX sym_path = tmp[0];
SX residual = sym_path - x(Slice(6,9));
SX lagrange = SX::sumRows( SX::mtimes(Q, pow(residual, 2)) ) + SX::sumRows( SX::mtimes(W, pow(reference_velocity - x[14], 2)) );
Function LagrangeTerm = Function("Lagrange", {x, u}, {lagrange});
/** trace functions */
PathError = Function("PathError", {x}, {residual});
VelError = Function("VelError", {x}, {reference_velocity - x[14]});
SX mayer = SX::sumRows( SX::mtimes(2 * Q, pow(residual, 2)) );
Function MayerTerm = Function("Mayer",{x}, {mayer});
SX performance_idx = spectral.CollocateCost(MayerTerm, LagrangeTerm, 0, 1);
SX varx = spectral.VarX();
SX varu = spectral.VarU();
SX opt_var = SX::vertcat(SXVector{varx, varu});
SX lbg = SX::zeros(diff_constr.size());
SX ubg = SX::zeros(diff_constr.size());
/** set inequality (box) constraints */
/** state */
SX lbx = SX::repmat(SX::mtimes(Scale_X, LBX), poly_order + 1, 1);
SX ubx = SX::repmat(SX::mtimes(Scale_X, UBX), poly_order + 1, 1);
/** control */
//lbx = SX::vertcat( {lbx, SX::repmat(SX::mtimes(Scale_U, LBU), poly_order, 1)} );
//ubx = SX::vertcat( {ubx, SX::repmat(SX::mtimes(Scale_U, UBU), poly_order, 1)} );
lbx = SX::vertcat( {lbx, SX::repmat(SX::mtimes(Scale_U, LBU), poly_order + 1, 1)} );
ubx = SX::vertcat( {ubx, SX::repmat(SX::mtimes(Scale_U, UBU), poly_order + 1, 1)} );
/** formulate NLP */
NLP["x"] = opt_var;
NLP["f"] = performance_idx;
NLP["g"] = diff_constr;
OPTS["ipopt.linear_solver"] = "ma97";
OPTS["ipopt.print_level"] = 0;
OPTS["ipopt.tol"] = 1e-5;
OPTS["ipopt.acceptable_tol"] = 1e-4;
//OPTS["ipopt.max_iter"] = 15;
NLP_Solver = nlpsol("solver", "ipopt", NLP, OPTS);
/** set default args */
ARG["lbx"] = lbx;
ARG["ubx"] = ubx;
ARG["lbg"] = lbg;
ARG["ubg"] = ubg;
DM feasible_state = DM::mtimes(Scale_X, (UBX + LBX) / 2);
DM feasible_control = DM::mtimes(Scale_U, (UBU + LBU) / 2);
ARG["x0"] = DM::vertcat(DMVector{DM::repmat(feasible_state, poly_order + 1, 1),
DM::repmat(feasible_control, poly_order + 1, 1)});
//DM::repmat(feasible_control, poly_order, 1)});
/** ----------------------------------------------------------------------------------*/
/** NLP formulation */
/** define optimisation variables */
/** SX z = SX::sym("z", n, N+1);
SX z_u = SX::sym("z_u", m, N); */
/** allocate vectors to store constraints */
/** SX vecx = {};
SX g = {}; */
/** state and control constraints */
/** SX lbx = {};
SX ubx = {}; */
/** nonlinear state constraints */
/** SX lbg = {};
SX ubg = {}; */
/** objective function */
/** SX objective = 0; */
/** Geometric path definition */
/** SX theta = SX::sym("theta"); */
/** NLP formulation */
/** Lagrangian term */
/** SX pos = z(Slice(6,9), Slice(1, z.size2()-1));
pos = SX::vec(pos); */
/** path following error */
/** DM ScaleXYZ = Scale_X(Slice(6,9), Slice(6,9));
SX path_scaled = SX::mtimes(ScaleXYZ, SX::vertcat(PathFunc(SXVector{theta})));
Function PathFuncScaled = Function("Scaled_Path",{theta},{path_scaled}); */
/** @badcode: better DM to double conversion should exist */
/** double scale_th = Scale_X(13,13).nonzeros()[0];
SX path = kmath::mat_func( scale_th * z(13, Slice(1, z.size2()-1)), PathFuncScaled);
SX err = pos - path;
SX Qn = SX::kron(SX::eye(N-1), Q); */
/** reference velocity following */
/** SX err_v = z(14, Slice(1, z.size2()-1)).T() - SX::repmat(reference_velocity, N-1, 1);
SX Qw = W * SX::eye(N-1);
SX L = SX::sumRows( SX::mtimes(Qn, pow(err, 2)) ) + SX::sumRows( SX::mtimes(Qw, pow(err_v, 2)) ); */
/** Mayer term */
/** path = SX::vertcat(PathFuncScaled(SXVector{scale_th * z(13,0)}));
SX err_n = z(Slice(6, 9), 0) - path;
err_n = SX::vertcat( SXVector{err_n, z(14, 0) - reference_velocity});
SX T = SX::mtimes((10 * SX::diagcat( {Q, W}) ), pow(err_n, 2)); */
/** objective function */
/** objective = L + SX::sumRows(T);
CostFunction = Function("Cost", {z, z_u}, {objective});
/** set equality constraints */
/** differentiation matrix */
/** DM _x, _D;
std::pair<double, double> interval = std::make_pair<double, double>(0,1);
kmath::cheb(_x, _D, N, interval);
SX D = _D;
D(D.size1()-1, Slice(0, D.size2())) = SX::zeros(N+1);
D(D.size1() - 1, D.size2() - 1) = 1;
SX Dn = SX::kron(D, SX::eye(n));
SX iSx = SX::kron(SX::eye(N+1), invSX);
SX iSu = SX::kron(SX::eye(N), invSU);
SX F = kmath::mat_dynamics(SX::mtimes(invSX, z), SX::mtimes(invSU, z_u), aug_dynamo);
SX G = SX::mtimes(SX::mtimes(Dn, iSx), SX::vec(z)) - F;
G = G(Slice(0, N * n), 0);
DynamicConstraints = Function("lox",{z, z_u},{G});
lbg = SX::zeros(( N ) * n, 1);
ubg = SX::zeros(( N ) * n, 1);
/** set inequality (box) constraints */
/** state */
/** @todo: SCALE CONSTRAINTS */
/** lbx = SX::repmat(SX::mtimes(Scale_X, LBX), N + 1, 1);
ubx = SX::repmat(SX::mtimes(Scale_X, UBX), N + 1, 1);
/** control */
/** lbx = SX::vertcat( {lbx, SX::repmat(SX::mtimes(Scale_U, LBU), N, 1)} );
ubx = SX::vertcat( {ubx, SX::repmat(SX::mtimes(Scale_U, UBU), N, 1)} );
/** set decision variable */
/** vecx = SX::vertcat({SX::vec(z), SX::vec(z_u)});
/** formulate NLP */
/** NLP["x"] = vecx;
NLP["f"] = objective;
NLP["g"] = G;
OPTS["ipopt.linear_solver"] = "ma97";
OPTS["ipopt.print_level"] = 0;
OPTS["ipopt.tol"] = 1e-5;
OPTS["ipopt.acceptable_tol"] = 1e-4;
//OPTS["ipopt.max_iter"] = 15;
NLP_Solver = nlpsol("solver", "ipopt", NLP, OPTS);
/** set default args */
/** ARG["lbx"] = lbx;
ARG["ubx"] = ubx;
ARG["lbg"] = lbg;
ARG["ubg"] = ubg;
DM feasible_state = DM::mtimes(Scale_X, (UBX + LBX) / 2);
DM feasible_control = DM::mtimes(Scale_U, (UBU + LBU) / 2);
ARG["x0"] = DM::vertcat(DMVector{DM::repmat(feasible_state, N + 1, 1),
DM::repmat(feasible_control, N, 1)}); */
}
void KiteNMPF::computeControl(const DM &_X0)
{
int N = NUM_SHOOTING_INTERVALS;
/** @badcode : remove magic constants */
/** number of states */
int nx = 15;
/** number of control inputs */
int nu = 4;
/** rectify virtual state */
DM X0 = _X0;
bool rectify = false;
if(X0[13].nonzeros()[0] > 2 * M_PI)
{
X0[13] -= 2 * M_PI;
rectify = true;
}
else if (X0[13].nonzeros()[0] < -2 * M_PI)
{
X0[13] += 2 * M_PI;
rectify = true;
}
/** scale input */
X0 = DM::mtimes(Scale_X, X0);
if(WARM_START)
{
int idx_in = N * nx;
int idx_out = idx_in + nx;
ARG["lbx"](Slice(idx_in, idx_out), 0) = X0;
ARG["ubx"](Slice(idx_in, idx_out), 0) = X0;
/** rectify initial guess */
if(rectify)
{
for(int i = 0; i < (N + 1) * nx; i += nx)
{
int idx = i + 13;
if(NLP_X[idx].nonzeros()[0] > 2 * M_PI)
NLP_X[idx] -= 2 * M_PI;
else if (NLP_X[idx].nonzeros()[0] < -2 * M_PI)
NLP_X[idx] += 2 * M_PI;
}
}
ARG["x0"] = NLP_X;
}
else
{
ARG["x0"](Slice(0, (N + 1) * nx), 0) = DM::repmat(X0, (N + 1), 1);
int idx_in = N * nx;
int idx_out = idx_in + nx;
ARG["lbx"](Slice(idx_in, idx_out), 0) = X0;
ARG["ubx"](Slice(idx_in, idx_out), 0) = X0;
}
/** store optimal solution */
DMDict res = NLP_Solver(ARG);
NLP_X = res.at("x");
DM opt_x = NLP_X(Slice(0, (N + 1) * nx ));
DM invSX = DM::solve(Scale_X, DM::eye(15));
OptimalTrajectory = DM::mtimes(invSX, DM::reshape(opt_x, nx, N + 1));
DM opt_u = NLP_X( Slice((N + 1) * nx, NLP_X.size1()) );
DM invSU = DM::solve(Scale_U, DM::eye(4));
OptimalControl = DM::mtimes(invSU, DM::reshape(opt_u, nu, N + 1));
stats = NLP_Solver.stats();
std::cout << stats << "\n";
enableWarmStart();
}
/** get path error */
double KiteNMPF::getPathError()
{
double error = 0;
if(!OptimalTrajectory.is_empty())
{
DM state = OptimalTrajectory(Slice(0, OptimalTrajectory.size1()), OptimalTrajectory.size2() - 1);
DMVector tmp = PathError(DMVector{state});
error = DM::norm_2( tmp[0] ).nonzeros()[0];
}
return error;
}
/** get velocity error */
double KiteNMPF::getVelocityError()
{
double error = 0;
if(!OptimalTrajectory.is_empty())
{
DM state = OptimalTrajectory(Slice(0, OptimalTrajectory.size1()), OptimalTrajectory.size2() - 1);
DMVector tmp = VelError(DMVector{state});
error = DM::norm_2( tmp[0] ).nonzeros()[0];
}
return error;
}
/** get virtual state */
double KiteNMPF::getVirtState()
{
double virt_state = 0;
if(!OptimalTrajectory.is_empty())
{
virt_state = OptimalTrajectory(13, OptimalTrajectory.size2() - 1).nonzeros()[0];
}
return virt_state;
}
/** compute intial guess for virtual state */
DM KiteNMPF::findClosestPointOnPath(const DM &position, const DM &init_guess)
{
SX theta = SX::sym("theta");
SX sym_residual = 0.5 * SX::norm_2(PathFunc(SXVector{theta})[0] - SX(position));
SX sym_gradient = SX::gradient(sym_residual,theta);
Function grad = Function("gradient", {theta}, {sym_gradient});
double tol = 1e-2;
int counter = 0;
DM theta_i = init_guess;
DMVector result = grad(DMVector{theta_i});
double step = 0.25;
/** check for local minimum */
if(fabs(result[0].nonzeros()[0]) < tol)
{
theta_i = {M_PI_2 + 0.1};
result = grad(DMVector{theta_i});
}
/** @badcode : shitty code */
while (fabs(result[0].nonzeros()[0]) >= tol)
{
counter++;
theta_i -= step * grad(theta_i);
/** gradient update */
result = grad(DMVector{theta_i});
if(counter > 10)
break;
}
std::cout << theta_i << "\n";
return theta_i;
}

Event Timeline

c4science · Help