| Line |
Branch |
Exec |
Source |
| 1 |
|
|
/////////////////////////////////////////////////////////////////////////////// |
| 2 |
|
|
// BSD 3-Clause License |
| 3 |
|
|
// |
| 4 |
|
|
// Copyright (C) 2019-2022, LAAS-CNRS, University of Edinburgh, |
| 5 |
|
|
// University of Oxford, Heriot-Watt University |
| 6 |
|
|
// Copyright note valid unless otherwise stated in individual files. |
| 7 |
|
|
// All rights reserved. |
| 8 |
|
|
/////////////////////////////////////////////////////////////////////////////// |
| 9 |
|
|
|
| 10 |
|
|
#include "crocoddyl/core/solvers/ddp.hpp" |
| 11 |
|
|
|
| 12 |
|
|
namespace crocoddyl { |
| 13 |
|
|
|
| 14 |
|
✗ |
SolverDDP::SolverDDP(std::shared_ptr<ShootingProblem> problem) |
| 15 |
|
|
: SolverAbstract(problem), |
| 16 |
|
✗ |
reg_incfactor_(10.), |
| 17 |
|
✗ |
reg_decfactor_(10.), |
| 18 |
|
✗ |
reg_min_(1e-9), |
| 19 |
|
✗ |
reg_max_(1e9), |
| 20 |
|
✗ |
cost_try_(0.), |
| 21 |
|
✗ |
th_grad_(1e-12), |
| 22 |
|
✗ |
th_stepdec_(0.5), |
| 23 |
|
✗ |
th_stepinc_(0.01) { |
| 24 |
|
✗ |
allocateData(); |
| 25 |
|
|
|
| 26 |
|
✗ |
const std::size_t n_alphas = 10; |
| 27 |
|
✗ |
alphas_.resize(n_alphas); |
| 28 |
|
✗ |
for (std::size_t n = 0; n < n_alphas; ++n) { |
| 29 |
|
✗ |
alphas_[n] = 1. / pow(2., static_cast<double>(n)); |
| 30 |
|
|
} |
| 31 |
|
✗ |
if (th_stepinc_ < alphas_[n_alphas - 1]) { |
| 32 |
|
✗ |
th_stepinc_ = alphas_[n_alphas - 1]; |
| 33 |
|
|
std::cerr << "Warning: th_stepinc has higher value than lowest alpha " |
| 34 |
|
|
"value, set to " |
| 35 |
|
✗ |
<< std::to_string(alphas_[n_alphas - 1]) << std::endl; |
| 36 |
|
|
} |
| 37 |
|
|
} |
| 38 |
|
|
|
| 39 |
|
✗ |
SolverDDP::~SolverDDP() {} |
| 40 |
|
|
|
| 41 |
|
✗ |
bool SolverDDP::solve(const std::vector<Eigen::VectorXd>& init_xs, |
| 42 |
|
|
const std::vector<Eigen::VectorXd>& init_us, |
| 43 |
|
|
const std::size_t maxiter, const bool is_feasible, |
| 44 |
|
|
const double init_reg) { |
| 45 |
|
✗ |
START_PROFILER("SolverDDP::solve"); |
| 46 |
|
✗ |
if (problem_->is_updated()) { |
| 47 |
|
✗ |
resizeData(); |
| 48 |
|
|
} |
| 49 |
|
✗ |
xs_try_[0] = |
| 50 |
|
✗ |
problem_->get_x0(); // it is needed in case that init_xs[0] is infeasible |
| 51 |
|
✗ |
setCandidate(init_xs, init_us, is_feasible); |
| 52 |
|
|
|
| 53 |
|
✗ |
if (std::isnan(init_reg)) { |
| 54 |
|
✗ |
preg_ = reg_min_; |
| 55 |
|
✗ |
dreg_ = reg_min_; |
| 56 |
|
|
} else { |
| 57 |
|
✗ |
preg_ = init_reg; |
| 58 |
|
✗ |
dreg_ = init_reg; |
| 59 |
|
|
} |
| 60 |
|
✗ |
was_feasible_ = false; |
| 61 |
|
|
|
| 62 |
|
✗ |
bool recalcDiff = true; |
| 63 |
|
✗ |
for (iter_ = 0; iter_ < maxiter; ++iter_) { |
| 64 |
|
|
while (true) { |
| 65 |
|
|
try { |
| 66 |
|
✗ |
computeDirection(recalcDiff); |
| 67 |
|
✗ |
} catch (std::exception& e) { |
| 68 |
|
✗ |
recalcDiff = false; |
| 69 |
|
✗ |
increaseRegularization(); |
| 70 |
|
✗ |
if (preg_ == reg_max_) { |
| 71 |
|
✗ |
return false; |
| 72 |
|
|
} else { |
| 73 |
|
✗ |
continue; |
| 74 |
|
|
} |
| 75 |
|
|
} |
| 76 |
|
✗ |
break; |
| 77 |
|
|
} |
| 78 |
|
✗ |
expectedImprovement(); |
| 79 |
|
|
|
| 80 |
|
|
// We need to recalculate the derivatives when the step length passes |
| 81 |
|
✗ |
recalcDiff = false; |
| 82 |
|
✗ |
for (std::vector<double>::const_iterator it = alphas_.begin(); |
| 83 |
|
✗ |
it != alphas_.end(); ++it) { |
| 84 |
|
✗ |
steplength_ = *it; |
| 85 |
|
|
|
| 86 |
|
|
try { |
| 87 |
|
✗ |
dV_ = tryStep(steplength_); |
| 88 |
|
✗ |
} catch (std::exception& e) { |
| 89 |
|
✗ |
continue; |
| 90 |
|
|
} |
| 91 |
|
✗ |
dVexp_ = steplength_ * (d_[0] + 0.5 * steplength_ * d_[1]); |
| 92 |
|
|
|
| 93 |
|
✗ |
if (dVexp_ >= 0) { // descend direction |
| 94 |
|
✗ |
if (std::abs(d_[0]) < th_grad_ || !is_feasible_ || |
| 95 |
|
✗ |
dV_ > th_acceptstep_ * dVexp_) { |
| 96 |
|
✗ |
was_feasible_ = is_feasible_; |
| 97 |
|
✗ |
setCandidate(xs_try_, us_try_, true); |
| 98 |
|
✗ |
cost_ = cost_try_; |
| 99 |
|
✗ |
recalcDiff = true; |
| 100 |
|
✗ |
break; |
| 101 |
|
|
} |
| 102 |
|
|
} |
| 103 |
|
|
} |
| 104 |
|
|
|
| 105 |
|
✗ |
if (steplength_ > th_stepdec_) { |
| 106 |
|
✗ |
decreaseRegularization(); |
| 107 |
|
|
} |
| 108 |
|
✗ |
if (steplength_ <= th_stepinc_) { |
| 109 |
|
✗ |
increaseRegularization(); |
| 110 |
|
✗ |
if (preg_ == reg_max_) { |
| 111 |
|
✗ |
STOP_PROFILER("SolverDDP::solve"); |
| 112 |
|
✗ |
return false; |
| 113 |
|
|
} |
| 114 |
|
|
} |
| 115 |
|
✗ |
stoppingCriteria(); |
| 116 |
|
|
|
| 117 |
|
✗ |
const std::size_t n_callbacks = callbacks_.size(); |
| 118 |
|
✗ |
for (std::size_t c = 0; c < n_callbacks; ++c) { |
| 119 |
|
✗ |
CallbackAbstract& callback = *callbacks_[c]; |
| 120 |
|
✗ |
callback(*this); |
| 121 |
|
|
} |
| 122 |
|
|
|
| 123 |
|
✗ |
if (was_feasible_ && stop_ < th_stop_) { |
| 124 |
|
✗ |
STOP_PROFILER("SolverDDP::solve"); |
| 125 |
|
✗ |
return true; |
| 126 |
|
|
} |
| 127 |
|
|
} |
| 128 |
|
✗ |
STOP_PROFILER("SolverDDP::solve"); |
| 129 |
|
✗ |
return false; |
| 130 |
|
|
} |
| 131 |
|
|
|
| 132 |
|
✗ |
void SolverDDP::computeDirection(const bool recalcDiff) { |
| 133 |
|
✗ |
START_PROFILER("SolverDDP::computeDirection"); |
| 134 |
|
✗ |
if (recalcDiff) { |
| 135 |
|
✗ |
calcDiff(); |
| 136 |
|
|
} |
| 137 |
|
✗ |
backwardPass(); |
| 138 |
|
✗ |
STOP_PROFILER("SolverDDP::computeDirection"); |
| 139 |
|
|
} |
| 140 |
|
|
|
| 141 |
|
✗ |
double SolverDDP::tryStep(const double steplength) { |
| 142 |
|
✗ |
START_PROFILER("SolverDDP::tryStep"); |
| 143 |
|
✗ |
forwardPass(steplength); |
| 144 |
|
✗ |
STOP_PROFILER("SolverDDP::tryStep"); |
| 145 |
|
✗ |
return cost_ - cost_try_; |
| 146 |
|
|
} |
| 147 |
|
|
|
| 148 |
|
✗ |
double SolverDDP::stoppingCriteria() { |
| 149 |
|
|
// This stopping criteria represents the expected reduction in the value |
| 150 |
|
|
// function. If this reduction is less than a certain threshold, then the |
| 151 |
|
|
// algorithm reaches the local minimum. For more details, see C. Mastalli et |
| 152 |
|
|
// al. "Inverse-dynamics MPC via Nullspace Resolution". |
| 153 |
|
✗ |
stop_ = std::abs(d_[0] + 0.5 * d_[1]); |
| 154 |
|
✗ |
return stop_; |
| 155 |
|
|
} |
| 156 |
|
|
|
| 157 |
|
✗ |
const Eigen::Vector2d& SolverDDP::expectedImprovement() { |
| 158 |
|
✗ |
d_.fill(0); |
| 159 |
|
✗ |
const std::size_t T = this->problem_->get_T(); |
| 160 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 161 |
|
✗ |
problem_->get_runningModels(); |
| 162 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 163 |
|
✗ |
const std::size_t nu = models[t]->get_nu(); |
| 164 |
|
✗ |
if (nu != 0) { |
| 165 |
|
✗ |
d_[0] += Qu_[t].dot(k_[t]); |
| 166 |
|
✗ |
d_[1] -= k_[t].dot(Quuk_[t]); |
| 167 |
|
|
} |
| 168 |
|
|
} |
| 169 |
|
✗ |
return d_; |
| 170 |
|
|
} |
| 171 |
|
|
|
| 172 |
|
✗ |
void SolverDDP::resizeData() { |
| 173 |
|
✗ |
START_PROFILER("SolverDDP::resizeData"); |
| 174 |
|
✗ |
SolverAbstract::resizeData(); |
| 175 |
|
|
|
| 176 |
|
✗ |
const std::size_t T = problem_->get_T(); |
| 177 |
|
✗ |
const std::size_t ndx = problem_->get_ndx(); |
| 178 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 179 |
|
✗ |
problem_->get_runningModels(); |
| 180 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 181 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& model = models[t]; |
| 182 |
|
✗ |
const std::size_t nu = model->get_nu(); |
| 183 |
|
✗ |
Qxu_[t].conservativeResize(ndx, nu); |
| 184 |
|
✗ |
Quu_[t].conservativeResize(nu, nu); |
| 185 |
|
✗ |
Qu_[t].conservativeResize(nu); |
| 186 |
|
✗ |
K_[t].conservativeResize(nu, ndx); |
| 187 |
|
✗ |
k_[t].conservativeResize(nu); |
| 188 |
|
✗ |
us_try_[t].conservativeResize(nu); |
| 189 |
|
✗ |
FuTVxx_p_[t].conservativeResize(nu, ndx); |
| 190 |
|
✗ |
Quuk_[t].conservativeResize(nu); |
| 191 |
|
✗ |
if (nu != 0) { |
| 192 |
|
✗ |
FuTVxx_p_[t].setZero(); |
| 193 |
|
|
} |
| 194 |
|
|
} |
| 195 |
|
✗ |
STOP_PROFILER("SolverDDP::resizeData"); |
| 196 |
|
|
} |
| 197 |
|
|
|
| 198 |
|
✗ |
double SolverDDP::calcDiff() { |
| 199 |
|
✗ |
START_PROFILER("SolverDDP::calcDiff"); |
| 200 |
|
✗ |
if (iter_ == 0) { |
| 201 |
|
✗ |
problem_->calc(xs_, us_); |
| 202 |
|
|
} |
| 203 |
|
✗ |
cost_ = problem_->calcDiff(xs_, us_); |
| 204 |
|
|
|
| 205 |
|
✗ |
ffeas_ = computeDynamicFeasibility(); |
| 206 |
|
✗ |
gfeas_ = computeInequalityFeasibility(); |
| 207 |
|
✗ |
hfeas_ = computeEqualityFeasibility(); |
| 208 |
|
✗ |
STOP_PROFILER("SolverDDP::calcDiff"); |
| 209 |
|
✗ |
return cost_; |
| 210 |
|
|
} |
| 211 |
|
|
|
| 212 |
|
✗ |
void SolverDDP::backwardPass() { |
| 213 |
|
✗ |
START_PROFILER("SolverDDP::backwardPass"); |
| 214 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d_T = problem_->get_terminalData(); |
| 215 |
|
✗ |
Vxx_.back() = d_T->Lxx; |
| 216 |
|
✗ |
Vx_.back() = d_T->Lx; |
| 217 |
|
|
|
| 218 |
|
✗ |
if (!std::isnan(preg_)) { |
| 219 |
|
✗ |
Vxx_.back().diagonal().array() += preg_; |
| 220 |
|
|
} |
| 221 |
|
|
|
| 222 |
|
✗ |
if (!is_feasible_) { |
| 223 |
|
✗ |
Vx_.back().noalias() += Vxx_.back() * fs_.back(); |
| 224 |
|
|
} |
| 225 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 226 |
|
✗ |
problem_->get_runningModels(); |
| 227 |
|
|
const std::vector<std::shared_ptr<ActionDataAbstract> >& datas = |
| 228 |
|
✗ |
problem_->get_runningDatas(); |
| 229 |
|
✗ |
for (int t = static_cast<int>(problem_->get_T()) - 1; t >= 0; --t) { |
| 230 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& m = models[t]; |
| 231 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = datas[t]; |
| 232 |
|
|
|
| 233 |
|
|
// Compute the linear-quadratic approximation of the control Hamiltonian |
| 234 |
|
|
// function |
| 235 |
|
✗ |
computeActionValueFunction(t, m, d); |
| 236 |
|
|
|
| 237 |
|
|
// Compute the feedforward and feedback gains |
| 238 |
|
✗ |
computeGains(t); |
| 239 |
|
|
|
| 240 |
|
|
// Compute the linear-quadratic approximation of the Value function |
| 241 |
|
✗ |
computeValueFunction(t, m); |
| 242 |
|
|
|
| 243 |
|
✗ |
if (raiseIfNaN(Vx_[t].lpNorm<Eigen::Infinity>())) { |
| 244 |
|
✗ |
throw_pretty("backward_error"); |
| 245 |
|
|
} |
| 246 |
|
✗ |
if (raiseIfNaN(Vxx_[t].lpNorm<Eigen::Infinity>())) { |
| 247 |
|
✗ |
throw_pretty("backward_error"); |
| 248 |
|
|
} |
| 249 |
|
|
} |
| 250 |
|
✗ |
STOP_PROFILER("SolverDDP::backwardPass"); |
| 251 |
|
|
} |
| 252 |
|
|
|
| 253 |
|
✗ |
void SolverDDP::forwardPass(const double steplength) { |
| 254 |
|
✗ |
if (steplength > 1. || steplength < 0.) { |
| 255 |
|
✗ |
throw_pretty("Invalid argument: " |
| 256 |
|
|
<< "invalid step length, value is between 0. to 1."); |
| 257 |
|
|
} |
| 258 |
|
✗ |
START_PROFILER("SolverDDP::forwardPass"); |
| 259 |
|
✗ |
cost_try_ = 0.; |
| 260 |
|
✗ |
const std::size_t T = problem_->get_T(); |
| 261 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 262 |
|
✗ |
problem_->get_runningModels(); |
| 263 |
|
|
const std::vector<std::shared_ptr<ActionDataAbstract> >& datas = |
| 264 |
|
✗ |
problem_->get_runningDatas(); |
| 265 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 266 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& m = models[t]; |
| 267 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = datas[t]; |
| 268 |
|
|
|
| 269 |
|
✗ |
m->get_state()->diff(xs_[t], xs_try_[t], dx_[t]); |
| 270 |
|
✗ |
if (m->get_nu() != 0) { |
| 271 |
|
✗ |
us_try_[t].noalias() = us_[t]; |
| 272 |
|
✗ |
us_try_[t].noalias() -= k_[t] * steplength; |
| 273 |
|
✗ |
us_try_[t].noalias() -= K_[t] * dx_[t]; |
| 274 |
|
✗ |
m->calc(d, xs_try_[t], us_try_[t]); |
| 275 |
|
|
} else { |
| 276 |
|
✗ |
m->calc(d, xs_try_[t]); |
| 277 |
|
|
} |
| 278 |
|
✗ |
xs_try_[t + 1] = d->xnext; |
| 279 |
|
✗ |
cost_try_ += d->cost; |
| 280 |
|
|
|
| 281 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 282 |
|
✗ |
STOP_PROFILER("SolverDDP::forwardPass"); |
| 283 |
|
✗ |
throw_pretty("forward_error"); |
| 284 |
|
|
} |
| 285 |
|
✗ |
if (raiseIfNaN(xs_try_[t + 1].lpNorm<Eigen::Infinity>())) { |
| 286 |
|
✗ |
STOP_PROFILER("SolverDDP::forwardPass"); |
| 287 |
|
✗ |
throw_pretty("forward_error"); |
| 288 |
|
|
} |
| 289 |
|
|
} |
| 290 |
|
|
|
| 291 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& m = problem_->get_terminalModel(); |
| 292 |
|
✗ |
const std::shared_ptr<ActionDataAbstract>& d = problem_->get_terminalData(); |
| 293 |
|
✗ |
m->calc(d, xs_try_.back()); |
| 294 |
|
✗ |
cost_try_ += d->cost; |
| 295 |
|
|
|
| 296 |
|
✗ |
if (raiseIfNaN(cost_try_)) { |
| 297 |
|
✗ |
STOP_PROFILER("SolverDDP::forwardPass"); |
| 298 |
|
✗ |
throw_pretty("forward_error"); |
| 299 |
|
|
} |
| 300 |
|
✗ |
STOP_PROFILER("SolverDDP::forwardPass"); |
| 301 |
|
|
} |
| 302 |
|
|
|
| 303 |
|
✗ |
void SolverDDP::computeActionValueFunction( |
| 304 |
|
|
const std::size_t t, const std::shared_ptr<ActionModelAbstract>& model, |
| 305 |
|
|
const std::shared_ptr<ActionDataAbstract>& data) { |
| 306 |
|
✗ |
assert_pretty(t < problem_->get_T(), |
| 307 |
|
|
"Invalid argument: t should be between 0 and " + |
| 308 |
|
|
std::to_string(problem_->get_T());); |
| 309 |
|
✗ |
const std::size_t nu = model->get_nu(); |
| 310 |
|
✗ |
const Eigen::MatrixXd& Vxx_p = Vxx_[t + 1]; |
| 311 |
|
✗ |
const Eigen::VectorXd& Vx_p = Vx_[t + 1]; |
| 312 |
|
|
|
| 313 |
|
✗ |
FxTVxx_p_.noalias() = data->Fx.transpose() * Vxx_p; |
| 314 |
|
✗ |
START_PROFILER("SolverDDP::Qx"); |
| 315 |
|
✗ |
Qx_[t] = data->Lx; |
| 316 |
|
✗ |
Qx_[t].noalias() += data->Fx.transpose() * Vx_p; |
| 317 |
|
✗ |
STOP_PROFILER("SolverDDP::Qx"); |
| 318 |
|
✗ |
START_PROFILER("SolverDDP::Qxx"); |
| 319 |
|
✗ |
Qxx_[t] = data->Lxx; |
| 320 |
|
✗ |
Qxx_[t].noalias() += FxTVxx_p_ * data->Fx; |
| 321 |
|
✗ |
STOP_PROFILER("SolverDDP::Qxx"); |
| 322 |
|
✗ |
if (nu != 0) { |
| 323 |
|
✗ |
FuTVxx_p_[t].noalias() = data->Fu.transpose() * Vxx_p; |
| 324 |
|
✗ |
START_PROFILER("SolverDDP::Qu"); |
| 325 |
|
✗ |
Qu_[t] = data->Lu; |
| 326 |
|
✗ |
Qu_[t].noalias() += data->Fu.transpose() * Vx_p; |
| 327 |
|
✗ |
STOP_PROFILER("SolverDDP::Qu"); |
| 328 |
|
✗ |
START_PROFILER("SolverDDP::Quu"); |
| 329 |
|
✗ |
Quu_[t] = data->Luu; |
| 330 |
|
✗ |
Quu_[t].noalias() += FuTVxx_p_[t] * data->Fu; |
| 331 |
|
✗ |
STOP_PROFILER("SolverDDP::Quu"); |
| 332 |
|
✗ |
START_PROFILER("SolverDDP::Qxu"); |
| 333 |
|
✗ |
Qxu_[t] = data->Lxu; |
| 334 |
|
✗ |
Qxu_[t].noalias() += FxTVxx_p_ * data->Fu; |
| 335 |
|
✗ |
STOP_PROFILER("SolverDDP::Qxu"); |
| 336 |
|
✗ |
if (!std::isnan(preg_)) { |
| 337 |
|
✗ |
Quu_[t].diagonal().array() += preg_; |
| 338 |
|
|
} |
| 339 |
|
|
} |
| 340 |
|
|
} |
| 341 |
|
|
|
| 342 |
|
✗ |
void SolverDDP::computeValueFunction( |
| 343 |
|
|
const std::size_t t, const std::shared_ptr<ActionModelAbstract>& model) { |
| 344 |
|
✗ |
assert_pretty(t < problem_->get_T(), |
| 345 |
|
|
"Invalid argument: t should be between 0 and " + |
| 346 |
|
|
std::to_string(problem_->get_T());); |
| 347 |
|
✗ |
const std::size_t nu = model->get_nu(); |
| 348 |
|
✗ |
Vx_[t] = Qx_[t]; |
| 349 |
|
✗ |
Vxx_[t] = Qxx_[t]; |
| 350 |
|
✗ |
if (nu != 0) { |
| 351 |
|
✗ |
START_PROFILER("SolverDDP::Vx"); |
| 352 |
|
✗ |
Quuk_[t].noalias() = Quu_[t] * k_[t]; |
| 353 |
|
✗ |
Vx_[t].noalias() -= K_[t].transpose() * Qu_[t]; |
| 354 |
|
✗ |
STOP_PROFILER("SolverDDP::Vx"); |
| 355 |
|
✗ |
START_PROFILER("SolverDDP::Vxx"); |
| 356 |
|
✗ |
Vxx_[t].noalias() -= Qxu_[t] * K_[t]; |
| 357 |
|
✗ |
STOP_PROFILER("SolverDDP::Vxx"); |
| 358 |
|
|
} |
| 359 |
|
✗ |
Vxx_tmp_ = 0.5 * (Vxx_[t] + Vxx_[t].transpose()); |
| 360 |
|
✗ |
Vxx_[t] = Vxx_tmp_; |
| 361 |
|
|
|
| 362 |
|
✗ |
if (!std::isnan(preg_)) { |
| 363 |
|
✗ |
Vxx_[t].diagonal().array() += preg_; |
| 364 |
|
|
} |
| 365 |
|
|
|
| 366 |
|
|
// Compute and store the Vx gradient at end of the interval (rollout state) |
| 367 |
|
✗ |
if (!is_feasible_) { |
| 368 |
|
✗ |
Vx_[t].noalias() += Vxx_[t] * fs_[t]; |
| 369 |
|
|
} |
| 370 |
|
|
} |
| 371 |
|
|
|
| 372 |
|
✗ |
void SolverDDP::computeGains(const std::size_t t) { |
| 373 |
|
✗ |
assert_pretty(t < problem_->get_T(), |
| 374 |
|
|
"Invalid argument: t should be between 0 and " + |
| 375 |
|
|
std::to_string(problem_->get_T())); |
| 376 |
|
✗ |
START_PROFILER("SolverDDP::computeGains"); |
| 377 |
|
✗ |
const std::size_t nu = problem_->get_runningModels()[t]->get_nu(); |
| 378 |
|
✗ |
if (nu > 0) { |
| 379 |
|
✗ |
START_PROFILER("SolverDDP::Quu_inv"); |
| 380 |
|
✗ |
Quu_llt_[t].compute(Quu_[t]); |
| 381 |
|
✗ |
STOP_PROFILER("SolverDDP::Quu_inv"); |
| 382 |
|
✗ |
const Eigen::ComputationInfo& info = Quu_llt_[t].info(); |
| 383 |
|
✗ |
if (info != Eigen::Success) { |
| 384 |
|
✗ |
STOP_PROFILER("SolverDDP::computeGains"); |
| 385 |
|
✗ |
throw_pretty("backward_error"); |
| 386 |
|
|
} |
| 387 |
|
✗ |
K_[t] = Qxu_[t].transpose(); |
| 388 |
|
|
|
| 389 |
|
✗ |
START_PROFILER("SolverDDP::Quu_inv_Qux"); |
| 390 |
|
✗ |
Quu_llt_[t].solveInPlace(K_[t]); |
| 391 |
|
✗ |
STOP_PROFILER("SolverDDP::Quu_inv_Qux"); |
| 392 |
|
✗ |
k_[t] = Qu_[t]; |
| 393 |
|
✗ |
Quu_llt_[t].solveInPlace(k_[t]); |
| 394 |
|
|
} |
| 395 |
|
✗ |
STOP_PROFILER("SolverDDP::computeGains"); |
| 396 |
|
|
} |
| 397 |
|
|
|
| 398 |
|
✗ |
void SolverDDP::increaseRegularization() { |
| 399 |
|
✗ |
preg_ *= reg_incfactor_; |
| 400 |
|
✗ |
if (preg_ > reg_max_) { |
| 401 |
|
✗ |
preg_ = reg_max_; |
| 402 |
|
|
} |
| 403 |
|
✗ |
dreg_ = preg_; |
| 404 |
|
|
} |
| 405 |
|
|
|
| 406 |
|
✗ |
void SolverDDP::decreaseRegularization() { |
| 407 |
|
✗ |
preg_ /= reg_decfactor_; |
| 408 |
|
✗ |
if (preg_ < reg_min_) { |
| 409 |
|
✗ |
preg_ = reg_min_; |
| 410 |
|
|
} |
| 411 |
|
✗ |
dreg_ = preg_; |
| 412 |
|
|
} |
| 413 |
|
|
|
| 414 |
|
✗ |
void SolverDDP::allocateData() { |
| 415 |
|
✗ |
const std::size_t T = problem_->get_T(); |
| 416 |
|
✗ |
Vxx_.resize(T + 1); |
| 417 |
|
✗ |
Vx_.resize(T + 1); |
| 418 |
|
✗ |
Qxx_.resize(T); |
| 419 |
|
✗ |
Qxu_.resize(T); |
| 420 |
|
✗ |
Quu_.resize(T); |
| 421 |
|
✗ |
Qx_.resize(T); |
| 422 |
|
✗ |
Qu_.resize(T); |
| 423 |
|
✗ |
K_.resize(T); |
| 424 |
|
✗ |
k_.resize(T); |
| 425 |
|
|
|
| 426 |
|
✗ |
xs_try_.resize(T + 1); |
| 427 |
|
✗ |
us_try_.resize(T); |
| 428 |
|
✗ |
dx_.resize(T); |
| 429 |
|
|
|
| 430 |
|
✗ |
FuTVxx_p_.resize(T); |
| 431 |
|
✗ |
Quu_llt_.resize(T); |
| 432 |
|
✗ |
Quuk_.resize(T); |
| 433 |
|
|
|
| 434 |
|
✗ |
const std::size_t ndx = problem_->get_ndx(); |
| 435 |
|
|
const std::vector<std::shared_ptr<ActionModelAbstract> >& models = |
| 436 |
|
✗ |
problem_->get_runningModels(); |
| 437 |
|
✗ |
for (std::size_t t = 0; t < T; ++t) { |
| 438 |
|
✗ |
const std::shared_ptr<ActionModelAbstract>& model = models[t]; |
| 439 |
|
✗ |
const std::size_t nu = model->get_nu(); |
| 440 |
|
✗ |
Vxx_[t] = Eigen::MatrixXd::Zero(ndx, ndx); |
| 441 |
|
✗ |
Vx_[t] = Eigen::VectorXd::Zero(ndx); |
| 442 |
|
✗ |
Qxx_[t] = Eigen::MatrixXd::Zero(ndx, ndx); |
| 443 |
|
✗ |
Qxu_[t] = Eigen::MatrixXd::Zero(ndx, nu); |
| 444 |
|
✗ |
Quu_[t] = Eigen::MatrixXd::Zero(nu, nu); |
| 445 |
|
✗ |
Qx_[t] = Eigen::VectorXd::Zero(ndx); |
| 446 |
|
✗ |
Qu_[t] = Eigen::VectorXd::Zero(nu); |
| 447 |
|
✗ |
K_[t] = MatrixXdRowMajor::Zero(nu, ndx); |
| 448 |
|
✗ |
k_[t] = Eigen::VectorXd::Zero(nu); |
| 449 |
|
|
|
| 450 |
|
✗ |
if (t == 0) { |
| 451 |
|
✗ |
xs_try_[t] = problem_->get_x0(); |
| 452 |
|
|
} else { |
| 453 |
|
✗ |
xs_try_[t] = model->get_state()->zero(); |
| 454 |
|
|
} |
| 455 |
|
✗ |
us_try_[t] = Eigen::VectorXd::Zero(nu); |
| 456 |
|
✗ |
dx_[t] = Eigen::VectorXd::Zero(ndx); |
| 457 |
|
|
|
| 458 |
|
✗ |
FuTVxx_p_[t] = MatrixXdRowMajor::Zero(nu, ndx); |
| 459 |
|
✗ |
Quu_llt_[t] = Eigen::LLT<Eigen::MatrixXd>(nu); |
| 460 |
|
✗ |
Quuk_[t] = Eigen::VectorXd(nu); |
| 461 |
|
|
} |
| 462 |
|
✗ |
Vxx_.back() = Eigen::MatrixXd::Zero(ndx, ndx); |
| 463 |
|
✗ |
Vxx_tmp_ = Eigen::MatrixXd::Zero(ndx, ndx); |
| 464 |
|
✗ |
Vx_.back() = Eigen::VectorXd::Zero(ndx); |
| 465 |
|
✗ |
xs_try_.back() = problem_->get_terminalModel()->get_state()->zero(); |
| 466 |
|
|
|
| 467 |
|
✗ |
FxTVxx_p_ = MatrixXdRowMajor::Zero(ndx, ndx); |
| 468 |
|
✗ |
fTVxx_p_ = Eigen::VectorXd::Zero(ndx); |
| 469 |
|
|
} |
| 470 |
|
|
|
| 471 |
|
✗ |
double SolverDDP::get_reg_incfactor() const { return reg_incfactor_; } |
| 472 |
|
|
|
| 473 |
|
✗ |
double SolverDDP::get_reg_decfactor() const { return reg_decfactor_; } |
| 474 |
|
|
|
| 475 |
|
✗ |
double SolverDDP::get_regfactor() const { return reg_incfactor_; } |
| 476 |
|
|
|
| 477 |
|
✗ |
double SolverDDP::get_reg_min() const { return reg_min_; } |
| 478 |
|
|
|
| 479 |
|
✗ |
double SolverDDP::get_regmin() const { return reg_min_; } |
| 480 |
|
|
|
| 481 |
|
✗ |
double SolverDDP::get_reg_max() const { return reg_max_; } |
| 482 |
|
|
|
| 483 |
|
✗ |
double SolverDDP::get_regmax() const { return reg_max_; } |
| 484 |
|
|
|
| 485 |
|
✗ |
const std::vector<double>& SolverDDP::get_alphas() const { return alphas_; } |
| 486 |
|
|
|
| 487 |
|
✗ |
double SolverDDP::get_th_stepdec() const { return th_stepdec_; } |
| 488 |
|
|
|
| 489 |
|
✗ |
double SolverDDP::get_th_stepinc() const { return th_stepinc_; } |
| 490 |
|
|
|
| 491 |
|
✗ |
double SolverDDP::get_th_grad() const { return th_grad_; } |
| 492 |
|
|
|
| 493 |
|
✗ |
const std::vector<Eigen::MatrixXd>& SolverDDP::get_Vxx() const { return Vxx_; } |
| 494 |
|
|
|
| 495 |
|
✗ |
const std::vector<Eigen::VectorXd>& SolverDDP::get_Vx() const { return Vx_; } |
| 496 |
|
|
|
| 497 |
|
✗ |
const std::vector<Eigen::MatrixXd>& SolverDDP::get_Qxx() const { return Qxx_; } |
| 498 |
|
|
|
| 499 |
|
✗ |
const std::vector<Eigen::MatrixXd>& SolverDDP::get_Qxu() const { return Qxu_; } |
| 500 |
|
|
|
| 501 |
|
✗ |
const std::vector<Eigen::MatrixXd>& SolverDDP::get_Quu() const { return Quu_; } |
| 502 |
|
|
|
| 503 |
|
✗ |
const std::vector<Eigen::VectorXd>& SolverDDP::get_Qx() const { return Qx_; } |
| 504 |
|
|
|
| 505 |
|
✗ |
const std::vector<Eigen::VectorXd>& SolverDDP::get_Qu() const { return Qu_; } |
| 506 |
|
|
|
| 507 |
|
|
const std::vector<typename MathBaseTpl<double>::MatrixXsRowMajor>& |
| 508 |
|
✗ |
SolverDDP::get_K() const { |
| 509 |
|
✗ |
return K_; |
| 510 |
|
|
} |
| 511 |
|
|
|
| 512 |
|
✗ |
const std::vector<Eigen::VectorXd>& SolverDDP::get_k() const { return k_; } |
| 513 |
|
|
|
| 514 |
|
✗ |
void SolverDDP::set_reg_incfactor(const double regfactor) { |
| 515 |
|
✗ |
if (regfactor <= 1.) { |
| 516 |
|
✗ |
throw_pretty( |
| 517 |
|
|
"Invalid argument: " << "reg_incfactor value is higher than 1."); |
| 518 |
|
|
} |
| 519 |
|
✗ |
reg_incfactor_ = regfactor; |
| 520 |
|
|
} |
| 521 |
|
|
|
| 522 |
|
✗ |
void SolverDDP::set_reg_decfactor(const double regfactor) { |
| 523 |
|
✗ |
if (regfactor <= 1.) { |
| 524 |
|
✗ |
throw_pretty( |
| 525 |
|
|
"Invalid argument: " << "reg_decfactor value is higher than 1."); |
| 526 |
|
|
} |
| 527 |
|
✗ |
reg_decfactor_ = regfactor; |
| 528 |
|
|
} |
| 529 |
|
|
|
| 530 |
|
✗ |
void SolverDDP::set_regfactor(const double regfactor) { |
| 531 |
|
✗ |
if (regfactor <= 1.) { |
| 532 |
|
✗ |
throw_pretty("Invalid argument: " << "regfactor value is higher than 1."); |
| 533 |
|
|
} |
| 534 |
|
✗ |
set_reg_incfactor(regfactor); |
| 535 |
|
✗ |
set_reg_decfactor(regfactor); |
| 536 |
|
|
} |
| 537 |
|
|
|
| 538 |
|
✗ |
void SolverDDP::set_reg_min(const double regmin) { |
| 539 |
|
✗ |
if (0. > regmin) { |
| 540 |
|
✗ |
throw_pretty("Invalid argument: " << "regmin value has to be positive."); |
| 541 |
|
|
} |
| 542 |
|
✗ |
reg_min_ = regmin; |
| 543 |
|
|
} |
| 544 |
|
|
|
| 545 |
|
✗ |
void SolverDDP::set_regmin(const double regmin) { |
| 546 |
|
✗ |
if (0. > regmin) { |
| 547 |
|
✗ |
throw_pretty("Invalid argument: " << "regmin value has to be positive."); |
| 548 |
|
|
} |
| 549 |
|
✗ |
reg_min_ = regmin; |
| 550 |
|
|
} |
| 551 |
|
|
|
| 552 |
|
✗ |
void SolverDDP::set_reg_max(const double regmax) { |
| 553 |
|
✗ |
if (0. > regmax) { |
| 554 |
|
✗ |
throw_pretty("Invalid argument: " << "regmax value has to be positive."); |
| 555 |
|
|
} |
| 556 |
|
✗ |
reg_max_ = regmax; |
| 557 |
|
|
} |
| 558 |
|
|
|
| 559 |
|
✗ |
void SolverDDP::set_regmax(const double regmax) { |
| 560 |
|
✗ |
if (0. > regmax) { |
| 561 |
|
✗ |
throw_pretty("Invalid argument: " << "regmax value has to be positive."); |
| 562 |
|
|
} |
| 563 |
|
✗ |
reg_max_ = regmax; |
| 564 |
|
|
} |
| 565 |
|
|
|
| 566 |
|
✗ |
void SolverDDP::set_alphas(const std::vector<double>& alphas) { |
| 567 |
|
✗ |
double prev_alpha = alphas[0]; |
| 568 |
|
✗ |
if (prev_alpha != 1.) { |
| 569 |
|
✗ |
std::cerr << "Warning: alpha[0] should be 1" << std::endl; |
| 570 |
|
|
} |
| 571 |
|
✗ |
for (std::size_t i = 1; i < alphas.size(); ++i) { |
| 572 |
|
✗ |
double alpha = alphas[i]; |
| 573 |
|
✗ |
if (0. >= alpha) { |
| 574 |
|
✗ |
throw_pretty("Invalid argument: " << "alpha values has to be positive."); |
| 575 |
|
|
} |
| 576 |
|
✗ |
if (alpha >= prev_alpha) { |
| 577 |
|
✗ |
throw_pretty( |
| 578 |
|
|
"Invalid argument: " << "alpha values are monotonously decreasing."); |
| 579 |
|
|
} |
| 580 |
|
✗ |
prev_alpha = alpha; |
| 581 |
|
|
} |
| 582 |
|
✗ |
alphas_ = alphas; |
| 583 |
|
|
} |
| 584 |
|
|
|
| 585 |
|
✗ |
void SolverDDP::set_th_stepdec(const double th_stepdec) { |
| 586 |
|
✗ |
if (0. >= th_stepdec || th_stepdec > 1.) { |
| 587 |
|
✗ |
throw_pretty( |
| 588 |
|
|
"Invalid argument: " << "th_stepdec value should between 0 and 1."); |
| 589 |
|
|
} |
| 590 |
|
✗ |
th_stepdec_ = th_stepdec; |
| 591 |
|
|
} |
| 592 |
|
|
|
| 593 |
|
✗ |
void SolverDDP::set_th_stepinc(const double th_stepinc) { |
| 594 |
|
✗ |
if (0. >= th_stepinc || th_stepinc > 1.) { |
| 595 |
|
✗ |
throw_pretty( |
| 596 |
|
|
"Invalid argument: " << "th_stepinc value should between 0 and 1."); |
| 597 |
|
|
} |
| 598 |
|
✗ |
th_stepinc_ = th_stepinc; |
| 599 |
|
|
} |
| 600 |
|
|
|
| 601 |
|
✗ |
void SolverDDP::set_th_grad(const double th_grad) { |
| 602 |
|
✗ |
if (0. > th_grad) { |
| 603 |
|
✗ |
throw_pretty("Invalid argument: " << "th_grad value has to be positive."); |
| 604 |
|
|
} |
| 605 |
|
✗ |
th_grad_ = th_grad; |
| 606 |
|
|
} |
| 607 |
|
|
|
| 608 |
|
|
} // namespace crocoddyl |
| 609 |
|
|
|