| Line |
Branch |
Exec |
Source |
| 1 |
|
|
// |
| 2 |
|
|
// Copyright (c) 2022-2024 INRIA |
| 3 |
|
|
// |
| 4 |
|
|
|
| 5 |
|
|
#ifndef __pinocchio_algorithm_admm_solver_hpp__ |
| 6 |
|
|
#define __pinocchio_algorithm_admm_solver_hpp__ |
| 7 |
|
|
|
| 8 |
|
|
#include "pinocchio/algorithm/constraints/fwd.hpp" |
| 9 |
|
|
#include "pinocchio/math/fwd.hpp" |
| 10 |
|
|
#include "pinocchio/math/comparison-operators.hpp" |
| 11 |
|
|
#include "pinocchio/math/eigenvalues.hpp" |
| 12 |
|
|
|
| 13 |
|
|
#include "pinocchio/algorithm/contact-solver-base.hpp" |
| 14 |
|
|
#include "pinocchio/algorithm/delassus-operator-base.hpp" |
| 15 |
|
|
|
| 16 |
|
|
#include <boost/optional.hpp> |
| 17 |
|
|
|
| 18 |
|
|
namespace pinocchio |
| 19 |
|
|
{ |
| 20 |
|
|
template<typename _Scalar> |
| 21 |
|
|
struct PINOCCHIO_UNSUPPORTED_MESSAGE("The API will change towards more flexibility") |
| 22 |
|
|
ADMMContactSolverTpl : ContactSolverBaseTpl<_Scalar> |
| 23 |
|
|
{ |
| 24 |
|
|
typedef _Scalar Scalar; |
| 25 |
|
|
typedef ContactSolverBaseTpl<_Scalar> Base; |
| 26 |
|
|
typedef Eigen::Matrix<Scalar, Eigen::Dynamic, 1> VectorXs; |
| 27 |
|
|
typedef const Eigen::Ref<const VectorXs> ConstRefVectorXs; |
| 28 |
|
|
typedef Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic> MatrixXs; |
| 29 |
|
|
typedef PowerIterationAlgoTpl<VectorXs> PowerIterationAlgo; |
| 30 |
|
|
|
| 31 |
|
|
using Base::problem_size; |
| 32 |
|
|
|
| 33 |
|
|
// struct SolverParameters |
| 34 |
|
|
// { |
| 35 |
|
|
// explicit SolverParameters(const int problem_dim) |
| 36 |
|
|
// : rho_power(Scalar(0.2)) |
| 37 |
|
|
// , ratio_primal_dual(Scalar(10)) |
| 38 |
|
|
// , mu_prox |
| 39 |
|
|
// { |
| 40 |
|
|
// |
| 41 |
|
|
// } |
| 42 |
|
|
// |
| 43 |
|
|
// /// \brief Rho solver ADMM |
| 44 |
|
|
// boost::optional<Scalar> rho; |
| 45 |
|
|
// /// \brief Power value associated to rho. This quantity will be automatically updated. |
| 46 |
|
|
// Scalar rho_power; |
| 47 |
|
|
// /// \brief Ratio primal/dual |
| 48 |
|
|
// Scalar ratio_primal_dual; |
| 49 |
|
|
// /// \brief Proximal value |
| 50 |
|
|
// Scalar mu_prox; |
| 51 |
|
|
// |
| 52 |
|
|
// /// \brief Largest eigenvalue |
| 53 |
|
|
// boost::optional<Scalar> L_value; |
| 54 |
|
|
// /// \brief Largest eigenvector |
| 55 |
|
|
// boost::optional<VectorXs> L_vector; |
| 56 |
|
|
// }; |
| 57 |
|
|
// |
| 58 |
|
|
struct SolverStats |
| 59 |
|
|
{ |
| 60 |
|
✗ |
explicit SolverStats(const int max_it) |
| 61 |
|
✗ |
: it(0) |
| 62 |
|
✗ |
, cholesky_update_count(0) |
| 63 |
|
|
{ |
| 64 |
|
✗ |
primal_feasibility.reserve(size_t(max_it)); |
| 65 |
|
✗ |
dual_feasibility.reserve(size_t(max_it)); |
| 66 |
|
✗ |
dual_feasibility_ncp.reserve(size_t(max_it)); |
| 67 |
|
✗ |
complementarity.reserve(size_t(max_it)); |
| 68 |
|
✗ |
rho.reserve(size_t(max_it)); |
| 69 |
|
|
} |
| 70 |
|
|
|
| 71 |
|
✗ |
void reset() |
| 72 |
|
|
{ |
| 73 |
|
✗ |
primal_feasibility.clear(); |
| 74 |
|
✗ |
dual_feasibility.clear(); |
| 75 |
|
✗ |
complementarity.clear(); |
| 76 |
|
✗ |
dual_feasibility_ncp.clear(); |
| 77 |
|
✗ |
rho.clear(); |
| 78 |
|
✗ |
it = 0; |
| 79 |
|
✗ |
cholesky_update_count = 0; |
| 80 |
|
|
} |
| 81 |
|
|
|
| 82 |
|
✗ |
size_t size() const |
| 83 |
|
|
{ |
| 84 |
|
✗ |
return primal_feasibility.size(); |
| 85 |
|
|
} |
| 86 |
|
|
|
| 87 |
|
|
/// \brief Number of total iterations. |
| 88 |
|
|
int it; |
| 89 |
|
|
|
| 90 |
|
|
/// \brief Number of Cholesky updates. |
| 91 |
|
|
int cholesky_update_count; |
| 92 |
|
|
|
| 93 |
|
|
/// \brief History of primal feasibility values. |
| 94 |
|
|
std::vector<Scalar> primal_feasibility; |
| 95 |
|
|
|
| 96 |
|
|
/// \brief History of dual feasibility values. |
| 97 |
|
|
std::vector<Scalar> dual_feasibility; |
| 98 |
|
|
std::vector<Scalar> dual_feasibility_ncp; |
| 99 |
|
|
|
| 100 |
|
|
/// \brief History of complementarity values. |
| 101 |
|
|
std::vector<Scalar> complementarity; |
| 102 |
|
|
|
| 103 |
|
|
/// \brief History of rho values. |
| 104 |
|
|
std::vector<Scalar> rho; |
| 105 |
|
|
}; |
| 106 |
|
|
// |
| 107 |
|
|
// struct SolverResults |
| 108 |
|
|
// { |
| 109 |
|
|
// explicit SolverResults(const int problem_dim, const int max_it) |
| 110 |
|
|
// : L_vector(problem_dim) |
| 111 |
|
|
// |
| 112 |
|
|
// /// \brief Largest eigenvalue |
| 113 |
|
|
// Scalar L_value; |
| 114 |
|
|
// /// \brief Largest eigenvector |
| 115 |
|
|
// VectorXs L_vector; |
| 116 |
|
|
// |
| 117 |
|
|
// SolverStats stats; |
| 118 |
|
|
// }; |
| 119 |
|
|
|
| 120 |
|
✗ |
explicit ADMMContactSolverTpl( |
| 121 |
|
|
int problem_dim, |
| 122 |
|
|
Scalar mu_prox = Scalar(1e-6), |
| 123 |
|
|
Scalar tau = Scalar(0.5), |
| 124 |
|
|
Scalar rho_power = Scalar(0.2), |
| 125 |
|
|
Scalar rho_power_factor = Scalar(0.05), |
| 126 |
|
|
Scalar ratio_primal_dual = Scalar(10), |
| 127 |
|
|
int max_it_largest_eigen_value_solver = 20) |
| 128 |
|
|
: Base(problem_dim) |
| 129 |
|
✗ |
, is_initialized(false) |
| 130 |
|
✗ |
, mu_prox(mu_prox) |
| 131 |
|
✗ |
, tau(tau) |
| 132 |
|
✗ |
, rho(Scalar(-1)) |
| 133 |
|
✗ |
, rho_power(rho_power) |
| 134 |
|
✗ |
, rho_power_factor(rho_power_factor) |
| 135 |
|
✗ |
, ratio_primal_dual(ratio_primal_dual) |
| 136 |
|
✗ |
, max_it_largest_eigen_value_solver(max_it_largest_eigen_value_solver) |
| 137 |
|
✗ |
, power_iteration_algo(problem_dim) |
| 138 |
|
✗ |
, x_(VectorXs::Zero(problem_dim)) |
| 139 |
|
✗ |
, y_(VectorXs::Zero(problem_dim)) |
| 140 |
|
✗ |
, x_previous(VectorXs::Zero(problem_dim)) |
| 141 |
|
✗ |
, y_previous(VectorXs::Zero(problem_dim)) |
| 142 |
|
✗ |
, z_previous(VectorXs::Zero(problem_dim)) |
| 143 |
|
✗ |
, z_(VectorXs::Zero(problem_dim)) |
| 144 |
|
✗ |
, s_(VectorXs::Zero(problem_dim)) |
| 145 |
|
✗ |
, rhs(problem_dim) |
| 146 |
|
✗ |
, primal_feasibility_vector(VectorXs::Zero(problem_dim)) |
| 147 |
|
✗ |
, dual_feasibility_vector(VectorXs::Zero(problem_dim)) |
| 148 |
|
✗ |
, stats(Base::max_it) |
| 149 |
|
|
{ |
| 150 |
|
✗ |
power_iteration_algo.max_it = max_it_largest_eigen_value_solver; |
| 151 |
|
|
} |
| 152 |
|
|
|
| 153 |
|
|
/// \brief Set the ADMM penalty value. |
| 154 |
|
✗ |
void setRho(const Scalar rho) |
| 155 |
|
|
{ |
| 156 |
|
✗ |
this->rho = rho; |
| 157 |
|
|
} |
| 158 |
|
|
/// \brief Get the ADMM penalty value. |
| 159 |
|
✗ |
Scalar getRho() const |
| 160 |
|
|
{ |
| 161 |
|
✗ |
return rho; |
| 162 |
|
|
} |
| 163 |
|
|
|
| 164 |
|
|
/// \brief Set the power associated to the problem conditionning. |
| 165 |
|
✗ |
void setRhoPower(const Scalar rho_power) |
| 166 |
|
|
{ |
| 167 |
|
✗ |
this->rho_power = rho_power; |
| 168 |
|
|
} |
| 169 |
|
|
/// \brief Get the power associated to the problem conditionning. |
| 170 |
|
✗ |
Scalar getRhoPower() const |
| 171 |
|
|
{ |
| 172 |
|
✗ |
return rho_power; |
| 173 |
|
|
} |
| 174 |
|
|
|
| 175 |
|
|
/// \brief Set the power factor associated to the problem conditionning. |
| 176 |
|
✗ |
void setRhoPowerFactor(const Scalar rho_power_factor) |
| 177 |
|
|
{ |
| 178 |
|
✗ |
this->rho_power_factor = rho_power_factor; |
| 179 |
|
|
} |
| 180 |
|
|
/// \brief Get the power factor associated to the problem conditionning. |
| 181 |
|
✗ |
Scalar getRhoPowerFactor() const |
| 182 |
|
|
{ |
| 183 |
|
✗ |
return rho_power_factor; |
| 184 |
|
|
} |
| 185 |
|
|
|
| 186 |
|
|
/// \brief Set the tau linear scaling factor. |
| 187 |
|
✗ |
void setTau(const Scalar tau) |
| 188 |
|
|
{ |
| 189 |
|
✗ |
this->tau = tau; |
| 190 |
|
|
} |
| 191 |
|
|
/// \brief Get the tau linear scaling factor. |
| 192 |
|
✗ |
Scalar getTau() const |
| 193 |
|
|
{ |
| 194 |
|
✗ |
return tau; |
| 195 |
|
|
} |
| 196 |
|
|
|
| 197 |
|
|
/// \brief Set the proximal value. |
| 198 |
|
✗ |
void setProximalValue(const Scalar mu) |
| 199 |
|
|
{ |
| 200 |
|
✗ |
this->mu_prox = mu; |
| 201 |
|
|
} |
| 202 |
|
|
/// \brief Get the proximal value. |
| 203 |
|
✗ |
Scalar getProximalValue() const |
| 204 |
|
|
{ |
| 205 |
|
✗ |
return mu_prox; |
| 206 |
|
|
} |
| 207 |
|
|
|
| 208 |
|
|
/// \brief Set the primal/dual ratio. |
| 209 |
|
✗ |
void setRatioPrimalDual(const Scalar ratio_primal_dual) |
| 210 |
|
|
{ |
| 211 |
|
✗ |
PINOCCHIO_CHECK_INPUT_ARGUMENT( |
| 212 |
|
|
ratio_primal_dual > 0., "The ratio primal/dual should be positive strictly"); |
| 213 |
|
✗ |
this->ratio_primal_dual = ratio_primal_dual; |
| 214 |
|
|
} |
| 215 |
|
|
/// \brief Get the primal/dual ratio. |
| 216 |
|
✗ |
Scalar getRatioPrimalDual() const |
| 217 |
|
|
{ |
| 218 |
|
✗ |
return ratio_primal_dual; |
| 219 |
|
|
} |
| 220 |
|
|
|
| 221 |
|
|
/// \returns the number of updates of the Cholesky factorization due to rho updates. |
| 222 |
|
✗ |
int getCholeskyUpdateCount() const |
| 223 |
|
|
{ |
| 224 |
|
✗ |
return cholesky_update_count; |
| 225 |
|
|
} |
| 226 |
|
|
|
| 227 |
|
|
/// |
| 228 |
|
|
/// \brief Solve the constrained conic problem composed of problem data (G,g,cones) and starting |
| 229 |
|
|
/// from the initial guess. |
| 230 |
|
|
/// |
| 231 |
|
|
/// \param[in] G Symmetric PSD matrix representing the Delassus of the contact problem. |
| 232 |
|
|
/// \param[in] g Free contact acceleration or velicity associted with the contact problem. |
| 233 |
|
|
/// \param[in] cones Vector of conic constraints. |
| 234 |
|
|
/// \param[in,out] x Initial guess and output solution of the problem |
| 235 |
|
|
/// \param[in] mu_prox Proximal smoothing value associated to the algorithm. |
| 236 |
|
|
/// \param[in] R Proximal regularization value associated to the compliant contacts (corresponds |
| 237 |
|
|
/// to the lowest non-zero). \param[in] tau Over relaxation value |
| 238 |
|
|
/// |
| 239 |
|
|
/// \returns True if the problem has converged. |
| 240 |
|
|
template< |
| 241 |
|
|
typename DelassusDerived, |
| 242 |
|
|
typename VectorLike, |
| 243 |
|
|
typename ConstraintAllocator, |
| 244 |
|
|
typename VectorLikeR> |
| 245 |
|
|
bool solve( |
| 246 |
|
|
DelassusOperatorBase<DelassusDerived> & delassus, |
| 247 |
|
|
const Eigen::MatrixBase<VectorLike> & g, |
| 248 |
|
|
const std::vector<CoulombFrictionConeTpl<Scalar>, ConstraintAllocator> & cones, |
| 249 |
|
|
const Eigen::MatrixBase<VectorLikeR> & R, |
| 250 |
|
|
const boost::optional<ConstRefVectorXs> primal_guess = boost::none, |
| 251 |
|
|
const boost::optional<ConstRefVectorXs> dual_guess = boost::none, |
| 252 |
|
|
bool compute_largest_eigen_values = true, |
| 253 |
|
|
bool stat_record = false); |
| 254 |
|
|
|
| 255 |
|
|
/// |
| 256 |
|
|
/// \brief Solve the constrained conic problem composed of problem data (G,g,cones) and starting |
| 257 |
|
|
/// from the initial guess. |
| 258 |
|
|
/// |
| 259 |
|
|
/// \param[in] G Symmetric PSD matrix representing the Delassus of the contact problem. |
| 260 |
|
|
/// \param[in] g Free contact acceleration or velicity associted with the contact problem. |
| 261 |
|
|
/// \param[in] cones Vector of conic constraints. |
| 262 |
|
|
/// \param[in,out] x Initial guess and output solution of the problem |
| 263 |
|
|
/// \param[in] mu_prox Proximal smoothing value associated to the algorithm. |
| 264 |
|
|
/// \param[in] tau Over relaxation value |
| 265 |
|
|
/// |
| 266 |
|
|
/// \returns True if the problem has converged. |
| 267 |
|
|
template< |
| 268 |
|
|
typename DelassusDerived, |
| 269 |
|
|
typename VectorLike, |
| 270 |
|
|
typename ConstraintAllocator, |
| 271 |
|
|
typename VectorLikeOut> |
| 272 |
|
|
bool solve( |
| 273 |
|
|
DelassusOperatorBase<DelassusDerived> & delassus, |
| 274 |
|
|
const Eigen::MatrixBase<VectorLike> & g, |
| 275 |
|
|
const std::vector<CoulombFrictionConeTpl<Scalar>, ConstraintAllocator> & cones, |
| 276 |
|
|
const Eigen::DenseBase<VectorLikeOut> & x) |
| 277 |
|
|
{ |
| 278 |
|
|
return solve( |
| 279 |
|
|
delassus.derived(), g.derived(), cones, x.const_cast_derived(), |
| 280 |
|
|
VectorXs::Zero(problem_size)); |
| 281 |
|
|
} |
| 282 |
|
|
|
| 283 |
|
|
/// \returns the primal solution of the problem |
| 284 |
|
✗ |
const VectorXs & getPrimalSolution() const |
| 285 |
|
|
{ |
| 286 |
|
✗ |
return y_; |
| 287 |
|
|
} |
| 288 |
|
|
/// \returns the dual solution of the problem |
| 289 |
|
✗ |
const VectorXs & getDualSolution() const |
| 290 |
|
|
{ |
| 291 |
|
✗ |
return z_; |
| 292 |
|
|
} |
| 293 |
|
|
/// \returns the complementarity shift |
| 294 |
|
|
const VectorXs & getComplementarityShift() const |
| 295 |
|
|
{ |
| 296 |
|
|
return s_; |
| 297 |
|
|
} |
| 298 |
|
|
|
| 299 |
|
|
/// \brief Compute the penalty ADMM value from the current largest and lowest eigenvalues and |
| 300 |
|
|
/// the scaling spectral factor. |
| 301 |
|
✗ |
static inline Scalar computeRho(const Scalar L, const Scalar m, const Scalar rho_power) |
| 302 |
|
|
{ |
| 303 |
|
✗ |
const Scalar cond = L / m; |
| 304 |
|
✗ |
const Scalar rho = math::sqrt(L * m) * math::pow(cond, rho_power); |
| 305 |
|
✗ |
return rho; |
| 306 |
|
|
} |
| 307 |
|
|
|
| 308 |
|
|
/// \brief Compute the scaling spectral factor of the ADMM penalty term from the current |
| 309 |
|
|
/// largest and lowest eigenvalues and the ADMM penalty term. |
| 310 |
|
✗ |
static inline Scalar computeRhoPower(const Scalar L, const Scalar m, const Scalar rho) |
| 311 |
|
|
{ |
| 312 |
|
✗ |
const Scalar cond = L / m; |
| 313 |
|
✗ |
const Scalar sqtr_L_m = math::sqrt(L * m); |
| 314 |
|
✗ |
const Scalar rho_power = math::log(rho / sqtr_L_m) / math::log(cond); |
| 315 |
|
✗ |
return rho_power; |
| 316 |
|
|
} |
| 317 |
|
|
|
| 318 |
|
✗ |
PowerIterationAlgo & getPowerIterationAlgo() |
| 319 |
|
|
{ |
| 320 |
|
✗ |
return power_iteration_algo; |
| 321 |
|
|
} |
| 322 |
|
|
|
| 323 |
|
✗ |
SolverStats & getStats() |
| 324 |
|
|
{ |
| 325 |
|
✗ |
return stats; |
| 326 |
|
|
} |
| 327 |
|
|
|
| 328 |
|
|
protected: |
| 329 |
|
|
bool is_initialized; |
| 330 |
|
|
|
| 331 |
|
|
/// \brief proximal value |
| 332 |
|
|
Scalar mu_prox; |
| 333 |
|
|
|
| 334 |
|
|
/// \brief Linear scaling of the ADMM penalty term |
| 335 |
|
|
Scalar tau; |
| 336 |
|
|
|
| 337 |
|
|
/// \brief Penalty term associated to the ADMM. |
| 338 |
|
|
Scalar rho; |
| 339 |
|
|
/// \brief Power value associated to rho. This quantity will be automatically updated. |
| 340 |
|
|
Scalar rho_power; |
| 341 |
|
|
/// \brief Update factor for the primal/dual update of rho. |
| 342 |
|
|
Scalar rho_power_factor; |
| 343 |
|
|
/// \brief Ratio primal/dual |
| 344 |
|
|
Scalar ratio_primal_dual; |
| 345 |
|
|
|
| 346 |
|
|
/// \brief Maximum number of iterarions called for the power iteration algorithm |
| 347 |
|
|
int max_it_largest_eigen_value_solver; |
| 348 |
|
|
|
| 349 |
|
|
/// \brief Power iteration algo. |
| 350 |
|
|
PowerIterationAlgo power_iteration_algo; |
| 351 |
|
|
|
| 352 |
|
|
/// \brief Primal variables (corresponds to the contact forces) |
| 353 |
|
|
VectorXs x_, y_; |
| 354 |
|
|
/// \brief Previous value of y. |
| 355 |
|
|
VectorXs x_previous, y_previous, z_previous; |
| 356 |
|
|
/// \brief Dual varible of the ADMM (corresponds to the contact velocity or acceleration). |
| 357 |
|
|
VectorXs z_; |
| 358 |
|
|
/// \brief De Saxé shift |
| 359 |
|
|
VectorXs s_; |
| 360 |
|
|
|
| 361 |
|
|
VectorXs rhs, primal_feasibility_vector, dual_feasibility_vector; |
| 362 |
|
|
|
| 363 |
|
|
int cholesky_update_count; |
| 364 |
|
|
|
| 365 |
|
|
/// \brief Stats of the solver |
| 366 |
|
|
SolverStats stats; |
| 367 |
|
|
|
| 368 |
|
|
#ifdef PINOCCHIO_WITH_HPP_FCL |
| 369 |
|
|
using Base::timer; |
| 370 |
|
|
#endif // PINOCCHIO_WITH_HPP_FCL |
| 371 |
|
|
}; // struct ADMMContactSolverTpl |
| 372 |
|
|
} // namespace pinocchio |
| 373 |
|
|
|
| 374 |
|
|
#include "pinocchio/algorithm/admm-solver.hxx" |
| 375 |
|
|
|
| 376 |
|
|
#endif // ifndef __pinocchio_algorithm_admm_solver_hpp__ |
| 377 |
|
|
|