| 1 | /* |
| 2 | * @file SimpleHelicopter.h |
| 3 | * @brief Implement SimpleHelicopter discrete dynamics model and variational integrator, |
| 4 | * following [Kobilarov09siggraph] |
| 5 | * @author Duy-Nguyen Ta |
| 6 | */ |
| 7 | |
| 8 | #pragma once |
| 9 | |
| 10 | #include <gtsam/nonlinear/NonlinearFactor.h> |
| 11 | #include <gtsam/geometry/Pose3.h> |
| 12 | #include <gtsam/base/numericalDerivative.h> |
| 13 | #include <cmath> |
| 14 | |
| 15 | namespace gtsam { |
| 16 | |
| 17 | /** |
| 18 | * Implement the Reconstruction equation: \f$ g_{k+1} = g_k \exp (h\xi_k) \f$, where |
| 19 | * \f$ h \f$: timestep (parameter) |
| 20 | * \f$ g_{k+1}, g_{k} \f$: poses at the current and the next timestep |
| 21 | * \f$ \xi_k \f$: the body-fixed velocity (Lie algebra) |
| 22 | * It is somewhat similar to BetweenFactor, but treats the body-fixed velocity |
| 23 | * \f$ \xi_k \f$ as a variable. So it is a three-way factor. |
| 24 | * Note: this factor is necessary if one needs to smooth the entire graph. It's not needed |
| 25 | * in sequential update method. |
| 26 | */ |
| 27 | class Reconstruction : public NoiseModelFactorN<Pose3, Pose3, Vector6> { |
| 28 | |
| 29 | double h_; // time step |
| 30 | typedef NoiseModelFactorN<Pose3, Pose3, Vector6> Base; |
| 31 | public: |
| 32 | |
| 33 | // Provide access to the Matrix& version of evaluateError: |
| 34 | using Base::evaluateError; |
| 35 | |
| 36 | Reconstruction(Key gKey1, Key gKey, Key xiKey, double h, double mu = 1000.0) : |
| 37 | Base(noiseModel::Constrained::All(dim: 6, mu: std::abs(x: mu)), gKey1, gKey, |
| 38 | xiKey), h_(h) { |
| 39 | } |
| 40 | ~Reconstruction() override {} |
| 41 | |
| 42 | /// @return a deep copy of this factor |
| 43 | gtsam::NonlinearFactor::shared_ptr clone() const override { |
| 44 | return std::static_pointer_cast<gtsam::NonlinearFactor>( |
| 45 | r: gtsam::NonlinearFactor::shared_ptr(new Reconstruction(*this))); } |
| 46 | |
| 47 | /** \f$ log((g_k\exp(h\xi_k))^{-1}g_{k+1}) = 0 \f$, with optional derivatives */ |
| 48 | Vector evaluateError(const Pose3& gk1, const Pose3& gk, const Vector6& xik, |
| 49 | OptionalMatrixType H1, OptionalMatrixType H2, |
| 50 | OptionalMatrixType H3) const override { |
| 51 | |
| 52 | Matrix6 D_exphxi_xi; |
| 53 | Pose3 exphxi = Pose3::Expmap(xi: h_ * xik, Hxi: H3 ? &D_exphxi_xi : 0); |
| 54 | |
| 55 | Matrix6 D_gkxi_gk, D_gkxi_exphxi; |
| 56 | Pose3 gkxi = gk.compose(g: exphxi, H1: D_gkxi_gk, H2: H3 ? &D_gkxi_exphxi : 0); |
| 57 | |
| 58 | Matrix6 D_hx_gk1, D_hx_gkxi; |
| 59 | Pose3 hx = gkxi.between(g: gk1, H1: (H2 || H3) ? &D_hx_gkxi : 0, H2: H1 ? &D_hx_gk1 : 0); |
| 60 | |
| 61 | Matrix6 D_log_hx; |
| 62 | Vector error = Pose3::Logmap(pose: hx, Hpose: D_log_hx); |
| 63 | |
| 64 | if (H1) *H1 = D_log_hx * D_hx_gk1; |
| 65 | if (H2 || H3) { |
| 66 | Matrix6 D_log_gkxi = D_log_hx * D_hx_gkxi; |
| 67 | if (H2) *H2 = D_log_gkxi * D_gkxi_gk; |
| 68 | if (H3) *H3 = D_log_gkxi * D_gkxi_exphxi * D_exphxi_xi * h_; |
| 69 | } |
| 70 | |
| 71 | return error; |
| 72 | } |
| 73 | |
| 74 | }; |
| 75 | |
| 76 | /** |
| 77 | * Implement the Discrete Euler-Poincare' equation: |
| 78 | */ |
| 79 | class DiscreteEulerPoincareHelicopter : public NoiseModelFactorN<Vector6, Vector6, Pose3> { |
| 80 | |
| 81 | double h_; /// time step |
| 82 | Matrix Inertia_; /// Inertia tensors Inertia = [ J 0; 0 M ] |
| 83 | Vector Fu_; /// F is the 6xc Control matrix, where c is the number of control variables uk, which directly change the vehicle pose (e.g., gas/brake/speed) |
| 84 | /// F(.) is actually a function of the shape variables, which do not change the pose, but affect the vehicle's shape, e.g. steering wheel. |
| 85 | /// Fu_ encodes everything we need to know about the vehicle's dynamics. |
| 86 | double m_; /// mass. For gravity external force f_ext, which has a fixed formula in this case. |
| 87 | |
| 88 | // TODO: Fk_ and f_ext should be generalized as functions (factor nodes) on control signals and poses/velocities. |
| 89 | // This might be needed in control or system identification problems. |
| 90 | // We treat them as constant here, since the control inputs are to specify. |
| 91 | |
| 92 | typedef NoiseModelFactorN<Vector6, Vector6, Pose3> Base; |
| 93 | |
| 94 | public: |
| 95 | |
| 96 | // Provide access to the Matrix& version of evaluateError: |
| 97 | using Base::evaluateError; |
| 98 | |
| 99 | DiscreteEulerPoincareHelicopter(Key xiKey1, Key xiKey_1, Key gKey, |
| 100 | double h, const Matrix& Inertia, const Vector& Fu, double m, |
| 101 | double mu = 1000.0) : |
| 102 | Base(noiseModel::Constrained::All(dim: 6, mu: std::abs(x: mu)), xiKey1, xiKey_1, gKey), |
| 103 | h_(h), Inertia_(Inertia), Fu_(Fu), m_(m) { |
| 104 | } |
| 105 | ~DiscreteEulerPoincareHelicopter() override {} |
| 106 | |
| 107 | /// @return a deep copy of this factor |
| 108 | gtsam::NonlinearFactor::shared_ptr clone() const override { |
| 109 | return std::static_pointer_cast<gtsam::NonlinearFactor>( |
| 110 | r: gtsam::NonlinearFactor::shared_ptr(new DiscreteEulerPoincareHelicopter(*this))); } |
| 111 | |
| 112 | /** DEP, with optional derivatives |
| 113 | * pk - pk_1 - h_*Fu_ - h_*f_ext = 0 |
| 114 | * where pk = CT_TLN(h*xi_k)*Inertia*xi_k |
| 115 | * pk_1 = CT_TLN(-h*xi_k_1)*Inertia*xi_k_1 |
| 116 | * */ |
| 117 | Vector evaluateError(const Vector6& xik, const Vector6& xik_1, const Pose3& gk, |
| 118 | OptionalMatrixType H1, OptionalMatrixType H2, |
| 119 | OptionalMatrixType H3) const override { |
| 120 | |
| 121 | Vector muk = Inertia_*xik; |
| 122 | Vector muk_1 = Inertia_*xik_1; |
| 123 | |
| 124 | // Apply the inverse right-trivialized tangent (derivative) map of the exponential map, |
| 125 | // using the trapezoidal Lie-Newmark (TLN) scheme, to a vector. |
| 126 | // TLN is just a first order approximation of the dExpInv_exp above, detailed in [Kobilarov09siggraph] |
| 127 | // C_TLN formula: I6 - 1/2 ad[xi]. |
| 128 | Matrix D_adjThxik_muk, D_adjThxik1_muk1; |
| 129 | Vector pk = muk - 0.5*Pose3::adjointTranspose(xi: h_*xik, y: muk, Hxi: D_adjThxik_muk); |
| 130 | Vector pk_1 = muk_1 - 0.5*Pose3::adjointTranspose(xi: -h_*xik_1, y: muk_1, Hxi: D_adjThxik1_muk1); |
| 131 | |
| 132 | Matrix D_gravityBody_gk; |
| 133 | Point3 gravityBody = gk.rotation().unrotate(p: Point3(0.0, 0.0, -9.81*m_), H1: D_gravityBody_gk, H2: {}); |
| 134 | Vector f_ext = (Vector(6) << 0.0, 0.0, 0.0, gravityBody.x(), gravityBody.y(), gravityBody.z()).finished(); |
| 135 | |
| 136 | Vector hx = pk - pk_1 - h_*Fu_ - h_*f_ext; |
| 137 | |
| 138 | if (H1) { |
| 139 | Matrix D_pik_xi = Inertia_-0.5*(h_*D_adjThxik_muk + Pose3::adjointMap(xi: h_*xik).transpose()*Inertia_); |
| 140 | *H1 = D_pik_xi; |
| 141 | } |
| 142 | |
| 143 | if (H2) { |
| 144 | Matrix D_pik1_xik1 = Inertia_-0.5*(-h_*D_adjThxik1_muk1 + Pose3::adjointMap(xi: -h_*xik_1).transpose()*Inertia_); |
| 145 | *H2 = -D_pik1_xik1; |
| 146 | } |
| 147 | |
| 148 | if (H3) { |
| 149 | *H3 = Z_6x6; |
| 150 | insertSub(fullMatrix&: *H3, subMatrix: -h_*D_gravityBody_gk, i: 3, j: 0); |
| 151 | } |
| 152 | |
| 153 | return hx; |
| 154 | } |
| 155 | |
| 156 | #if 0 |
| 157 | Vector computeError(const Vector6& xik, const Vector6& xik_1, const Pose3& gk) const { |
| 158 | Vector pk = Pose3::dExpInv_exp(h_*xik).transpose()*Inertia_*xik; |
| 159 | Vector pk_1 = Pose3::dExpInv_exp(-h_*xik_1).transpose()*Inertia_*xik_1; |
| 160 | |
| 161 | Point3 gravityBody = gk.rotation().unrotate(Point3(0.0, 0.0, -9.81*m_)); |
| 162 | Vector f_ext = (Vector(6) << 0.0, 0.0, 0.0, gravityBody.x(), gravityBody.y(), gravityBody.z()); |
| 163 | |
| 164 | Vector hx = pk - pk_1 - h_*Fu_ - h_*f_ext; |
| 165 | |
| 166 | return hx; |
| 167 | } |
| 168 | |
| 169 | Vector evaluateError(const Vector6& xik, const Vector6& xik_1, const Pose3& gk, |
| 170 | OptionalMatrixType H1, OptionalMatrixType H2, |
| 171 | OptionalMatrixType H3) const { |
| 172 | if (H1) { |
| 173 | (*H1) = numericalDerivative31( |
| 174 | std::function<Vector(const Vector6&, const Vector6&, const Pose3&)>( |
| 175 | std::bind(&DiscreteEulerPoincareHelicopter::computeError, *this, _1, _2, _3) |
| 176 | ), |
| 177 | xik, xik_1, gk, 1e-5 |
| 178 | ); |
| 179 | } |
| 180 | if (H2) { |
| 181 | (*H2) = numericalDerivative32( |
| 182 | std::function<Vector(const Vector6&, const Vector6&, const Pose3&)>( |
| 183 | std::bind(&DiscreteEulerPoincareHelicopter::computeError, *this, _1, _2, _3) |
| 184 | ), |
| 185 | xik, xik_1, gk, 1e-5 |
| 186 | ); |
| 187 | } |
| 188 | if (H3) { |
| 189 | (*H3) = numericalDerivative33( |
| 190 | std::function<Vector(const Vector6&, const Vector6&, const Pose3&)>( |
| 191 | std::bind(&DiscreteEulerPoincareHelicopter::computeError, *this, _1, _2, _3) |
| 192 | ), |
| 193 | xik, xik_1, gk, 1e-5 |
| 194 | ); |
| 195 | } |
| 196 | |
| 197 | return computeError(xik, xik_1, gk); |
| 198 | } |
| 199 | #endif |
| 200 | |
| 201 | }; |
| 202 | |
| 203 | } // namespace gtsam |
| 204 | |