Program Listing for File log.hxx

Return to documentation for file (include/pinocchio/spatial/log.hxx)

//
// Copyright (c) 2015-2021 CNRS INRIA
//

#ifndef __pinocchio_spatial_log_hxx__
#define __pinocchio_spatial_log_hxx__

namespace pinocchio
{

  namespace internal
  {
    template<long i0, typename Matrix3, typename Vector3>
    void compute_theta_axis(
      const typename Matrix3::Scalar & val,
      const Eigen::MatrixBase<Matrix3> & R,
      typename Matrix3::Scalar & angle,
      const Eigen::MatrixBase<Vector3> & _axis)
    {
      typedef typename Matrix3::Scalar Scalar;
      static const Scalar eps = Eigen::NumTraits<Scalar>::epsilon();

      static const long i1 = (i0 + 1) % 3;
      static const long i2 = (i0 + 2) % 3;
      Vector3 & axis = _axis.const_cast_derived();

      const Scalar s =
        math::sqrt(val + eps + eps * eps)
        * if_then_else(
          GE, R.coeff(i2, i1), R.coeff(i1, i2), Scalar(1.),
          Scalar(-1.)); // Ensure value in sqrt is non negative and that s is non zero
      axis[i0] = s / Scalar(2);
      axis[i1] = Scalar(1) / (2 * s) * (R.coeff(i1, i0) + R.coeff(i0, i1));
      axis[i2] = Scalar(1) / (2 * s) * (R.coeff(i2, i0) + R.coeff(i0, i2));
      const Scalar w = Scalar(1) / (2 * s) * (R.coeff(i2, i1) - R.coeff(i1, i2));

      const Scalar axis_norm = axis.norm();
      angle = 2 * math::atan2(axis_norm, w);
      axis /= axis_norm;
    }
  } // namespace internal

  template<typename Matrix3>
  inline typename PINOCCHIO_EIGEN_PLAIN_TYPE(Matrix3)
    renormalize_rotation_matrix(const Eigen::MatrixBase<Matrix3> & R)
  {
    typename PINOCCHIO_EIGEN_PLAIN_TYPE(Matrix3) Rout;
    Rout.col(0).noalias() = R.col(0) / R.col(0).norm();
    Rout.col(1).noalias() = R.col(1) / R.col(1).norm();
    Rout.col(2).noalias() = Rout.col(0).cross(Rout.col(1));
    Rout.col(0).noalias() = Rout.col(1).cross(Rout.col(2));
    return Rout;
  }

  template<typename _Scalar>
  struct log3_impl
  {
    template<typename Matrix3Like, typename Vector3Out>
    static void run(
      const Eigen::MatrixBase<Matrix3Like> & R,
      typename Matrix3Like::Scalar & theta,
      const Eigen::MatrixBase<Vector3Out> & angle_axis)
    {
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix3Like, R, 3, 3);
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Vector3Out, angle_axis, 3, 1);
      using namespace internal;

      typedef typename Matrix3Like::Scalar Scalar;
      typedef Eigen::Matrix<Scalar, 3, 1, PINOCCHIO_EIGEN_PLAIN_TYPE(Matrix3Like)::Options> Vector3;
      static const Scalar eps = Eigen::NumTraits<Scalar>::epsilon();

      const static Scalar PI_value = PI<Scalar>();
      Vector3Out & angle_axis_ = angle_axis.const_cast_derived();

      typedef typename PINOCCHIO_EIGEN_PLAIN_TYPE(Matrix3Like) Matrix3;
      const Matrix3 Rnormed = renormalize_rotation_matrix(R);

      const Scalar tr = Rnormed.trace();
      const Scalar cos_value = (tr - Scalar(1)) / Scalar(2);

      const Scalar prec = TaylorSeriesExpansion<Scalar>::template precision<2>();
      // Singular cases when theta == PI
      Vector3 angle_axis_singular;
      Scalar theta_singular;

      {
        Vector3 val_singular;
        val_singular.array() = 2 * Rnormed.diagonal().array() - tr + Scalar(1);
        Vector3 axis_0, axis_1, axis_2;
        Scalar theta_0, theta_1, theta_2;

        internal::compute_theta_axis<0>(val_singular[0], Rnormed, theta_0, axis_0);
        internal::compute_theta_axis<1>(val_singular[1], Rnormed, theta_1, axis_1);
        internal::compute_theta_axis<2>(val_singular[2], Rnormed, theta_2, axis_2);

        theta_singular = if_then_else(
          GE, val_singular[0], val_singular[1],
          if_then_else(GE, val_singular[0], val_singular[2], theta_0, theta_2),
          if_then_else(GE, val_singular[1], val_singular[2], theta_1, theta_2));

        for (int k = 0; k < 3; ++k)
          angle_axis_singular[k] = if_then_else(
            GE, val_singular[0], val_singular[1],
            if_then_else(GE, val_singular[0], val_singular[2], axis_0[k], axis_2[k]),
            if_then_else(GE, val_singular[1], val_singular[2], axis_1[k], axis_2[k]));
      }
      const Scalar acos_expansion = math::sqrt(2 * (1 - cos_value) + eps * eps);
      const Scalar theta_nominal = if_then_else(
        LE, tr, static_cast<Scalar>(Scalar(3) - prec),
        if_then_else(
          GE, tr, static_cast<Scalar>(Scalar(-1) + prec),
          math::acos(cos_value),                         // then
          static_cast<Scalar>(PI_value - acos_expansion) // else
          ),
        static_cast<Scalar>(acos_expansion) // else
      );
      assert(
        check_expression_if_real<Scalar>(theta_nominal == theta_nominal)
        && "theta contains some NaN"); // theta != NaN

      Vector3 antisymmetric_R;
      unSkew(Rnormed, antisymmetric_R);
      const Scalar norm_antisymmetric_R_squared = antisymmetric_R.squaredNorm();

      const Scalar t = if_then_else(
        GE, theta_nominal, prec,
        static_cast<Scalar>(theta_nominal / sin(theta_nominal)), // then
        static_cast<Scalar>(
          Scalar(1.) + norm_antisymmetric_R_squared / Scalar(6)
          + norm_antisymmetric_R_squared * norm_antisymmetric_R_squared * Scalar(3)
              / Scalar(40)) // else
      );

      theta = if_then_else(
        GE, cos_value, static_cast<Scalar>(Scalar(-1.) + prec), theta_nominal, theta_singular);

      for (int k = 0; k < 3; ++k)
        angle_axis_[k] = if_then_else(
          GE, cos_value, static_cast<Scalar>(Scalar(-1.) + prec),
          static_cast<Scalar>(t * antisymmetric_R[k]),
          static_cast<Scalar>(theta_singular * angle_axis_singular[k]));
    }
  };

  template<typename _Scalar>
  struct Jlog3_impl
  {
    template<typename Scalar, typename Vector3Like, typename Matrix3Like>
    static void run(
      const Scalar & theta,
      const Eigen::MatrixBase<Vector3Like> & log,
      const Eigen::MatrixBase<Matrix3Like> & Jlog)
    {
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Vector3Like, log, 3, 1);
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix3Like, Jlog, 3, 3);

      using namespace internal;
      Scalar ct, st;
      SINCOS(theta, &st, &ct);
      const Scalar st_1mct = st / (Scalar(1) - ct);
      const Scalar prec = TaylorSeriesExpansion<Scalar>::template precision<3>();

      const Scalar alpha = if_then_else(
        LT, theta, prec,
        static_cast<Scalar>(Scalar(1) / Scalar(12) + theta * theta / Scalar(720)),       // then
        static_cast<Scalar>(Scalar(1) / (theta * theta) - st_1mct / (Scalar(2) * theta)) // else
      );

      const Scalar diag_value = if_then_else(
        LT, theta, prec,
        static_cast<Scalar>(Scalar(0.5) * (2 - theta * theta / Scalar(6))), // then
        static_cast<Scalar>(Scalar(0.5) * (theta * st_1mct))                // else
      );

      Matrix3Like & Jlog_ = PINOCCHIO_EIGEN_CONST_CAST(Matrix3Like, Jlog);
      Jlog_.noalias() = alpha * log * log.transpose();
      Jlog_.diagonal().array() += diag_value;

      // Jlog += r_{\times}/2
      addSkew(Scalar(0.5) * log, Jlog_);
    }
  };

  template<typename _Scalar>
  struct log6_impl
  {
    template<typename Scalar, int Options, typename MotionDerived>
    static void run(const SE3Tpl<Scalar, Options> & M, MotionDense<MotionDerived> & mout)
    {
      typedef SE3Tpl<Scalar, Options> SE3;
      typedef typename SE3::Vector3 Vector3;

      typename SE3::ConstAngularRef R = M.rotation();
      typename SE3::ConstLinearRef p = M.translation();

      using namespace internal;

      Vector3 antisymmetric_R;
      unSkew(R, antisymmetric_R);
      const Scalar norm_antisymmetric_R_squared = antisymmetric_R.squaredNorm();

      Scalar theta;
      const Scalar tr = R.trace();
      const Vector3 w(log3(R, theta)); // t in [0,π]
      const Scalar & t2 = norm_antisymmetric_R_squared;

      Scalar st, ct;
      SINCOS(theta, &st, &ct);
      const Scalar alpha = if_then_else(
        GE, tr,
        static_cast<Scalar>(Scalar(3) - TaylorSeriesExpansion<Scalar>::template precision<2>()),
        static_cast<Scalar>(Scalar(1) - t2 / Scalar(12) - t2 * t2 / Scalar(720)), // then
        static_cast<Scalar>(theta * st / (Scalar(2) * (Scalar(1) - ct)))          // else
      );

      const Scalar beta = if_then_else(
        GE, tr,
        static_cast<Scalar>(Scalar(3) - TaylorSeriesExpansion<Scalar>::template precision<2>()),
        static_cast<Scalar>(Scalar(1) / Scalar(12) + t2 / Scalar(720)), // then
        static_cast<Scalar>(
          Scalar(1) / (theta * theta) - st / (Scalar(2) * theta * (Scalar(1) - ct))) // else
      );

      mout.linear().noalias() = alpha * p - Scalar(0.5) * w.cross(p) + (beta * w.dot(p)) * w;
      mout.angular() = w;
    }

    template<typename Vector3Like, typename QuaternionLike, typename MotionDerived>
    static void run(
      const Eigen::QuaternionBase<QuaternionLike> & quat,
      const Eigen::MatrixBase<Vector3Like> & vec,
      MotionDense<MotionDerived> & mout)
    {
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Vector3Like, vec, 3, 1);

      typedef typename Vector3Like::Scalar Scalar;
      enum
      {
        Options = PINOCCHIO_EIGEN_PLAIN_TYPE(Vector3Like)::Options
      };
      typedef Eigen::Matrix<Scalar, 3, 1, Options> Vector3;
      const Scalar eps = Eigen::NumTraits<Scalar>::epsilon();

      using namespace internal;

      const Scalar pos_neg = if_then_else(GE, quat.w(), Scalar(0), Scalar(+1), Scalar(-1));

      Scalar theta;
      Vector3 w(quaternion::log3(quat, theta)); // theta nonsingular by construction
      const Scalar t2 = w.squaredNorm();

      // Scalar st,ct; SINCOS(theta,&st,&ct);
      Scalar st_2, ct_2;
      ct_2 = pos_neg * quat.w();
      st_2 = math::sqrt(quat.vec().squaredNorm() + eps * eps);
      const Scalar cot_th_2 = ct_2 / st_2;
      // const Scalar cot_th_2 = ( st / (Scalar(1) - ct) ); // cotan of half angle

      // we use formula (9.26) from
      // https://ingmec.ual.es/~jlblanco/papers/jlblanco2010geometry3D_techrep.pdf for the linear
      // part of the Log map. A Taylor series expansion of cotan can be used up to order 4
      const Scalar th_2_squared = t2 / Scalar(4); // (theta / 2) squared

      //      const Scalar alpha = if_then_else(LE,theta,TaylorSeriesExpansion<Scalar>::template
      //      precision<3>(),
      //                                        static_cast<Scalar>(Scalar(1) - t2/Scalar(12) -
      //                                        t2*t2/Scalar(720)), // then
      //                                        static_cast<Scalar>(theta * cot_th_2 /(Scalar(2)))
      //                                        // else
      //                                        );

      const Scalar beta_alt = (Scalar(1) / Scalar(3) - th_2_squared / Scalar(45)) / Scalar(4);
      const Scalar beta = if_then_else(
        LE, theta, TaylorSeriesExpansion<Scalar>::template precision<3>(),
        static_cast<Scalar>(beta_alt),                                       // then
        static_cast<Scalar>(Scalar(1) / t2 - cot_th_2 * Scalar(0.5) / theta) // else
        // static_cast<Scalar>(Scalar(1) / t2 - st/(Scalar(2)*theta*(Scalar(1)-ct))) // else
      );

      // mout.linear().noalias() = alpha * vec - Scalar(0.5) * w.cross(vec) + (beta * w.dot(vec)) *
      // w;
      mout.linear().noalias() = vec - Scalar(0.5) * w.cross(vec) + beta * w.cross(w.cross(vec));
      mout.angular() = w;
    }
  };

  template<typename _Scalar>
  struct Jlog6_impl
  {
    template<typename Scalar, int Options, typename Matrix6Like>
    static void run(const SE3Tpl<Scalar, Options> & M, const Eigen::MatrixBase<Matrix6Like> & Jlog)
    {
      PINOCCHIO_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix6Like, Jlog, 6, 6);

      typedef SE3Tpl<Scalar, Options> SE3;
      typedef typename SE3::Vector3 Vector3;
      Matrix6Like & value = PINOCCHIO_EIGEN_CONST_CAST(Matrix6Like, Jlog);

      typename SE3::ConstAngularRef R = M.rotation();
      typename SE3::ConstLinearRef p = M.translation();

      using namespace internal;

      Scalar t;
      Vector3 w(log3(R, t));

      // value is decomposed as following:
      // value = [ A, B;
      //           C, D ]
      typedef Eigen::Block<Matrix6Like, 3, 3> Block33;
      Block33 A = value.template topLeftCorner<3, 3>();
      Block33 B = value.template topRightCorner<3, 3>();
      Block33 C = value.template bottomLeftCorner<3, 3>();
      Block33 D = value.template bottomRightCorner<3, 3>();

      Jlog3(t, w, A);
      D = A;

      const Scalar t2 = t * t;
      const Scalar tinv = Scalar(1) / t, t2inv = tinv * tinv;

      Scalar st, ct;
      SINCOS(t, &st, &ct);
      const Scalar inv_2_2ct = Scalar(1) / (Scalar(2) * (Scalar(1) - ct));

      const Scalar beta = if_then_else(
        LT, t, TaylorSeriesExpansion<Scalar>::template precision<3>(),
        static_cast<Scalar>(Scalar(1) / Scalar(12) + t2 / Scalar(720)), // then
        static_cast<Scalar>(t2inv - st * tinv * inv_2_2ct)              // else
      );

      const Scalar beta_dot_over_theta = if_then_else(
        LT, t, TaylorSeriesExpansion<Scalar>::template precision<3>(),
        static_cast<Scalar>(Scalar(1) / Scalar(360)), // then
        static_cast<Scalar>(
          -Scalar(2) * t2inv * t2inv + (Scalar(1) + st * tinv) * t2inv * inv_2_2ct) // else
      );

      const Scalar wTp = w.dot(p);
      const Vector3 v3_tmp(
        (beta_dot_over_theta * wTp) * w - (t2 * beta_dot_over_theta + Scalar(2) * beta) * p);
      // C can be treated as a temporary variable
      C.noalias() = v3_tmp * w.transpose();
      C.noalias() += beta * w * p.transpose();
      C.diagonal().array() += wTp * beta;
      addSkew(Scalar(.5) * p, C);

      B.noalias() = C * A;
      C.setZero();
    }
  };

} // namespace pinocchio

#endif // ifndef __pinocchio_spatial_log_hxx__