零、任务介绍
- 补全
src/ros-bridge/carla_shenlan_projects/carla_shenlan_lqr_pid_controller/src/lqr_controller.cpp
中的TODO
部分,实现基于LQR的车辆横向控制。
一、系统建模
1.1 连续模型
选取系统状态变量为 x = [ e c g , e ˙ c g , e θ , e ˙ θ ] ⊤ x = [e_{cg}, \dot e_{cg}, e_\theta, \dot e_\theta]^\top x=[ecg,e˙cg,eθ,e˙θ]⊤,控制输入为 u = δ u = \delta u=δ。系统参考轨迹每个点可以给出位置、偏航角、速度、加速度、曲率,这些量可以表示为沿轨迹行驶路程 s s s的函数。车辆的横向误差动力学可以用如下的线性模型描述
x ˙ = A x + B 1 δ + B 2 r d e s \dot x = Ax + B_1\delta + B_2 r_{des} x˙=Ax+B1δ+B2rdes
其中
A = [ 0 1 0 0 0 − ( c f + c r ) m v c f + c r m l r c r − l f c f m v 0 0 0 1 0 l r c r − l f c f I z v l r c r − l f c f I z − ( l f 2 c f + l r 2 c r ) I z v ] A = \left[\begin{matrix} 0 & 1 & 0 & 0\\ 0 & -\frac{(c_f + c_r)}{mv} & \frac{c_f + c_r}{m} & \frac{l_r c_r - l_f c_f}{mv}\\ 0 & 0 & 0 & 1\\ 0 & \frac{l_r c_r - l_f c_f}{I_z v} & \frac{l_r c_r - l_f c_f}{I_z} & -\frac{(l_f^2 c_f + l_r^2 c_r)}{I_z v} \end{matrix}\right] A=
00001−mv(cf+cr)0Izvlrcr−lfcf0mcf+cr0Izlrcr−lfcf0mvlrcr−lfcf1−Izv(lf2cf+lr2cr)
B 1 = [ 0 c f / m 0 l f c f / I z ] B_1 = \left[\begin{matrix} 0\\ c_f / m\\ 0\\ l_f c_f /I_z \end{matrix}\right] B1= 0cf/m0lfcf/Iz
B 2 = [ 0 ( l r c r − l f c f ) / ( m v ) − v 0 − ( l f 2 c f + l r 2 c r ) / ( I z v ) ] B_2 = \left[ \begin{matrix} 0\\ (l_r c_r - l_f c_f)/(mv) - v\\ 0\\ -(l_f^2 c_f + l_r^2 c_r)/(I_z v) \end{matrix} \right] B2=
0(lrcr−lfcf)/(mv)−v0−(lf2cf+lr2cr)/(Izv)
这里实际上是假设在一小段时间内车辆纵向速度 v v v保持不变,然后在工作点线性化,参考轨迹的偏航角速度 r d e s r_{des} rdes可以由参考点速度 v d e s v_{des} vdes和参考点曲率 κ d e s \kappa_{des} κdes计算
r d e s = v d e s κ d e s r_{des} = v_{des} \kappa_{des} rdes=vdesκdes
1.2 离散化
实际实现时,需要将系统离散化,离散化可以采取前向欧拉法和零阶保持器精确离散化等方法。采样周期为 T T T,前向欧拉法得到的离散系统矩阵和输入矩阵为
A d = I + T A B d 1 = T B 1 B d 2 = T B 2 \begin{aligned} A_d &= I + TA \\ B_{d1} &= TB_1 \\ B_{d2} &= TB_2 \end{aligned} AdBd1Bd2=I+TA=TB1=TB2
为了达到更好的准确性和稳定性,可以使用零阶保持器进行精确离散化,精确离散化的计算公式如下:
A d = e A T B d 1 = ( ∫ 0 T e A τ d τ ) B 1 B d 2 = ( ∫ 0 T e A τ d τ ) B 2 \begin{aligned} A_d &= e^{AT}\\ B_{d1} &= (\int_0^T e^{A\tau} d\tau)B_1 \\ B_{d2} &= (\int_0^T e^{A\tau} d\tau)B_2 \end{aligned} AdBd1Bd2=eAT=(∫0TeAτdτ)B1=(∫0TeAτdτ)B2
矩阵指数可以用如下的拉普拉斯逆变换得到
e A t = L − 1 [ ( s I − A ) − 1 ] e^{At} = \mathcal L^{-1}[(sI - A)^{-1}] eAt=L−1[(sI−A)−1]
本次project中的参数如下
参数名称 | 符号 | 数值 |
---|---|---|
前轮侧偏刚度(左右轮之和) | c f c_f cf | 155494.663 |
后轮侧偏刚度(左右轮之和) | c r c_r cr | 155494.663 |
前轮到质心的距离 | l f l_f lf | 1.426 |
后轮到质心的距离 | l r l_r lr | 1.426 |
整车质量 | m m m | 1845.0 |
Z轴转动惯量 | I z I_z Iz | 3751.76 |
控制周期 | t s t_s ts | 0.01 |
使用mathematica计算矩阵拉普拉斯反变换
ts = 0.01;
cf = 155494.663;
cr = 155494.663;
mfl = 1845.0/4;
mfr = 1845.0/4;
mrl = 1845.0/4;
mrr = 1845.0/4;
mf = mfl + mfr;
mr = mrl + mrr;
m = mf + mr
wheelbase = 2.852;
lf = wheelbase * (1.0 - mf/m)
lr = wheelbase * (1.0 - mr /m)
iz = lf*lf*mf + lr*lr*mr
A = {
{0, 1, 0, 0},
{0, -(cf + cr)/(m*v), (cf + cr)/m, (lr*cr - lf*cf)/(m*v)},
{0, 0, 0, 1},
{0, (lr*cr - lf*cf)/(iz * v), (lr*cr - lf*cf)/
iz, -(lf^2*cf + lr^2*cr)/(iz * v)}
}
InverseLaplaceTransform[Inverse[s*IdentityMatrix[4] - A], s, 0.01]
计算得到的矩阵 A d A_d Ad如下
A d = ( 1.0 v ( 0.00593268 − 0.00593268 e − 1.68558 / v ) v ( ( 0.00593268 e − 1.68558 / v − 0.00593268 ) v + 0.01 ) 168.558 ( e − 1.68558 / v ( 4.176213162156246 × 1 0 − 7 v + 3.519668608485906 × 1 0 − 7 ) − 4.176213162156246 × 1 0 − 7 v + 3.519668608485956 × 1 0 − 7 ) v 2 0. 1. e − 1.68558 / v v ( 1. − 1. e − 1.68558 / v ) 168.558 ( e − 1.68558 / v ( − 0.0000351967 v − 0.0000593268 ) + 0.0000351967 v ) v 0. 0. 1. e − 1.68558 / v ( − 0.00593268 v − 3.3723378595376876 × 1 0 − 18 ) + 0.00593268 v 0. 0. 0. e − 1.68558 / v ( 1. v + 2.842170943040401 × 1 0 − 16 ) v ) A_d = \left( \begin{array}{cccc} 1.0 & v \left(0.00593268 - 0.00593268 e^{-1.68558/v}\right) & v \left(\left(0.00593268 e^{-1.68558/v}-0.00593268\right) v + 0.01\right) & 168.558 \left(e^{-1.68558/v} \left(4.176213162156246 \times 10^{-7} v + 3.519668608485906 \times 10^{-7}\right) - 4.176213162156246 \times 10^{-7} v + 3.519668608485956 \times 10^{-7}\right) v^2 \\ 0. & 1. e^{-1.68558/v} & v \left(1. - 1. e^{-1.68558/v}\right) & 168.558 \left(e^{-1.68558/v} \left(-0.0000351967 v - 0.0000593268\right) + 0.0000351967 v\right) v \\ 0. & 0. & 1. & e^{-1.68558/v} \left(-0.00593268 v - 3.3723378595376876 \times 10^{-18}\right) + 0.00593268 v \\ 0. & 0. & 0. & \frac{e^{-1.68558/v} \left(1. v + 2.842170943040401 \times 10^{-16}\right)}{v} \end{array} \right) Ad=
1.00.0.0.v(0.00593268−0.00593268e−1.68558/v)1.e−1.68558/v0.0.v((0.00593268e−1.68558/v−0.00593268)v+0.01)v(1.−1.e−1.68558/v)1.0.168.558(e−1.68558/v(4.176213162156246×10−7v+3.519668608485906×10−7)−4.176213162156246×10−7v+3.519668608485956×10−7)v2168.558(e−1.68558/v(−0.0000351967v−0.0000593268)+0.0000351967v)ve−1.68558/v(−0.00593268v−3.3723378595376876×10−18)+0.00593268vve−1.68558/v(1.v+2.842170943040401×10−16)
通过观察可以得到,矩阵 A A A不可逆,因此计算 B d B_d Bd时采用幂级数展开求 e A τ e^{A\tau} eAτ的积分,即
B d 1 = ∫ 0 T e A τ d τ B 1 = ∫ 0 T ( I + A τ + ( A τ ) 2 2 ! + ( A τ ) 3 3 ! + ⋯ ) d τ B 1 = ( T I + T 2 2 A + T 3 6 A 2 + ⋯ ) B 1 ≈ T B 1 \begin{aligned} B_{d1} &= \int_0^T e^{A\tau} d\tau B_1 \\ &= \int_0^T(I + A\tau + \frac{(A\tau)^2}{2!} + \frac{(A\tau)^3}{3!} + \cdots)d\tau B_1 \\ & = (TI + \frac{T^2}{2}A + \frac{T^3}{6}A^2 + \cdots) B_1 \\ &\approx TB_1 \end{aligned} Bd1=∫0TeAτdτB1=∫0T(I+Aτ+2!(Aτ)2+3!(Aτ)3+⋯)dτB1=(TI+2T2A+6T3A2+⋯)B1≈TB1
类似地有 B d 2 = T B 2 B_{d2} = TB_2 Bd2=TB2。
二、算法
2.1 离散时间LQR
离散时间LQR的最优控制律为
u ∗ ( k ) = − K x ( k ) u^*(k) = - K x(k) u∗(k)=−Kx(k)
其中反馈增益 K K K为
K = ( R + B d ⊤ P B d ) − 1 B d ⊤ P A d K = (R + B_d^\top P B_d)^{-1} B_d^\top P A_d K=(R+Bd⊤PBd)−1Bd⊤PAd
其中正定矩阵P可以使用动态规划算法得出,迭代形式如下
P t − 1 = A d ⊤ P t A d − A d ⊤ P t B d ( R + B d ⊤ P t B d ) − 1 B d ⊤ P t A d + Q P_{t-1} = A_d^\top P_{t} A_d - A_d^\top P_{t} B_d (R + B_d^\top P_{t} B_d)^{-1}B_d^\top P_{t} A_d + Q Pt−1=Ad⊤PtAd−Ad⊤PtBd(R+Bd⊤PtBd)−1Bd⊤PtAd+Q
此处的推导细节参考基础算法 - LQR - 离散时间有限边界
2.2 前馈控制
采用上述状态反馈控制律后,系统的模型可以改写成
x ˙ = ( A − B 1 K ) x + B 2 r d e s \dot x = (A - B_1 K) x + B_2 r_{des} x˙=(A−B1K)x+B2rdes
闭环系统 A − B 1 K A - B_1K A−B1K是渐进稳定的,但是由于 B 2 r d e s B_2 r_{des} B2rdes的存在,无法保证系统状态量收敛至0。因此需要在控制律中加入前馈项
δ = − K x + δ f f \delta = -Kx + \delta_{ff} δ=−Kx+δff
前馈项的设计参考自动驾驶控制算法 —— 横向LQR控制+前馈控制
δ f f = L κ + K v a y − k 3 κ ( l r − l f m v x 2 2 c α r L ) \begin{aligned} \delta_{ff} = L \kappa + K_v a_y - k_3 \kappa (l_r - \frac{l_f mv_x^2}{2c_{\alpha r} L}) \end{aligned} δff=Lκ+Kvay−k3κ(lr−2cαrLlfmvx2)
其中
K v = l r m 2 c α f ( l f + l r ) − l f m 2 c α r ( l f + l r ) K_v = \frac{l_r m}{2 c_{\alpha f} (l_f + l_r)} - \frac{l_f m}{2 c_{\alpha r} (l_f + l_r)} Kv=2cαf(lf+lr)lrm−2cαr(lf+lr)lfm
三、代码实现
计算控制量
bool LqrController::ComputeControlCommand(const VehicleState &localization, const TrajectoryData &planning_published_trajectory, ControlCmd &cmd) {
// 规划轨迹
trajectory_points_ = planning_published_trajectory.trajectory_points;
/*
A matrix (Gear Drive)
[0.0, 1.0, 0.0, 0.0;
0.0, (-(c_f + c_r) / m) / v, (c_f + c_r) / m, (l_r * c_r - l_f * c_f) / m / v;
0.0, 0.0, 0.0, 1.0;
0.0, ((lr * cr - lf * cf) / i_z) / v, (l_f * c_f - l_r * c_r) / i_z, (-1.0 * (l_f^2 * c_f + l_r^2 * c_r) / i_z) / v;]
*/
// TODO 01 配置状态矩阵A
double v = max(localization.velocity, minimum_speed_protection_); // 避免除0
matrix_a_(1, 1) = matrix_a_coeff_(1, 1) / v;
matrix_a_(1, 3) = matrix_a_coeff_(1, 3) / v;
matrix_a_(3, 1) = matrix_a_coeff_(3, 1) / v;
matrix_a_(3, 3) = matrix_a_coeff_(3, 3) / v;
/*
b = [0.0, c_f / m, 0.0, l_f * c_f / i_z]^T
*/
// TODO 02 动力矩阵B
matrix_bd_ = matrix_b_ * ts_;
// cout << "matrix_bd_.row(): " << matrix_bd_.rows() << endl;
// cout << "matrix_bd_.col(): " << matrix_bd_.cols() << endl;
// Update state = [Lateral Error, Lateral Error Rate, Heading Error, Heading Error Rate]
// TODO 03 计算横向误差并且更新状态向量x
UpdateState(localization);
// TODO 04 更新状态矩阵A并将状态矩阵A离散化
UpdateMatrix(localization);
// cout << "matrix_bd_.row(): " << matrix_bd_.rows() << endl;
// cout << "matrix_bd_.col(): " << matrix_bd_.cols() << endl;
// TODO 05 Solve Lqr Problem
SolveLQRProblem(matrix_ad_, matrix_bd_, matrix_q_, matrix_r_, lqr_eps_, lqr_max_iteration_, &matrix_k_);
// TODO 06 计算feedback, 根据反馈对计算状态变量(误差状态)的时候的符号的理解:K里面的值实际运算中全部为正值,steer = -K *
// state,按照代码中采用的横向误差的计算方式,横向误差为正值的时候(state中的第一项为正),参考点位于车辆左侧,车辆应该向左转以减小误差,而根据试验,仿真器中,给正值的时候,车辆向右转,给负值的时候,车辆向左转。
// feedback = - K * state
// Convert vehicle steer angle from rad to degree and then to steer degrees
// then to 100% ratio
std::cout << "matrix_k_: " << matrix_k_ << std::endl;
cout << "matrix_k * matrix_state: " << (matrix_k_ * matrix_state_) << endl;
double steer_angle_feedback = -(matrix_k_ * matrix_state_)(0, 0);
// TODO 07 计算前馈控制,计算横向转角的反馈量
double steer_angle_feedforward = 0.0;
steer_angle_feedforward = ComputeFeedForward(localization, ref_curv_);
double steer_angle = steer_angle_feedback - 0.9 * steer_angle_feedforward;
cout << "steer_angle_feedforward: " << steer_angle_feedforward << endl;
// 限制前轮最大转角,这里定义前轮最大转角位于 [-20度~20度]
if (steer_angle >= atan2_to_PI(20.0)) {
steer_angle = atan2_to_PI(20.0);
} else if (steer_angle <= -atan2_to_PI(20.0)) {
steer_angle = -atan2_to_PI(20.0);
}
// Set the steer commands
cmd.steer_target = steer_angle;
return true;
}
计算横向误差
// TODO 03 计算误差
void LqrController::ComputeLateralErrors(const double x, const double y, const double theta, const double linear_v, const double angular_v, const double linear_a, LateralControlErrorPtr &lat_con_err) {
// 轨迹上最近的点
TrajectoryPoint target_point = QueryNearestPointByPosition(x, y);
double dx = target_point.x - x;
double dy = target_point.y - y;
double theta_p = target_point.heading;
double e_cg = cos(theta_p) * dy - sin(theta_p) * dx;
double e_theta = NormalizeAngle(theta_p - theta);
double diff_e_cg = linear_v * sin(e_theta);
double diff_e_theta = angular_v - target_point.kappa * target_point.v;
lat_con_err->heading_error = e_theta;
lat_con_err->lateral_error = e_cg;
lat_con_err->heading_error_rate = diff_e_theta;
lat_con_err->lateral_error_rate = diff_e_cg;
}
更新状态矩阵并离散化
// TODO 04 更新状态矩阵A并将状态矩阵A离散化
void LqrController::UpdateMatrix(const VehicleState &vehicle_state) {
double v = max(vehicle_state.velocity, minimum_speed_protection_);
// 更新A矩阵
matrix_a_(1, 1) = matrix_a_coeff_(1, 1) / v;
matrix_a_(1, 3) = matrix_a_coeff_(1, 3) / v;
matrix_a_(3, 1) = matrix_a_coeff_(3, 1) / v;
matrix_a_(3, 3) = matrix_a_coeff_(3, 3) / v;
// 更新A_d矩阵,前向欧拉法
// matrix_ad_ = Matrix::Identity(basic_state_size_, basic_state_size_) + matrix_a_ * ts_;
// 更新A_d矩阵,ZOH精确离散化
// 计算指数项
double exp_term = exp(-1.68558 / v);
// 使用ZOH精确离散化
matrix_ad_(0, 0) = 1.0;
matrix_ad_(0, 1) = 0.00593268 * (1.0 - exp_term);
matrix_ad_(0, 2) = 0.010000000000000142 + (-0.00593268 + 0.00593268 * exp_term) / v;
matrix_ad_(0, 3) = 168.55790027100272 * (3.519668608485956e-7 + exp_term * (3.519668608485906e-7 + 4.176213162156246e-7 * v)) - 4.176213162156246e-7 * v;
matrix_ad_(1, 0) = 0.0;
matrix_ad_(1, 1) = 1.0 * exp_term;
matrix_ad_(1, 2) = 168.55790027100272 * (exp_term * (-0.000059326795029614625 - 0.000035196686084859064 * v) + 0.000035196686084859064 * v);
matrix_ad_(1, 3) = 0.0;
matrix_ad_(2, 0) = 0.0;
matrix_ad_(2, 1) = 0.0;
matrix_ad_(2, 2) = 1.0;
matrix_ad_(2, 3) = exp_term * (-0.00593268 * v - 3.3723378595376876e-18) + 0.00593268 * v;
matrix_ad_(3, 0) = 0.0;
matrix_ad_(3, 1) = 0.0;
matrix_ad_(3, 2) = 0.0;
matrix_ad_(3, 3) = exp_term * (1.0 * v + 2.842170943040401e-16) / v;
}
求解LQR问题
// TODO 05:求解LQR方程
void LqrController::SolveLQRProblem(const Matrix &A, const Matrix &B, const Matrix &Q, const Matrix &R, const double tolerance, const uint max_num_iteration, Matrix *ptr_K) {
// 防止矩阵的维数出错导致后续的运算失败
if (A.rows() != A.cols() || B.rows() != A.rows() || Q.rows() != Q.cols() || Q.rows() != A.rows() || R.rows() != R.cols() || R.rows() != B.cols()) {
std::cout << "LQR solver: one or more matrices have incompatible dimensions." << std::endl;
return;
}
Matrix P = Q;
Matrix P_new = Q;
Matrix At = A.transpose();
Matrix Bt = B.transpose();
for (int i = 0; i < max_num_iteration; i++) {
P_new = At * P * A - At * P * B * (R + Bt * P * B).inverse() * Bt * P * A + Q;
double diff = fabs((P_new - P).maxCoeff());
P = P_new;
if (diff < tolerance) {
cout << "diff: " << diff << endl;
*ptr_K = (R + Bt * P * B).inverse() * Bt * P * A;
return;
}
}
cout << "exceed max iteration" << endl;
}
计算前馈控制量
// TODO 07 前馈控制,计算横向转角的反馈量
double LqrController::ComputeFeedForward(const VehicleState &localization, double ref_curvature) {
if (isnan(ref_curvature)) {
ref_curvature = 0;
}
const double K_v = lr_ * mass_ / (2 * cf_ * (lf_ + lr_));
const double v = localization.velocity;
double steer_angle_feedforward = ref_curvature * wheelbase_
+ K_v * v * v * ref_curvature
- matrix_k_(0, 2) * ref_curvature * (lr_ - lf_ * mass_ * v * v / (2 * cr_ * wheelbase_));
return steer_angle_feedforward;
}
四、效果展示
ZOH离散化(无前馈)
ZOH离散化(前馈)
前向欧拉法离散化(前馈)
通过观察可以发现
- 使用ZOH离散化后,转弯过后的横向振荡明显减少。
- 当轨迹的偏航角速度不为0时(转弯),未加入前馈时出现较大的横向误差,加入前馈后横向误差显著减小。
自动驾驶控制与规划——Project 3:LQR车辆横向控制