[wpimath] Add ImplicitModelFollower (#4056)

This commit is contained in:
Tyler Veness
2022-03-20 00:36:12 -07:00
committed by GitHub
parent 78108c2aba
commit 8d79dc8738
4 changed files with 493 additions and 0 deletions

View File

@@ -0,0 +1,138 @@
// Copyright (c) FIRST and other WPILib contributors.
// Open Source Software; you can modify and/or share it under the terms of
// the WPILib BSD license file in the root directory of this project.
package edu.wpi.first.math.controller;
import edu.wpi.first.math.Matrix;
import edu.wpi.first.math.Num;
import edu.wpi.first.math.numbers.N1;
import edu.wpi.first.math.system.Discretization;
import edu.wpi.first.math.system.LinearSystem;
import org.ejml.simple.SimpleMatrix;
/**
* Contains the controller coefficients and logic for an implicit model follower.
*
* <p>Implicit model following lets us design a feedback controller that erases the dynamics of our
* system and makes it behave like some other system. This can be used to make a drivetrain more
* controllable during teleop driving by making it behave like a slower or more benign drivetrain.
*
* <p>For more on the underlying math, read appendix B.3 in
* https://file.tavsys.net/control/controls-engineering-in-frc.pdf.
*/
@SuppressWarnings("ClassTypeParameterName")
public class ImplicitModelFollower<States extends Num, Inputs extends Num, Outputs extends Num> {
// Computed controller output
@SuppressWarnings("MemberName")
private Matrix<Inputs, N1> m_u;
// State space conversion gain
@SuppressWarnings("MemberName")
private Matrix<Inputs, States> m_A;
// Input space conversion gain
@SuppressWarnings("MemberName")
private Matrix<Inputs, Inputs> m_B;
/**
* Constructs a controller with the given coefficients and plant.
*
* @param plant The plant being controlled.
* @param plantRef The plant whose dynamics should be followed.
* @param dtSeconds Discretization timestep.
*/
public ImplicitModelFollower(
LinearSystem<States, Inputs, Outputs> plant,
LinearSystem<States, Inputs, Outputs> plantRef,
double dtSeconds) {
this(plant.getA(), plant.getB(), plantRef.getA(), plantRef.getB(), dtSeconds);
}
/**
* Constructs a controller with the given coefficients and plant.
*
* @param A Continuous system matrix of the plant being controlled.
* @param B Continuous input matrix of the plant being controlled.
* @param Aref Continuous system matrix whose dynamics should be followed.
* @param Bref Continuous input matrix whose dynamics should be followed.
* @param dtSeconds Discretization timestep.
*/
@SuppressWarnings("ParameterName")
public ImplicitModelFollower(
Matrix<States, States> A,
Matrix<States, Inputs> B,
Matrix<States, States> Aref,
Matrix<States, Inputs> Bref,
double dtSeconds) {
m_u = new Matrix<>(new SimpleMatrix(B.getNumCols(), 1));
// Discretize real dynamics
var discABPair = Discretization.discretizeAB(A, B, dtSeconds);
var discA = discABPair.getFirst();
var discB = discABPair.getSecond();
// Discretize desired dynamics
var discABrefPair = Discretization.discretizeAB(Aref, Bref, dtSeconds);
var discAref = discABrefPair.getFirst();
var discBref = discABrefPair.getSecond();
// Find u_imf that makes real model match reference model.
//
// x_k+1 = Ax_k + Bu_imf
// z_k+1 = Aref z_k + Bref u_k
//
// Let x_k = z_k.
//
// x_k+1 = z_k+1
// Ax_k + Bu_imf = Aref x_k + Bref u_k
// Bu_imf = Aref x_k - Ax_k + Bref u_k
// Bu_imf = (Aref - A)x_k + Bref u_k
// u_imf = B^+ ((Aref - A)x_k + Bref u_k)
// u_imf = -B^+ (A - Aref)x_k + B^+ Bref u_k
// The first term makes the open-loop poles that of the reference
// system, and the second term makes the input behave like that of the
// reference system.
m_A = discB.solve(discA.minus(discAref)).times(-1.0);
m_B = discB.solve(discBref);
reset();
}
/**
* Returns the control input vector u.
*
* @return The control input.
*/
public Matrix<Inputs, N1> getU() {
return m_u;
}
/**
* Returns an element of the control input vector u.
*
* @param i Row of u.
* @return The row of the control input vector.
*/
public double getU(int i) {
return m_u.get(i, 0);
}
/** Resets the controller. */
public void reset() {
m_u.fill(0.0);
}
/**
* Returns the next output of the controller.
*
* @param x The current state x.
* @param u The current input for the original model.
* @return The next controller output.
*/
public Matrix<Inputs, N1> calculate(Matrix<States, N1> x, Matrix<Inputs, N1> u) {
m_u = m_A.times(x).plus(m_B.times(u));
return m_u;
}
}

View File

@@ -0,0 +1,137 @@
// Copyright (c) FIRST and other WPILib contributors.
// Open Source Software; you can modify and/or share it under the terms of
// the WPILib BSD license file in the root directory of this project.
#pragma once
#include <frc/system/Discretization.h>
#include <frc/system/LinearSystem.h>
#include "Eigen/Core"
#include "Eigen/QR"
#include "units/time.h"
namespace frc {
/**
* Contains the controller coefficients and logic for an implicit model
* follower.
*
* Implicit model following lets us design a feedback controller that erases the
* dynamics of our system and makes it behave like some other system. This can
* be used to make a drivetrain more controllable during teleop driving by
* making it behave like a slower or more benign drivetrain.
*
* For more on the underlying math, read appendix B.3 in
* https://file.tavsys.net/control/controls-engineering-in-frc.pdf.
*/
template <int States, int Inputs>
class ImplicitModelFollower {
public:
/**
* Constructs a controller with the given coefficients and plant.
*
* @param plant The plant being controlled.
* @param plantRef The plant whose dynamics should be followed.
* @param dt Discretization timestep.
*/
template <int Outputs>
ImplicitModelFollower(const LinearSystem<States, Inputs, Outputs>& plant,
const LinearSystem<States, Inputs, Outputs>& plantRef,
units::second_t dt)
: ImplicitModelFollower<States, Inputs>(plant.A(), plant.B(),
plantRef.A(), plantRef.B(), dt) {}
/**
* Constructs a controller with the given coefficients and plant.
*
* @param A Continuous system matrix of the plant being controlled.
* @param B Continuous input matrix of the plant being controlled.
* @param Aref Continuous system matrix whose dynamics should be followed.
* @param Bref Continuous input matrix whose dynamics should be followed.
* @param dt Discretization timestep.
*/
ImplicitModelFollower(const Eigen::Matrix<double, States, States>& A,
const Eigen::Matrix<double, States, Inputs>& B,
const Eigen::Matrix<double, States, States>& Aref,
const Eigen::Matrix<double, States, Inputs>& Bref,
units::second_t dt) {
// Discretize real dynamics
Eigen::Matrix<double, States, States> discA;
Eigen::Matrix<double, States, Inputs> discB;
frc::DiscretizeAB<States, Inputs>(A, B, dt, &discA, &discB);
// Discretize desired dynamics
Eigen::Matrix<double, States, States> discAref;
Eigen::Matrix<double, States, Inputs> discBref;
frc::DiscretizeAB<States, Inputs>(Aref, Bref, dt, &discAref, &discBref);
// Find u_imf that makes real model match reference model.
//
// x_k+1 = Ax_k + Bu_imf
// z_k+1 = Aref z_k + Bref u_k
//
// Let x_k = z_k.
//
// x_k+1 = z_k+1
// Ax_k + Bu_imf = Aref x_k + Bref u_k
// Bu_imf = Aref x_k - Ax_k + Bref u_k
// Bu_imf = (Aref - A)x_k + Bref u_k
// u_imf = B^+ ((Aref - A)x_k + Bref u_k)
// u_imf = -B^+ (A - Aref)x_k + B^+ Bref u_k
// The first term makes the open-loop poles that of the reference
// system, and the second term makes the input behave like that of the
// reference system.
m_A = -discB.householderQr().solve(discA - discAref);
m_B = discB.householderQr().solve(discBref);
Reset();
}
/**
* Returns the control input vector u.
*
* @return The control input.
*/
const Eigen::Vector<double, Inputs>& U() const { return m_u; }
/**
* Returns an element of the control input vector u.
*
* @param i Row of u.
*
* @return The row of the control input vector.
*/
double U(int i) const { return m_u(i); }
/**
* Resets the controller.
*/
void Reset() { m_u.setZero(); }
/**
* Returns the next output of the controller.
*
* @param x The current state x.
* @param u The current input for the original model.
*/
Eigen::Vector<double, Inputs> Calculate(
const Eigen::Vector<double, States>& x,
const Eigen::Vector<double, Inputs>& u) {
m_u = m_A * x + m_B * u;
return m_u;
}
private:
// Computed controller output
Eigen::Vector<double, Inputs> m_u;
// State space conversion gain
Eigen::Matrix<double, Inputs, States> m_A;
// Input space conversion gain
Eigen::Matrix<double, Inputs, Inputs> m_B;
};
} // namespace frc

View File

@@ -0,0 +1,109 @@
// Copyright (c) FIRST and other WPILib contributors.
// Open Source Software; you can modify and/or share it under the terms of
// the WPILib BSD license file in the root directory of this project.
package edu.wpi.first.math.controller;
import static org.junit.jupiter.api.Assertions.assertEquals;
import static org.junit.jupiter.api.Assertions.assertTrue;
import edu.wpi.first.math.MatBuilder;
import edu.wpi.first.math.Nat;
import edu.wpi.first.math.numbers.N2;
import edu.wpi.first.math.system.plant.LinearSystemId;
import org.junit.jupiter.api.Test;
class ImplicitModelFollowerTest {
@Test
@SuppressWarnings("LocalVariableName")
void testSameModel() {
final double dt = 0.005;
var plant = LinearSystemId.identifyDrivetrainSystem(1.0, 1.0, 1.0, 1.0);
var imf = new ImplicitModelFollower<N2, N2, N2>(plant, plant, dt);
var x = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(0.0, 0.0);
var xImf = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(0.0, 0.0);
// Forward
var u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(12.0, 12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertEquals(x.get(0, 0), xImf.get(0, 0));
assertEquals(x.get(1, 0), xImf.get(1, 0));
}
// Backward
u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(-12.0, -12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertEquals(x.get(0, 0), xImf.get(0, 0));
assertEquals(x.get(1, 0), xImf.get(1, 0));
}
// Rotate CCW
u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(-12.0, 12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertEquals(x.get(0, 0), xImf.get(0, 0));
assertEquals(x.get(1, 0), xImf.get(1, 0));
}
}
@Test
@SuppressWarnings("LocalVariableName")
void testSlowerRefModel() {
final double dt = 0.005;
var plant = LinearSystemId.identifyDrivetrainSystem(1.0, 1.0, 1.0, 1.0);
// Linear acceleration is slower, but angular acceleration is the same
var plantRef = LinearSystemId.identifyDrivetrainSystem(1.0, 2.0, 1.0, 1.0);
var imf = new ImplicitModelFollower<N2, N2, N2>(plant, plantRef, dt);
var x = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(0.0, 0.0);
var xImf = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(0.0, 0.0);
// Forward
var u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(12.0, 12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertTrue(x.get(0, 0) >= xImf.get(0, 0));
assertTrue(x.get(1, 0) >= xImf.get(1, 0));
}
// Backward
x.fill(0.0);
xImf.fill(0.0);
u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(-12.0, -12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertTrue(x.get(0, 0) <= xImf.get(0, 0));
assertTrue(x.get(1, 0) <= xImf.get(1, 0));
}
// Rotate CCW
x.fill(0.0);
xImf.fill(0.0);
u = new MatBuilder<>(Nat.N2(), Nat.N1()).fill(-12.0, 12.0);
for (double t = 0.0; t < 3.0; t += dt) {
x = plant.calculateX(x, u, dt);
xImf = plant.calculateX(xImf, imf.calculate(xImf, u), dt);
assertEquals(x.get(0, 0), xImf.get(0, 0), 1e-5);
assertEquals(x.get(1, 0), xImf.get(1, 0), 1e-5);
}
}
}

View File

@@ -0,0 +1,109 @@
// Copyright (c) FIRST and other WPILib contributors.
// Open Source Software; you can modify and/or share it under the terms of
// the WPILib BSD license file in the root directory of this project.
#include <gtest/gtest.h>
#include "frc/controller/ImplicitModelFollower.h"
#include "frc/system/plant/LinearSystemId.h"
namespace frc {
TEST(ImplicitModelFollowerTest, SameModel) {
constexpr auto dt = 5_ms;
using Kv_t = decltype(1_V / 1_mps);
using Ka_t = decltype(1_V / 1_mps_sq);
auto plant = LinearSystemId::IdentifyDrivetrainSystem(Kv_t{1.0}, Ka_t{1.0},
Kv_t{1.0}, Ka_t{1.0});
ImplicitModelFollower<2, 2> imf{plant, plant, dt};
Eigen::Vector<double, 2> x{0.0, 0.0};
Eigen::Vector<double, 2> xImf{0.0, 0.0};
// Forward
Eigen::Vector<double, 2> u{12.0, 12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_DOUBLE_EQ(x(0), xImf(0));
EXPECT_DOUBLE_EQ(x(1), xImf(1));
}
// Backward
u = Eigen::Vector<double, 2>{-12.0, -12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_DOUBLE_EQ(x(0), xImf(0));
EXPECT_DOUBLE_EQ(x(1), xImf(1));
}
// Rotate CCW
u = Eigen::Vector<double, 2>{-12.0, 12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_DOUBLE_EQ(x(0), xImf(0));
EXPECT_DOUBLE_EQ(x(1), xImf(1));
}
}
TEST(ImplicitModelFollowerTest, SlowerRefModel) {
constexpr auto dt = 5_ms;
using Kv_t = decltype(1_V / 1_mps);
using Ka_t = decltype(1_V / 1_mps_sq);
auto plant = LinearSystemId::IdentifyDrivetrainSystem(Kv_t{1.0}, Ka_t{1.0},
Kv_t{1.0}, Ka_t{1.0});
// Linear acceleration is slower, but angular acceleration is the same
auto plantRef = LinearSystemId::IdentifyDrivetrainSystem(
Kv_t{1.0}, Ka_t{2.0}, Kv_t{1.0}, Ka_t{1.0});
ImplicitModelFollower<2, 2> imf{plant, plantRef, dt};
Eigen::Vector<double, 2> x{0.0, 0.0};
Eigen::Vector<double, 2> xImf{0.0, 0.0};
// Forward
Eigen::Vector<double, 2> u{12.0, 12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_GE(x(0), xImf(0));
EXPECT_GE(x(1), xImf(1));
}
// Backward
x.setZero();
xImf.setZero();
u = Eigen::Vector<double, 2>{-12.0, -12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_LE(x(0), xImf(0));
EXPECT_LE(x(1), xImf(1));
}
// Rotate CCW
x.setZero();
xImf.setZero();
u = Eigen::Vector<double, 2>{-12.0, 12.0};
for (auto t = 0_s; t < 3_s; t += dt) {
x = plant.CalculateX(x, u, dt);
xImf = plant.CalculateX(xImf, imf.Calculate(xImf, u), dt);
EXPECT_NEAR(x(0), xImf(0), 1e-5);
EXPECT_NEAR(x(1), xImf(1), 1e-5);
}
}
} // namespace frc