bullet3/test/InverseDynamics/test_invdyn_kinematics.cpp
erwincoumans e387baf213 add collision and inverse dynamics unit tests to cmake, .travis.yml and appveyor.yml
(more unit tests will follow, once this setup works)
2015-11-24 16:13:21 -08:00

374 lines
14 KiB
C++

// Test of kinematic consistency: check if finite differences of velocities, accelerations
// match positions
#include <cmath>
#include <cstdio>
#include <cstdlib>
#include <iostream>
#include <gtest/gtest.h>
#include "../Extras/InverseDynamics/CoilCreator.hpp"
#include "../Extras/InverseDynamics/DillCreator.hpp"
#include "../Extras/InverseDynamics/SimpleTreeCreator.hpp"
#include "BulletInverseDynamics/MultiBodyTree.hpp"
using namespace btInverseDynamics;
const int kLevel = 5;
const int kNumBodies = BT_ID_POW(2, kLevel);
// template function for calculating the norm
template <typename T>
idScalar calculateNorm(T&);
// only implemented for vec3
template <>
idScalar calculateNorm(vec3& v) {
return std::sqrt(BT_ID_POW(v(0), 2) + BT_ID_POW(v(1), 2) + BT_ID_POW(v(2), 2));
}
// template function to convert a DiffType (finite differences)
// to a ValueType. This is for angular velocity calculations
// via finite differences.
template <typename ValueType, typename DiffType>
DiffType toDiffType(ValueType& fd, ValueType& val);
// vector case: just return finite difference approximation
template <>
vec3 toDiffType(vec3& fd, vec3& val) {
return fd;
}
// orientation case: calculate spin tensor and extract angular velocity
template <>
vec3 toDiffType(mat33& fd, mat33& val) {
// spin tensor
mat33 omega_tilde = fd * val.transpose();
// extract vector from spin tensor
vec3 omega;
omega(0) = 0.5 * (omega_tilde(2, 1) - omega_tilde(1, 2));
omega(1) = 0.5 * (omega_tilde(0, 2) - omega_tilde(2, 0));
omega(2) = 0.5 * (omega_tilde(1, 0) - omega_tilde(0, 1));
return omega;
}
/// Class for calculating finite difference approximation
/// of time derivatives and comparing it to an analytical solution
/// DiffType and ValueType can be different, to allow comparison
/// of angular velocity vectors and orientations given as transform matrices.
template <typename ValueType, typename DiffType>
class DiffFD {
public:
DiffFD() : m_dt(0.0), m_num_updates(0), m_max_error(0.0), m_max_value(0.0), m_valid_fd(false) {}
void init(std::string name, idScalar dt) {
m_name = name;
m_dt = dt;
m_num_updates = 0;
m_max_error = 0.0;
m_max_value = 0.0;
m_valid_fd = false;
}
void update(const ValueType& val, const DiffType& true_diff) {
m_val = val;
if (m_num_updates > 2) {
// 2nd order finite difference approximation for d(value)/dt
ValueType diff_value_fd = (val - m_older_val) / (2.0 * m_dt);
// convert to analytical diff type. This is for angular velocities
m_diff_fd = toDiffType<ValueType, DiffType>(diff_value_fd, m_old_val);
// now, calculate the error
DiffType error_value_type = m_diff_fd - m_old_true_diff;
idScalar error = calculateNorm<DiffType>(error_value_type);
if (error > m_max_error) {
m_max_error = error;
}
idScalar value = calculateNorm<DiffType>(m_old_true_diff);
if (value > m_max_value) {
m_max_value = value;
}
m_valid_fd = true;
}
m_older_val = m_old_val;
m_old_val = m_val;
m_old_true_diff = true_diff;
m_num_updates++;
m_time += m_dt;
}
void printMaxError() {
printf("max_error: %e dt= %e max_value= %e fraction= %e\n", m_max_error, m_dt, m_max_value,
m_max_value > 0.0 ? m_max_error / m_max_value : 0.0);
}
void printCurrent() {
if (m_valid_fd) {
// note: m_old_true_diff already equals m_true_diff here, so values are not aligned.
// (but error calculation takes this into account)
printf("%s time: %e fd: %e %e %e true: %e %e %e\n", m_name.c_str(), m_time,
m_diff_fd(0), m_diff_fd(1), m_diff_fd(2), m_old_true_diff(0), m_old_true_diff(1),
m_old_true_diff(2));
}
}
idScalar getMaxError() const { return m_max_error; }
idScalar getMaxValue() const { return m_max_value; }
private:
idScalar m_dt;
ValueType m_val;
ValueType m_old_val;
ValueType m_older_val;
DiffType m_old_true_diff;
DiffType m_diff_fd;
int m_num_updates;
idScalar m_max_error;
idScalar m_max_value;
idScalar m_time;
std::string m_name;
bool m_valid_fd;
};
template <typename ValueType, typename DiffType>
class VecDiffFD {
public:
VecDiffFD(std::string name, int dim, idScalar dt) : m_name(name), m_fd(dim), m_dt(dt) {
for (int i = 0; i < m_fd.size(); i++) {
char buf[256];
BT_ID_SNPRINTF(buf, 256, "%s-%.2d", name.c_str(), i);
m_fd[i].init(buf, dt);
}
}
void update(int i, ValueType& val, DiffType& true_diff) { m_fd[i].update(val, true_diff); }
idScalar getMaxError() const {
idScalar max_error = 0;
for (int i = 0; i < m_fd.size(); i++) {
const idScalar error = m_fd[i].getMaxError();
if (error > max_error) {
max_error = error;
}
}
return max_error;
}
idScalar getMaxValue() const {
idScalar max_value = 0;
for (int i = 0; i < m_fd.size(); i++) {
const idScalar value = m_fd[i].getMaxValue();
if (value > max_value) {
max_value= value;
}
}
return max_value;
}
void printMaxError() {
printf("%s: total dt= %e max_error= %e\n", m_name.c_str(), m_dt, getMaxError());
}
void printCurrent() {
for (int i = 0; i < m_fd.size(); i++) {
m_fd[i].printCurrent();
}
}
private:
std::string m_name;
std::vector<DiffFD<ValueType, DiffType> > m_fd;
const idScalar m_dt;
idScalar m_max_error;
};
// calculate maximum difference between finite difference and analytical differentiation
int calculateDifferentiationError(const MultiBodyTreeCreator& creator, idScalar deltaT,
idScalar endTime, idScalar* max_linear_velocity_error,
idScalar* max_angular_velocity_error,
idScalar* max_linear_acceleration_error,
idScalar* max_angular_acceleration_error) {
// setup system
MultiBodyTree* tree = CreateMultiBodyTree(creator);
if (0x0 == tree) {
return -1;
}
// set gravity to zero, so nothing is added to accelerations in forward kinematics
vec3 gravity_zero;
gravity_zero(0) = 0;
gravity_zero(1) = 0;
gravity_zero(2) = 0;
tree->setGravityInWorldFrame(gravity_zero);
//
const idScalar kAmplitude = 1.0;
const idScalar kFrequency = 1.0;
vecx q(tree->numDoFs());
vecx dot_q(tree->numDoFs());
vecx ddot_q(tree->numDoFs());
vecx joint_forces(tree->numDoFs());
VecDiffFD<vec3, vec3> fd_vel("linear-velocity", tree->numBodies(), deltaT);
VecDiffFD<vec3, vec3> fd_acc("linear-acceleration", tree->numBodies(), deltaT);
VecDiffFD<mat33, vec3> fd_omg("angular-velocity", tree->numBodies(), deltaT);
VecDiffFD<vec3, vec3> fd_omgd("angular-acceleration", tree->numBodies(), deltaT);
for (idScalar t = 0.0; t < endTime; t += deltaT) {
for (int body = 0; body < tree->numBodies(); body++) {
q(body) = kAmplitude * sin(t * 2.0 * BT_ID_PI * kFrequency);
dot_q(body) = kAmplitude * 2.0 * BT_ID_PI * kFrequency * cos(t * 2.0 * BT_ID_PI * kFrequency);
ddot_q(body) =
-kAmplitude * pow(2.0 * BT_ID_PI * kFrequency, 2) * sin(t * 2.0 * BT_ID_PI * kFrequency);
}
if (-1 == tree->calculateInverseDynamics(q, dot_q, ddot_q, &joint_forces)) {
delete tree;
return -1;
}
// position/velocity
for (int body = 0; body < tree->numBodies(); body++) {
vec3 pos;
vec3 vel;
mat33 world_T_body;
vec3 omega;
vec3 dot_omega;
vec3 acc;
tree->getBodyOrigin(body, &pos);
tree->getBodyTransform(body, &world_T_body);
tree->getBodyLinearVelocity(body, &vel);
tree->getBodyAngularVelocity(body, &omega);
tree->getBodyLinearAcceleration(body, &acc);
tree->getBodyAngularAcceleration(body, &dot_omega);
fd_vel.update(body, pos, vel);
fd_omg.update(body, world_T_body, omega);
fd_acc.update(body, vel, acc);
fd_omgd.update(body, omega, dot_omega);
// fd_vel.printCurrent();
//fd_acc.printCurrent();
//fd_omg.printCurrent();
//fd_omgd.printCurrent();
}
}
*max_linear_velocity_error = fd_vel.getMaxError()/fd_vel.getMaxValue();
*max_angular_velocity_error = fd_omg.getMaxError()/fd_omg.getMaxValue();
*max_linear_acceleration_error = fd_acc.getMaxError()/fd_acc.getMaxValue();
*max_angular_acceleration_error = fd_omgd.getMaxError()/fd_omgd.getMaxValue();
delete tree;
return 0;
}
// first test: absolute difference between numerical and numerial
// differentiation should be small
TEST(InvDynKinematicsDifferentiation, errorAbsolute) {
//CAVEAT:these values are hand-tuned to work for the specific trajectory defined above.
#ifdef BT_ID_USE_DOUBLE_PRECISION
const idScalar kDeltaT = 1e-7;
const idScalar kAcceptableError = 1e-4;
#else
const idScalar kDeltaT = 1e-4;
const idScalar kAcceptableError = 5e-3;
#endif
const idScalar kDuration = 0.01;
CoilCreator coil_creator(kNumBodies);
DillCreator dill_creator(kLevel);
SimpleTreeCreator simple_creator(kNumBodies);
idScalar max_linear_velocity_error;
idScalar max_angular_velocity_error;
idScalar max_linear_acceleration_error;
idScalar max_angular_acceleration_error;
// test serial chain
calculateDifferentiationError(coil_creator, kDeltaT, kDuration, &max_linear_velocity_error,
&max_angular_velocity_error, &max_linear_acceleration_error,
&max_angular_acceleration_error);
EXPECT_LT(max_linear_velocity_error, kAcceptableError);
EXPECT_LT(max_angular_velocity_error, kAcceptableError);
EXPECT_LT(max_linear_acceleration_error, kAcceptableError);
EXPECT_LT(max_angular_acceleration_error, kAcceptableError);
// test branched tree
calculateDifferentiationError(dill_creator, kDeltaT, kDuration, &max_linear_velocity_error,
&max_angular_velocity_error, &max_linear_acceleration_error,
&max_angular_acceleration_error);
EXPECT_LT(max_linear_velocity_error, kAcceptableError);
EXPECT_LT(max_angular_velocity_error, kAcceptableError);
EXPECT_LT(max_linear_acceleration_error, kAcceptableError);
EXPECT_LT(max_angular_acceleration_error, kAcceptableError);
// test system with different joint types
calculateDifferentiationError(simple_creator, kDeltaT, kDuration, &max_linear_velocity_error,
&max_angular_velocity_error, &max_linear_acceleration_error,
&max_angular_acceleration_error);
EXPECT_LT(max_linear_velocity_error, kAcceptableError);
EXPECT_LT(max_angular_velocity_error, kAcceptableError);
EXPECT_LT(max_linear_acceleration_error, kAcceptableError);
EXPECT_LT(max_angular_acceleration_error, kAcceptableError);
}
// second test: check if the change in the differentiation error
// is consitent with the second order approximation, ie, error ~ O(dt^2)
TEST(InvDynKinematicsDifferentiation, errorOrder) {
const idScalar kDeltaTs[2] = {1e-4, 1e-5};
const idScalar kDuration = 1e-2;
CoilCreator coil_creator(kNumBodies);
// DillCreator dill_creator(kLevel);
// SimpleTreeCreator simple_creator(kNumBodies);
idScalar max_linear_velocity_error[2];
idScalar max_angular_velocity_error[2];
idScalar max_linear_acceleration_error[2];
idScalar max_angular_acceleration_error[2];
// test serial chain
calculateDifferentiationError(coil_creator, kDeltaTs[0], kDuration,
&max_linear_velocity_error[0], &max_angular_velocity_error[0],
&max_linear_acceleration_error[0],
&max_angular_acceleration_error[0]);
calculateDifferentiationError(coil_creator, kDeltaTs[1], kDuration,
&max_linear_velocity_error[1], &max_angular_velocity_error[1],
&max_linear_acceleration_error[1],
&max_angular_acceleration_error[1]);
const idScalar expected_linear_velocity_error_1 =
max_linear_velocity_error[0] * pow(kDeltaTs[1] / kDeltaTs[0], 2);
const idScalar expected_angular_velocity_error_1 =
max_angular_velocity_error[0] * pow(kDeltaTs[1] / kDeltaTs[0], 2);
const idScalar expected_linear_acceleration_error_1 =
max_linear_acceleration_error[0] * pow(kDeltaTs[1] / kDeltaTs[0], 2);
const idScalar expected_angular_acceleration_error_1 =
max_angular_acceleration_error[0] * pow(kDeltaTs[1] / kDeltaTs[0], 2);
/* printf("linear vel error: %e %e %e\n", max_linear_velocity_error[1],
expected_linear_velocity_error_1,
max_linear_velocity_error[1] - expected_linear_velocity_error_1);
printf("angular vel error: %e %e %e\n", max_angular_velocity_error[1],
expected_angular_velocity_error_1,
max_angular_velocity_error[1] - expected_angular_velocity_error_1);
printf("linear acc error: %e %e %e\n", max_linear_acceleration_error[1],
expected_linear_acceleration_error_1,
max_linear_acceleration_error[1] - expected_linear_acceleration_error_1);
printf("angular acc error: %e %e %e\n", max_angular_acceleration_error[1],
expected_angular_acceleration_error_1,
max_angular_acceleration_error[1] - expected_angular_acceleration_error_1);
*/
}
int main(int argc, char** argv) {
::testing::InitGoogleTest(&argc, argv);
return RUN_ALL_TESTS();
return EXIT_SUCCESS;
}