Only Truth Matters

QT

Posted on 2023-03-09 Edited on 2023-09-23 In MarkdownNotes , Engineering , Computer Science , Miscellaneous

1. QT

signal

when signal emit, it would run instantly.

Current Problem

Line mode could be activated multiple times
Point mode only 1 times
Rect mode can’t be activated after Line mode

2. How it Works?

Make the Line work as Rect
add custom text for label

Pybind11

Posted on 2023-03-09 Edited on 2023-09-23 In MarkdownNotes , Engineering , Computer Science , Miscellaneous

1. What

Header only library
Expose C++ function, classes and objects as Python module

2. How to Use

Include the pybind11
define python modules using macros
export to python module

3. Example

my_class.h

#pragma once

#include <string>

class MyClass {
public:
MyClass(const std::string& name);
void setName(const std::string& name);
std::string getName() const;

private:
std::string name_;
};

my_class.cpp

#include "my_class.h"

MyClass::MyClass(const std::string& name) : name_(name) {}

void MyClass::setName(const std::string& name) {
name_ = name;
}

std::string MyClass::getName() const {
return name_;
}

My_module.cpp

Dynamic? Static?

Posted on 2023-03-08 Edited on 2023-09-20 In MarkdownNotes , Engineering , Computer Science , Miscellaneous

1. Goal

Understand the difference between dynamic and static libraries in software

Static: link at compile time, make the final program larger
Dynamic: link at runtime, require the system to have the lib

When to use?

Static: lib is small, not change, need to be optimized
Dynamic: lib is large, frequently update

2. Compile Process

Preprocessing, remove comments, expand macro
Compilation, translate into assembly
Assembly, translatie into machine code
Linking, combine it with other libraries and modules

3. Linking?

Static linking is usually used for local
Dynamic linking is usually used for external libraries

How it works?

ALOAM Learning

Posted on 2023-03-07 Edited on 2023-09-20 In MarkdownNotes , Engineering , Robotics , Perception , Lidar

1. Goal

Understanding the overall structure of the LOAM project and try to compile it without ROS

2. Theory

Paper: Lidar Odometry and Mapping in Real-time

Explanation

Problem: error in motion estimation cause mis-registration

Solution: propose 2 odometry

High frequency, low accuracy
Low frequency, high accuracy

Some key point

人类毛发疑惑

Posted on 2023-03-07 Edited on 2023-09-20 In MarkdownNotes , Life Skill , Interests

1. 问题

为什么头发没有在进化过程中减少
为什么头发会一直生长而其他毛发不会

2. 头发？

头发比较重要

保护，吸收紫外线
维持体温，大脑比较重要，且散热较多
社交信号

3. 生长？

生物结构不一样

头皮的毛囊更大，生长周期较长
其他部位毛发生长数月，就会停止生长

Lidar SLAM System Review

Posted on 2023-03-06 Edited on 2023-09-20 In MarkdownNotes , Engineering , Robotics , Perception , Lidar

1. Intro

Review Lidar based SLAM algorithm and libraries

Some of the popular lib:

LOAM
LEGO LOAM
LIO SAM

2. LOAM

Year: 2014

Lidar Odometry and Mapping

Scan Matching
Odometry
Fast and efficient

Con:

Require careful parameter tuning

Problem Calibrating Lidar and IMU

Posted on 2023-03-06 Edited on 2023-09-20 In MarkdownNotes , Engineering , Robotics , Perception , Lidar

1. 背景

硬件：

雷神雷达C16
16线
每线2000点
北斗星通MS-6111 六轴IMU
xyz加速度
角速度

软件

校准：lidar_imu_calib
开启雷达和IMU的ros节点
rosbag记录，同时移动雷达和IMU
计算IMU和雷达之间的位姿转换矩阵（只计算旋转）
SLAM：liorf
LIO-SAM的变种，可以接收六轴传感器

2. 问题

SLAM能够正常启动和运行，但是里程计会漂移，有很大的误差

观察到的现象

在平地时，IMU数据的x,y均有值。0.1m/s^2左右
里程计显示速度越来越快

BFGS in Optimization

Posted on 2023-03-04 Edited on 2023-09-20 In MarkdownNotes , Nature Science , Mathematics , Optimization

1. What

Broyden-Fletcher-Goldfarb-Shanno

Quasi-Newton optimization algorithm
Gradient based: use first derivative of th objective function to improve the solution
differnce of different algorithm: how to update the parameters
first derivative
hessian matrix – BFGS
Maintain an approximation of the Hessian

2. How

Find the minimum, the gradient is known or calculated
$$
B_{k+1} = B_k + \frac{(\Delta \theta_k)(\Delta \theta_k)^T}{(\Delta \theta_k)^T \Delta g_k} - \frac{B_k \Delta g_k (\Delta g_k)^T B_k}{(\Delta g_k)^T B_k \Delta g_k}
$$

import numpy as np

def bfgs(f, grad_f, x0, epsilon=1e-6, max_iter=1000):
n = len(x0)
Bk = np.eye(n) # initialize Hessian approximation
xk = x0
gk = grad_f(xk)
for k in range(max_iter):
if np.linalg.norm(gk) < epsilon:
break
pk = -np.dot(Bk, gk)
alpha = backtracking_line_search(f, grad_f, xk, pk)
xk_next = xk + alpha*pk
sk = xk_next - xk
yk = grad_f(xk_next) - gk # update the second order derivative
Bk_next = Bk + np.outer(yk, yk)/np.dot(yk, sk) - np.outer(np.dot(Bk, sk), np.dot(Bk, sk))/np.dot(sk, np.dot(Bk, sk))
xk, gk, Bk = xk_next, grad_f(xk_next), Bk_next
return xk

3. Ref

Wiki

ChatGPT

Trac-IK Paper Review

Posted on 2023-03-04 Edited on 2023-09-20 In MarkdownNotes , Engineering , Robotics , Action , Kinematics

1. Goal

In the paper, author compared some common open-source project address the kinematics problem. I want to know

Which optimization method is used behind these popular project
How these project compared to each other

2. Introduciton

The main purpose if trac_ik is to improve the popular KDL

Issue with KDL

convergence failure
stuck in local minima
loop for maximum time rather than iteration
Detect by $q_{next}-q_{prev}\approx 0$
Change to a random seed
Failure with joint limit
solved with SQP sequential quadratic programming
no utilization of tolerance

3. Candidate IK Algorithm

KDL-RR
Random Restart
maximum time
SQP: SLSQP with BFGS
$$
\arg \min {q \in \mathbf{R}^n}\left(q{\text {seed }}-q\right)^T\left(q_{\text {seed }}-q\right),
$$
SQP-DQ: dual quaternion
combine the translation and orientation
$$
\phi_{D Q}=4\left((\log \hat{e}) \cdot(\log \hat{e})^T\right)
$$
SQP-SS and SQP-L2, $p$ as simple 6-element vector
SS: $\phi_{S S}=p_{e r r} \cdot p_{e r r}^T$
L2: $\phi_{L 2}=\sqrt{p_{e r r} \cdot p_{e r r}^T} .$
Trac-ik
Run two algorithm at once
SQP-SS
KDL-RR
Use the fastest one