mapping drift when using the dataset provided by lio-sam

[xiaoxueshengyao]

Thans for your great work. While the robot will drift immediately when using the dataset "Campus dataset (small)" like this. What should I modify the project.

[qdLMF]

Try use config/params-campus.yaml

[xiaoxueshengyao]

Try use config/params-campus.yaml

Thanks for your reply. I had tried to use the yaml, but the drift is still occured. I tried to use more iteration, while the optimization can not converge even if the iteration is setted to 300.

[qdLMF]

Try use config/params-campus.yaml

Thanks for your reply. I had tried to use the yaml, but the drift is still occured. I tried to use more iteration, while the optimization can not converge even if the iteration is setted to 300.

Does it drift when using other bag files?

[xiaoxueshengyao]

Yes, I had tried several bags. All of them drifted.

[xiaoxueshengyao]

方便交流我用中文了，在我测试发现calc_surf_coeff.cu中计算面方程式使用qr分解去计算系数，但是所有得到的系数都是nan，两种方法都是同样的结果
代码如下
`for(int i = 0; i<5; i++)
{
printf("matA0: %f %f %f \n",matA0(i,0),matA0(i,1),matA0(i,2));
}

// ColPivHouseholderQR5x3 qr_decomp(matA0);
ColPivHouseholderQRNx3<5> qr_decomp(matA0);
matX0 = qr_decomp.Solve(matB0);
printf("matX0 %f %f %f\n",matX0(0,0),matX0(1,0),matX0(2,0));`

对应打印如下

[qdLMF]

尝试将cuda_plane_line_odometry/CMakeLists.txt中的set(CMAKE_CUDA_ARCHITECTURES 52)中的52改成你的GPU的compute capability
或者，使用注释掉的// ColPivHouseholderQR5x3 qr_decomp(matA0);
这里的qr分解，我原先想直接用Eigen的，但是报了超多模板错误，实在调不通无奈之下才写了这么个5x3的版本
我在其它项目中也碰上过关于Eigen的诡异错误
看来在kernel内盲目使用Eigen也许是个不成熟的决定

[tangqitao]

我也遇到了同样的问题，我的显卡是RTXA1000，查询到我是8.6,我尝试把 set(CMAKE_CUDA_ARCHITECTURES 86)和使用例外一个ColPivHouseholderQR5x3 ，这个类，也还是不行，跑compus这个数据集
for(int i = 0; i<5; i++)
{
printf("matA0: %f %f %f \n",matA0(i,0),matA0(i,1),matA0(i,2));
}
ColPivHouseholderQR5x3 qr_decomp(matA0);
// 使用列主元Householder QR分解求解线性方程组 Ax = B
//ColPivHouseholderQRNx3<5> qr_decomp(matA0);
matX0 = qr_decomp.Solve(matB0);
printf("matX0 %f %f %f\n",matX0(0,0),matX0(1,0),matX0(2,0));
// if(i == 0) {
// Eigen::Matrix<float, 5, 1> matB1 = matA0 * matX0;
// printf(
// "matB1 : %.10f , %.10f , %.10f , %.10f , %.10f \n",
// matB1(0), matB1(1), matB1(2), matB1(3), matB1(4)
// );
// }
// 验证求解结果（可选的调试代码）
打印结果和上面一致
matA0: 31.037735 57.008869 14.663079
matA0: 32.185959 56.992977 14.846636
matX0 nan nan nan
matX0 nan nan nan
matX0 nan nan nan

[tangqitao]

后面我定位到的原因是使用了EIGEN 的乘法，而在gpu上是不被实现的，我想问问是否是因为我们使用的eigen版本有差异

[qdLMF]

The old branch's 5x3 QR decomposition uses rather complicated Eigen expressions, which works fine on my desktop, but failed on my ORIN-NX-8GB. I can only guess ORIN's enviroment can't handle these Eigen expressions. So I wrote a new version of 5x3 QR decomposion using only simple array-based operations. The new version works fine on my ORIN-NX-8GB and is now uploaded to this repo.