[aktenlage]

Depends on what you want to learn. The SLAM Frontend (computing motion information from sensors) offers a lot of variety through the chosen combination of sensors (wheel odometer, IMU, mono/stereo/multi camera, lidar, radar, sonar, to name a few).

At least for vision, deep learning should be very useful. For the others I have no experience how much machine learning is relevant. Geometry and physics based methods should work well here, but there is probably much room to tack on some ML.

The SLAM backend (optimization) is mostly old-school methods like nonlinear least squares optimization or particle filters. Not sure if that counts as ML today.

I'd go for g2o or Ceres for mapping (unless you expect to have no loop closures), as there's really no need to reinvent that. It's definitely useful to learn about the backend, but usually the combination of sensors and their properties will demand more algorithmic tailoring than the backend, which gets more abstract input (i e. motions & uncertainties) and can be used more black-boxy.

[yalogin]

Thanks for the input, would you recommend developing slam from scratch or start with some libraries and tailor it to my sensors?