Lidar point clouds, high resolution imagery and driver inputs mostly. The high level idea being you want to train the neural network so that when it sees x on the sensors it should do y thing the human did.

As for why other self driving cars don't have lidar that's... A contentious topic. The argument for lidar is that it captures the environment in 3d unlike cameras that only capture in 2d and thus you can save on computation trying to convert 2d images into a 3d scene. On the other hand lidar is very expensive, has limitations like not being able to see through rain and maybe most importantly is believed to be unnecessary since "humans drive with just their eyes".

My personal belief is its a matter of cost vs time to market - lidar makes locatization and object recognition a little bit easier and if for example it saves 1 year of development time at a cost of 100k per car produced it's worth it if you're only making a small fleet of cars anyway.