Camera vs. LiDAR

The age old debate, which one is better?

Most of my opinions comes from listening to Andrej Karpathy podcast on Lex Fridman.

Tesla is going for a camera-only approach.

  • I firmly believed in this. Argument (in my own words, opinion probably obtained from listening to other people): There is additional overhead from adding other sensors that you need to introduce into your pipeline, from data collection, training, and deployment.
    • The layer deeper than this that Gordon made to me today (Oct 24th 2023) is that all of this adds additional cost. The biggest cost being that you now need to add a LiDAR to every single robot

The way NVIDIA Robotics is going to be different is that … see the private page

I was making the argument that you want to have multiple modalities to create redundancy into your system.

  • That is the whole safety idea

Gordon was telling me about how a safety rated LiDAR actually has 2 different algorithms running, and they make physics argument for why it works.