As it turns out, the Zed camera ecosystem is painfully, unequivocally unportable.

Their [https://github.com/stereolabs/zed-ros2-wrapper package] is particularly bad. You can't "just" add it to your package.xml and call it a day - nope! You've got to do a boatload of setup on each computer you install it on. That's not workable for Autonomous.

A generic solution

I'm thinking of a more generic solution instead. Unfortunately, it only runs on the CPU unless you perform some annoying NVIDIA setup, but we can look into those optimizations later on...

image_publisher

A few instances of these nodes will publish the required data (like in this file):

• /zed_2i_left_image_publisher: publishes /sensors/depth_image/left

• /zed_2i_right_image_publisher: publishes `/sensors/depth_image/right

stereo_image_proc

This package processes camera output into a

Optimizations (hopefully for later)

• We could use isaac_ros_stereo_image_proc for CUDA acceleration
  • Still not very portable :(
  • Would need some very hacky colcon shenanigans to replace it with the non-NVIDIA version if hardware acceleration isn't detected
    • Could be possible to do this at runtime? Sounds hacky, impossible to test, and still incompatible with almost all hardware that Autonomous members own lol
• Consider a Rust node making use of wgpu + compute shaders?
  • Would be 9000x more portable; probably eligible to be distributed through rosdep
  • Hilariously, about the same amount of documentation (very little on Rust compute shaders past this intro guide... by Google Fonts? 🤨)
  • This'd probably be one of the first (fast) implementations of a hardware-accelerated camera node that can run on anything!
  • Would need to take some time to write this