I've recently changed my setup so that all the fixed tfs are derived from a carefully constructed URDF. Before this I had launch files with static_transfor_publisherm nodes and this led to a conflict.

My initial approach is to process the urdf from the "bringup" on the robot. I do this so that I can just run rviz or other nodes on the remote computer without having to separately load the urdf. But I am not sure this is the right approach.

Processing the urdf on the robot itself (using robot_state_publisher and joint_state_publisher) means that the transforms are being published on the Raspberry Pi. It doesn't seem overloaded (based on idle time). But still I wonder what the best practice is.

What do you do?

Run the urdf robot state publisher onboard the robot or off on a remote computer? Is there a best practice? Can you think why it may matter either way?