I definitely believe it’s vital to bring up the concept of tensor integration with all of the ML stuff becoming more prominent in the industry. But I do have a lot of doubts about the need to integrate rcl::tensor into ROS’ core libraries. As mentioned:
That seems to be the most justifiable reason as to not integrate directly, just like how there are no data containers: rcl::vector or rcl::span. Why increase bloat, especially for those who will not need to use it?
That being said, it would be very beneficial to the entire community to have a separate library with things like tensors in mind, allowing easier integration of neural networks. This is amplified by the fact we could use (for the C++ side) TypeAdaptations, which have a big boost in performance. This would work well for external interfaces which use PyTorch, LibTorch or another already very optimized library.
Certainly seems necessary to give users that option as it’s been the standard for many other extended types and could be specific to hardware too.
Yes here’s the Original PR & corresponding REP 2007