[joshvm]

Robotics is a big field and encompasses a huge range of things. Similarly a lot of computer vision is pretty technical, particularly for robotics applications. Have a look at ROS and play around, you can simulate a lot of stuff. I'd also suggest Nvidia's Jetbot but the chip shortage has killed it.

AI-wise, nowadays model training stacks are practically codeless. Training a state of the art model on custom data usually boils down to changing some parameters in a config file, plus some boilerplate code. Not always that easy of course, but training an image classifier today can be a single line of Python.

It's not so much piping data around but dealing with datasets in any kind of non-trivial problem is where the hard work is. Neural nets are just another tool for most data scientists. For example exploring Kaggle will get you familiar with weird datasets, but actually collecting, arranging for data to be labelled, and then cleaning it is a skill on its own.

That's usually where experience comes in. Hence there are a lot of startups charging frankly obscene amounts of money for data labelling and curation tools.