When should you not use AI on your robot?
It's easy to get swept up in the hype around robotics right now. There is a big push for end-to-end learning, which means feeding in all of your data (or access to a simulation) to a neural network and getting it to figure out how to control the robot.
But is this always a good idea?
Wheels Are Easy To Control Without AI
If you can, just use wheels instead of legs.
Wheels can be controlled reliably with a PID loop and are pretty versatile. If you can design your robots environment so its flat, then you won't ever need to worry about stepping over things.
Robots Arms Can Be Programmed
Industrial robot arms have teach pendants.
These devices allow you to give precise control command to your robot. If you can engineer out the variability in your process, you won't need a fancy neural network to control it reliably and productively.
Path Planning, Obstacle Avoidance and Localization
Sometimes you need a robot to drive around and avoid obstacles.
You don't need a neural network for this if you can control the environment. You can use a 2D lidar, the ROS slam_toolbox and ROS navigation stack to get a collision avoiding, autonomous robot without needing to train a neural network.