I ran a simplified 2D robotics contest to automate code where the task was a simple 2D javascript canvas with a robot that had just two segments that could move (the arm and elbow). A piece of fruit was placed somewhat randomly within its reach, and when its hand was close enough it could be instructed to grasp the fruit. (No grasping code needed to be written, the hand just needed to be close enough) The task was then to bring it to its mouth.
There were no sensors involved, the positions were available as global values.
About twenty people looked at the contest (number of concurrent viewers on the dedicated subreddit I posted it on) but nobody submitted a successful entry. When I tried to solve it myself, I found I couldn't automate the task either, neither geometrically nor with a neural network, which turned out to be too difficult to train. The closest I got was being able to grasp the fruit about 7 out of 8 times or so (I forget the exact number), without being able to bring it to the mouth successfully.
By contrast, solving it by hand on the keyboard (moving the two joints and grasping using keyboard shortcuts) like remote controlling the robot (with a human in the loop) is easy.
I had no idea that even such a toy problem with a tiny 2D universe, only two joints, and perfect information, would prove to be so difficult to automate without a robotics framework.