It's really easy to look at a container ship and say, "autopilot has already been done, lets just apply that." I think the initial bit of getting the boat to drive between waypoints is a pretty easy problem, but once you put this boat out on the open ocean and something goes wrong, you're going to need some very serious fault handling, which means built-in redundancy and consideration for a huge number of possible failures. What happens when you catch RudderNotRespondingException("I tried to turn the rudder to 24.3 degrees and nothing happened.") Most likely you lost feedback, but maybe the rudder actually didn't turn to 24.3 degrees. When does the autopilot realize that it can't make it to the destination. Does the ship stop and wait for a tech to be helicoptered in to plug the rudder feedback cable back in?

Thats one of about 1000000 things that could go wrong that an experienced engineer would have little trouble diagnosing and fixing or working around, that are much more significant challenges for a computer making decisions. Just looking at the complexity of the engine room you can imagine all the possible points of failure, and the number of required contingency behaviors for the autopilot explodes when you start taking 2, 3, or more at a time.