Driving a car is in many ways harder than detecting subs. In particular, the car needs to make a lot of fine-grained decisions very fast in a feedback loop that affects what it sees. A sub-detecting algorithm only needs to output Option<vec3>. There's a lot of noise, yes, but I don't think it's crazy to think that can be solved by just having a honking huge model and adding more data, e.g. chemical traces, surface lidar, etc...