I've always thought that building a MUD AI would be extremely fun, and a good test for language understanding and reinforcement learning.

Here are a couple of related papers in the deep learning space:

https://arxiv.org/abs/1812.01628 https://arxiv.org/abs/1903.03094