While working on PULSE, I found out that you CAN train a single MLP to reach very very high imitation success rate on AMASS with the right training procedure.
Basically, no MCP/MOE is needed as long as you train it for long enough...
The PNN stuff is still very useful for failure state recovery, which degrades the imitation result if not handled properly.
Code is out for training a single MLP model just for the imitation task (no recovery), and I am training a model I can release: