Here is an update on the progress of my ML efforts.
Started training a giant colossus sized rig (30 meters tall).
Finally upgraded to the latest version of ML Agents. Upgrading is something I generally do infrequently because it usually generates a lot of re-work, but there have been a lot of improvements added.
One improvement to the training scenes is the ability to layer animations over top of the machine learned movements. The animations can be played during training so that the models can learn to walk and do arm movements at the same time (without falling over). Here is the troll learning to swing his arm and and toss while walking without falling over. The semi-transparent rig on the left is the reference animation. The arms of the troll are trying to copy this motion. The legs are under control of the neural net.
One of the techniques I am using in the training scenes it to use one model to generate initial conditions for another, training, model. Here is tumbling, and jump being used to generate start conditions for the walk model. The walk needs to learn to take-over/recover from the transition position.
