Newly trained walker agent skips using default setup

I used the default setup for the Walker example in the official ML Agent repo.
When I try to retrain the agent, I don't get the same result as the pre trained model in the repo.
The biggest issue is the agent skips instead of alternating steps.
I'm curious if there's documentation on how that model was trained. Is there a more consistent way to get the character to walk in a more human-like fashion?