One more thing from yesterday.
What makes that update special is that it's zero shot.
The policy was trained entirely in simulation, then run on the real robot with no tuning in between.
And it ran in IsaacLab, which we only moved to this week. Different physics engine, same configs, worked first try.
The hardware seems to be doing the quiet work here. Motor communication and firmware behave the same every run, so sim and real stay close.