Overview#
How wbc-mjlab is put together — the ideas behind the commands. Read these when you want to change the stack rather than just run it.
The one-paragraph version#
One shared MDP (motion command, rewards, observations, terminations) is written once and knows nothing about any robot or paper. A robot plugs in its model, actuators and body names. A preset configures the shared terms the way a method needs (weights, sampling, observation layout). A task names one robot + preset stack so the CLI can train it. Training samples clips from a motion library, starting episodes where the policy fails most (adaptive sampling).
The four layers and what each one owns.
How a --task becomes an environment config; how the Studio is built.
Methods as config mutators; composing them; Studio presets.
Why robots register, and the ways to add one.
Where episodes start in a clip library, and why it matters.
Modular reference terms — why papers’ actor layouts are just term sets.