Shared MDP

Shared MDP#

Design and term catalogs for the shared manager-based MDP (how terms compose, presets, and robot wiring). For live signatures and docstrings, see MDP terms.

wbc-mjlab builds on mjlab’s composable RL environment API: each MDP facet (actions, observations, rewards, …) is a set of named terms wired on ManagerBasedRlEnvCfg. Term callables live in env/mdp/; managers invoke them each step and aggregate batched tensors.

For mjlab’s upstream environment docs, see mjlab documentation.

Design rules#

Rule	Implication
Terms live in `env/mdp/`	Robot-agnostic; parameterized by `command_name`, body lists, weights
Template in `wbc_env_cfg.py`	Default term slots; robots fill entity-specific params
Presets mutate cfg	Change weights, thresholds, or remove terms — not Python forks
Robot entity wires names	Anchor body, keybody tuple, action scale, sensors — see The robot entity

Environment configuration#

make_base_wbc_env_cfg() returns a ManagerBasedRlEnvCfg with all manager groups pre-populated. Robot builders (<robot>_base_cfg) attach the entity and set motion-command body names; presets finish the recipe.

Term modules (`env/mdp/`)#

Module	Manager / topic
`actions.py`	Actions
`commands.py`	Motion command
`observations.py`	Observations
`rewards.py`	Rewards
`terminations.py`	Terminations
`assistive_wrench.py`, `sampling.py`	Events, RSI (reference-state initialization)
`metrics.py`	Logging / evaluation helpers

Import surface: wbc_mjlab.env.mdp (re-exports mjlab MDP helpers plus wbc terms).

MDP components#

Manager	wbc-mjlab contents
Actions	Reference-residual joint position (default)
Commands	`MotionCommandCfg` — multi-clip playback + RSI
Observations	Actor: reference + proprio. Critic: actor + privileged features
Rewards	`motion_*` tracking exponentials + regularizers
Terminations	Timeout, anchor drift, EE height, contact force
Events	Assistive wrench, push, COM, friction, encoder bias

Components