Presets

Presets#

Paper / deploy stacks as in-place env mutators. Task builders call one or more apply_* functions on the base template — no MDP forks.

Design: Presets and tasks. Task catalog: Choose a task.

wbc_mjlab.presets.apply_wbc(cfg: mjlab.envs.ManagerBasedRlEnvCfg, *, motion_body_names: tuple[str, ...], ee_termination_bodies: tuple[str, ...]) → None[source]#

Default WBC stack: Zest tracking + RSI on all keybodies, EE height, light regularizers.

Parameters:

cfg – Env config to mutate in place.
motion_body_names – Keybodies for tracking rewards / RSI.
ee_termination_bodies – Bodies for z-only EE height termination.

wbc_mjlab.presets.apply_zest(cfg: mjlab.envs.ManagerBasedRlEnvCfg, *, reward_body_names: tuple[str, ...], contact_body_names: tuple[str, ...], force_threshold: float = 2000.0) → None[source]#

Zest paper repro: no actor SE terms, reward-aligned RSI, assistive wrench.

Parameters:

cfg – Env config to mutate in place.
reward_body_names – Keybodies for tracking rewards.
contact_body_names – Bodies for keybody ground-contact termination.
force_threshold – Contact force threshold (N).

wbc_mjlab.presets.apply_end_effector(cfg: mjlab.envs.ManagerBasedRlEnvCfg, *, body_names: tuple[str, ...] | None = None) → None[source]#

Swap actor joint-reference command for keybody pose refs.

Leaves rewards, RSI, and terminations untouched. Critic keeps ref_joint_pos and full ref_body_*.

Parameters:: body_names – Bodies included in actor ref_body_pos / ref_body_ori. None uses all motion-command bodies. Pass end-effector names to restrict the actor command to EEs only.

wbc_mjlab.presets.apply_se_actor(cfg: mjlab.envs.ManagerBasedRlEnvCfg) → None[source]#: State-estimation actor layout: drop height/gravity proxies; add anchor pose error + base lin vel.

wbc_mjlab.presets.apply_binary_failure(cfg: mjlab.envs.ManagerBasedRlEnvCfg) → None[source]#: BeyondMimic-style stack: full obs, binary-failure RSI, assistive wrench.