Actions#
Manager: ActionManager · Module: env/mdp/actions.py
Default action#
ReferenceJointPositionActionCfg — the policy outputs residual joint deltas
Δq that are added to reference joint positions from the motion command:
q_target = q_ref(motion command) + scale · action
This keeps the action space centered on the current clip frame rather than absolute joint targets.
Per-robot wiring#
Each robot entity sets actions["joint_pos"].scale to a per-DoF or grouped scale
tensor (e.g. <ROBOT>_ACTION_SCALE in the robot package).
Alternative#
make_base_wbc_env_cfg(use_reference_residual_action=False) switches to mjlab’s
absolute JointPositionActionCfg with default joint offsets.
apply_end_effector swaps to DefaultOffsetJointPositionActionCfg:
q_target = q_default + scale · action
On RSI resample (training), a is seeded to (q_ref − q_default) / scale
so the first PD target matches the clip. Deploy uses the same action equation
with measured joints, not q_ref.
API: MDP terms (Actions).
See also