End-effector actor command#
Task id: Wbc-G1-EE · In-tree reference on robot g1.
Same WBC rewards / RSI / terminations as Default WBC stack. The actor drops
ref_joint_pos and instead sees the existing keybody pose reference
(ref_body_pos / ref_body_ori) on wrists and feet. The critic and
joint-tracking rewards still see the motion joint reference during training.
Presets:
apply_wbc→apply_end_effectorBuilder:
g1_wbc_ee_env_cfg()
What this task changes#
Actor observations#
Removes ref_joint_pos / ref_joint_vel. Adds the already-defined
critic terms ref_body_pos / ref_body_ori (same functions, no MDP
changes). These are reference keybody poses — not current robot state.
Critic and rewards#
Unchanged from Default WBC stack.
Actions#
Switches to default-offset actions (q_cmd = q_default + scale · a). After
RSI, a is seeded from q_ref for the first frame only. Deploy does not
need the joint stream.
Dim rules: Observations.
Train & play#
uv run wbc-mjlab-train --task Wbc-G1-EE --dataset samples
uv run wbc-mjlab-play --task Wbc-G1-EE --dataset samples --viewer viser
Logs: logs/rsl_rl/wbc_g1_ee/<run>/