Default WBC stack

Default WBC stack#

Task id: Wbc-G1 · In-tree reference on robot g1.

Default WBC + deploy stack — ZEST Table S4 tracking on all motion keybodies, reward-aligned RSI, deploy-style actor obs.

  • Preset: apply_wbc (presets/wbc.py) — see Tasks and presets

  • Builder: g1_wbc_env_cfg() = g1_base_cfg() + apply_wbc(...)

  • Paper link: ZEST Table S4 (with deploy extras) — arXiv:2602.00401

What this task changes#

Rewards#

  • Table S4 on all 14 motion keybodies (G1_MOTION_BODY_NAMES in core)

  • Extra: motion_body_lin_vel / motion_body_ang_vel @ 0.5

  • motion_joint_vel, foot_slip, angular_momentum @ 0

  • Regularizers: survival 1.0, action_rate_l1 −0.1, joint_acc −5e−6, joint_limit −1.0, actuator_torque_soft_limit −0.1

RSI#

Reward-aligned similarity_ema — RSI (reference-state initialization).

Terminations#

  • anchor_pos z 0.35

  • ee_body_pos z-only on ankles + wrists (0.25)

  • keybody_ground_contact_force > 2000 N

Actor observations#

  • Removes ref_joint_vel (deploy-style)

Dim rules and preset deltas: Observations.

Train & play#

uv run wbc-mjlab-train --task Wbc-G1 --dataset samples
uv run wbc-mjlab-train --robot g1 --dataset samples   # shorthand
uv run wbc-mjlab-play --task Wbc-G1 --dataset samples --viewer viser

Logs: logs/rsl_rl/wbc_g1/<run>/ · params/policy.onnx · params/config.yaml

Tips#

  • Best starting point for multi-clip training and hardware deploy export.

  • Compare against ZEST reproduction for ZEST paper repro (4 EE bodies, no EE-z term).

  • Add --cache-motion-bundle on train/play for faster restarts.

  • See Deploy export for export handoff.