ZEST reproduction

ZEST reproduction#

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

ZEST paper reproduction — Table S4 on end-effector bodies only, reward-aligned RSI, no EE height termination.

  • Preset: apply_zest (presets/zest.py)

  • Builder: g1_wbc_zest_env_cfg()

  • Paper: arXiv:2602.00401

Preset wiring#

The task builder calls the preset with robot-specific body tuples:

def g1_wbc_zest_env_cfg() -> ManagerBasedRlEnvCfg:
  cfg = g1_base_cfg()
  apply_zest(
    cfg,
    reward_body_names=G1_ENDEFFECTOR_BODY_NAMES,  # 4 EE bodies
    contact_body_names=G1_MOTION_BODY_NAMES,
  )
  return cfg

What apply_zest mutates on cfg — see Tasks and presets for annotated preset source and the preset ↔ task model.

Summary#

Rewards#

  • Same Table S4 σ/κ as Default WBC stack, but body pos/ori on 4 EE bodies only

  • Disables extra vel terms and foot_slip / angular_momentum (weight 0)

RSI#

Identical reward-aligned settings to Default WBC stack — RSI (reference-state initialization).

Terminations#

  • Removes ee_body_pos (no EE height cutoff)

  • Keeps anchor_pos + keybody_ground_contact_force on full keybody set

Actor observations#

Dim rules: Observations.

Dim rules: Observations.

Train & play#

uv run wbc-mjlab-train --task Wbc-G1-Zest --dataset samples
uv run wbc-mjlab-train --robot g1 --no-state-estimation --dataset samples
uv run wbc-mjlab-play --task Wbc-G1-Zest --dataset samples --viewer viser

Logs: logs/rsl_rl/wbc_g1_zest/<run>/

Tips#