BeyondMimic-style RSI

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

BeyondMimic-style adaptive RSI — body pose/vel rewards, binary_failure RSI, full base actor obs (includes ref_joint_vel).

• Preset: apply_binary_failure (presets/binary_failure.py)

• Builder: g1_wbc_binary_failure_env_cfg()

• Paper: arXiv:2508.08241

What this task changes

Mutates rewards + RSI only; keeps base terminations and full actor layout from g1_base_cfg() (including ee_body_pos and ref_joint_vel).

Actor dim rules: Observations.

Rewards

Term	Weight
motion_global_root_pos / motion_global_root_ori	0.5
motion_body_pos / motion_body_ori / motion_body_lin_vel / motion_body_ang_vel	1.0 (all motion keybodies)
motion_joint_pos / motion_joint_vel	0
survival	0
action_rate_l1 / joint_acc	−0.1 / −2e−6

RSI

• strategy="binary_failure" — failure on early termination

• similarity_terms = keybody_similarity_preset()

See RSI (reference-state initialization).

Train & play

uv run wbc-mjlab-train --task Wbc-G1-BinaryFailure --dataset samples
uv run wbc-mjlab-play --task Wbc-G1-BinaryFailure --dataset samples --viewer viser

Logs: logs/rsl_rl/wbc_g1_binary/<run>/

Tips

• Watch adaptive RSI panels in Viser — binary-failure bins differ from ZEST RSI.