Wbc-G1-Lookahead — the SuperMimic paper’s task#
Task id: Wbc-G1-Lookahead · In-tree reference on robot g1.
A policy that looks ahead: it sees 10 steps of proprioception and the next 10 reference frames, with BeyondMimic’s tracking reward, actions about the default pose and episode starts where it earns low return. It is the task of the SuperMimic paper (Appendix A and D.1); the name is the task’s, not the method’s. Trained with the world model (World-model training) it is SuperMimic; with PPO, its baseline.
Preset:
apply_lookahead(presets/lookahead.py)Builder:
g1_wbc_lookahead_env_cfg()Paper: arXiv:2610.28480
uv run wbc-mjlab-train --task Wbc-G1-Lookahead --dataset lafan_seed --algo wm # SuperMimic
uv run wbc-mjlab-train --task Wbc-G1-Lookahead --dataset lafan_seed # PPO
Logs: logs/rsl_rl/wbc_g1_lookahead/<run>/
Note
The reference terms look 10 frames ahead (Looking ahead: frames and stride): the deploy
parameters record frames: 10 for each, so a robot runtime has to feed the next 10
reference frames. A preset can change the window per term (fewer frames, a stride, other
reference terms) like any other observation.
Observation (Table A.1)#
Term |
Noise |
Notes |
|---|---|---|
|
— |
tracked joint positions · velocities of the reference, 10 frames from the current
one ( |
|
— |
the reference anchor’s orientation in the robot anchor frame (6D), 10 frames |
|
— |
the reference anchor’s displacement from where the reference is now, in the robot anchor frame, 10 frames |
|
±0.01 · ±0.5 |
relative to the default pose; 10-step history |
|
±0.2 · ±0.05 |
IMU; 10-step history |
|
— |
previous action; 10-step history |
The actor needs only the robot’s joints and IMU (no state estimation). Actions:
q_target = q_default + scale · a.
Reward (Table A.2)#
All terms exp(−mean_b ‖e_b‖² / σ²) over the tracked bodies, × the control period.
Term |
σ |
Weight |
|---|---|---|
|
0.3 · 0.4 |
1.0 |
|
1.0 · 3.14 |
1.0 |
|
0.3 · 0.4 |
0.5 |
|
— |
−0.1 · −2 |
Termination, randomization, starts#
Termination: anchor height off by more than 0.25 m, or tilt (gravity z in the anchor frame) off by more than 0.8; no end-effector termination.
Randomization (Table A.3): foot friction U[0.3, 1.2], encoder offsets ±0.01 rad, torso center-of-mass offset; no pushes, no assistive wrench.
Starts (Appendix D.1): 1 s bins, each with a running estimate of the discounted return earned from it; starts are 50 % uniform, 40 %
softmax(−25 G / G_max), 10 % in proportion to each bin’s past softmax weight (rsi.strategy = "discounted_return", RSI configuration).