would pass the page gate in the drive expert of the rhythm set of g1ours
turn_clock (passed, score 1.0): rung_eval turn v10 in Mjlab-Velocity-Flat-Unitree-G1-Turn-Rhythm, 32 envs: in place 94% of commanded yaw (left 105/99%, right 84/89% at 0.5/1.0 rad/s), drifting 0.12 m/s; on arcs 83% of yaw at 85% of forward speed; 0% of envs fell. #104: 0.198, failing.
drive (failed, score 0.977): rung_eval drive v10: forward 68% of command against the 70% bar, stop 0.7 cm/s residual, left 106%, right 84%, 0% fell
terrain (measured): scripts/terrain_eval.py at 0.6 m/s, cadence 1.5 Hz, 128 envs, 12 s, in Mjlab-Velocity-HillyHoldout-Unitree-G1-Rhythm (the page's map mirrored, never trained on): drive 74.5% of command, falls 0.008; #104 on the same spawns 75.1%, 0.008.
Open caveats
scored only by an older evaluator than v12, owed a re-score: cockpit v10, drive v10, turn_clock v10
red flag support_off_feet: 1 knocks end upright with a mean 12% of support off the feet
amber flag reward_blind_to_support: no positive reward term reads the feet or contacts (track_linear_velocity, track_angular_velocity, upright, pose, step_schedule, beat_lock): standing can be rewarded however it is held up
amber flag support_off_feet: 1 knocks end upright with a mean 44% of support off the feet (in a world a declared flaw waives FEET)
amber flag power_spike: peak 5331 W while another expert drove (knocks, get-ups): not this slot's, over the 3000 W supply peak
waiver in force: FEET 12/16 back on its feet alone, under this policy's 0.80 (14% of its support off the feet once up)
Lineage
Oldest first, each resumed from the one above. The first predates LifTrain's control DB: anything before it is not recorded.
before the DB 2026-10-15_g1-rhythm-warm · unrecorded · 0 iterations
the run's own end-of-run export: scripts/check_onnx.py --task Mjlab-Velocity-Terrain-Unitree-G1-Turn-Rhythm, max |torch - onnx| 1.91e-06 over 256 observations, metadata matching the task's layout