would pass the page gate in the get-up expert of g1ours
standup_hilly (passed, score 1.0): rung_eval standup v12 in Mjlab-StandUp-Hilly-Unitree-G1-Rise, unaided: 93% of 196 episodes (3 per env x 64) end up and on the feet alone (0% of support off the feet once up); rises from 94% of non-standing starts (every kind 87-100%), median rise 0.54 s, stays up from 91% of standing starts; 132 W mean, 2350 W p95 peak (g1-getup-rise-s2 on the same rung: 1.000, 125 W mean, 2937 W p95 peak). On flat -Reverse 1.000: every episode ends up on the feet alone, 109 W mean, 2366 W p95 peak.
page_gate (passed, score 1.0): scripts/page_gate.py q159 in the get-up slot of mjweb-play's G1 (ours), mujoco 3.17.0, the turning walker g1-walk-turn-s1 driving: RECOVER knocks at 12x body mass, 16 seeded headings per world, end on its feet alone 15/15 on domino_land, 14/15 on hilly and 12/12 on halloween (the rest shrugged off), with no waiver. Cockpit and drive unchanged (the walker's). obsguard: page and gate agree. Red flags: the get-up drew a 3420 W peak in 1 sample over the 3 kW supply guess (g1-getup-rise-s2, q131: 4514 W in 11), so stage 2e's own gate, no power_spike, is not met; it is served for the drop from 4514 W in 11 samples to 3420 W in 1; and 14 of 44 knocks were partly carried by the stander (q131: 12 of 43, the same harness artefact); evidence in scratch/gate/q159/.
Open caveats
red flag expert_share_mismatch: 14 of 44 trials judged on the standup expert were partly driven by another (e.g. shove world_scene_domino_land.xml s: drive 2%, stand 89%, standup 9%; its result counts as the standup expert's)
red flag power_spike: peak 3420 W while the standup expert drove, over the 3000 W supply peak (data/rung-budgets.json), in 1 samples
amber flag identical_across_seeds: all 8 envs end with exactly the same support (0.0): the starts may not have varied
amber flag reward_blind_to_support: no positive reward term reads the feet or contacts (upright_progress, height_progress, standing, rise_level, upright_one_sided): standing can be rewarded however it is held up
amber flag power_spike: peak 5331 W while another expert drove (knocks, get-ups): not this slot's, over the 3000 W supply peak
Lineage
Oldest first, each resumed from the one above; the first trained from scratch.
scripts/export_onnx.py (the run left no export), through scripts/page_gate.py: max |torch - onnx| 9.54e-06 over 256 observations (scripts/check_onnx.py --task Mjlab-StandUp-Terrain-Unitree-G1-RisePower), metadata identical to g1-getup-rise-s2's but for run_path