Run acceptance under measured contact parameters - honest condim/torsional-friction flipped a false PASS into a real-matching FAIL
eval-plant-honesty-contact-params.Pin the evaluation plant's contact model to measured values (contact dimension, torsional/rolling friction, mu) before trusting any gate that involves slip, impact, or drift - a gate can only fail a policy for physics its simulator contains.
Symptom
walk_v5 passed the old acceptance battery yet failed on the real robot (footfall force, drift, kicking) - the evaluation plant was flattering the policy.
Context
The battery was re-run under "honest contact parameters" - condim 4 (adding torsional contact) with measured torsional friction 0.035 - and v5 then FAILED exactly the rows corresponding to its real problems: heading 185 deg (limit 30), support-foot yaw slip 284 deg (limit 80), landing force 1.72x (limit 1.5x), suspended tilt 45.9 deg (limit 10). The slip physics depends on torsional friction, which the default contact model (condim 3) does not even simulate - a slip problem is invisible to an evaluator that cannot represent yaw friction at the foot. Term-sizing measurements for the new rewards were likewise taken under the same honest parameters (cmd 0.45, skipping the 5 s start transient).
Change
Acceptance harness pinned to condim 4 / torsion 0.035 (measured); verdicts issued under defaults declared non-citable for these rows.
Outcome
Sim acceptance verdicts began agreeing with hardware ("现在失败, 与真机一致"); the v6 fixes could be developed and validated against an evaluator that could actually see the disease.
Mechanism
An evaluator is a plant model too: contact dimensionality and friction values decide which failure modes exist in the simulation at all. Evaluating under default contact parameters tests the policy in a world where its real failure is physically impossible, producing structurally false PASSes.
Applies when
- sim acceptance passes policies that fail on hardware
- slip/drift/impact gates run under default simulator contact settings
- setting up a cross-simulator evaluation harness
“accept_v2.py 已加三条判据, walk_v5 在诚实的接触参数下(--condim 4 --torsion 0.035)现在失败, 与真机一致:直行 15s 航向累计 <30° | 185° ✗ … 落脚力峰值 <1.5× 体重 | 1.72× ✗”