Two simulators disagreed 1.8x on one policy's torque demand - four explanations were eliminated with numbers, the surviving suspect was never tested, and neither reading was allowed to cancel the other
torque-disagreement-between-simulators-unresolved.When two simulators disagree on a safety-relevant quantity, eliminate the measurement explanations one at a time with numbers, name the survivor as a hypothesis, keep the pessimistic reading binding, and run the direct test (replay one action sequence open-loop through both plants) before the quantity is used to pass a gate for hardware.
Symptom
For R3.0, Isaac read hip_pitch torque demand at 75-80% of the limit (PASS); MuJoCo read the same policy at 148-152% (1.5x over the limit).
Context
Candidates were eliminated one by one: PD gains (identical, RS06 kp 30 / kd 1.5), the settle window (both skip the first 0.5 s), the statistic (worse-of-two-legs vs per-joint - explains ~15%), self-collision (the no-contact subset still read 152%), sampling (500 Hz peak vs 50 Hz sample - ~9%). About 1.8x remained. The prime suspect - Isaac's implicit actuator (PD solved inside the PhysX integration, chosen to mimic kHz firmware PD) against MuJoCo's 500 Hz explicit PD - was named but untested.
Change
The Isaac PASS was recorded as valid for the Isaac plant only; the prescribed decider was an open-loop replay of one action sequence through both plants, compared step by step, needing no training. It was upgraded to a precondition of R3.1.
Outcome
R3.1 ran in parallel with, not after, the replay, and the replay never appears as done. By §40 it was "the oldest open account" on the line, to be fed by real-robot logs - which the spec also never records. Meanwhile the 50 Hz sample under-read the peak by 34% as smoothing narrowed the spikes, so Isaac-side readings grew more optimistic exactly when the comparison mattered more.
Mechanism
Each measurement artifact explained a slice of the gap; what remained is a plant difference, and an untested plant hypothesis cannot license discarding the pessimistic simulator on a safety-relevant quantity.
Conflicts
The spec prescribes the open-loop replay (§23), upgrades it to an R3.1 precondition, then records that R3.1 ran in parallel without it (§24), and lists it as the oldest open account before hardware (§40); nothing through §50 (2026-08-14) records a result. Every later "torque gate PASS" on this line is an Isaac-plant reading plus a MuJoCo check, never a reconciled one.
Applies when
- Isaac and MuJoCo (or sim and hardware) report different torques, contacts or slips
- implicit vs explicit actuator models are in play
- a gate passes on one simulator only
“**同一个策略,一侧判 PASS 一侧判超限 1.5 倍。** … 口径项全部扣掉后仍剩 **~1.8×** 没有解释。 … 但**这条没有验证,不能拿它当结论去抵消 MuJoCo 的读数**。 … (做法:拿同一条 动作序列在两边开环回放,逐拍比 τ —— 不需要重训)”