A torque-tail penalty was paid for by bracing the legs against each other - the second simulator's leg-contact count caught it, and the first explanation ("the trainer can't see self-collision") was retracted from the run's own config
torque-penalty-bought-by-leg-bracing.When a penalty lowers a demand metric, look for what the policy traded to get there - keep self-contact frames and foot spacing as standing sim2sim readouts - and check any "the trainer cannot see X" explanation against the run's resolved config before it enters the record.
Symptom
After R3.1's torque_headroom term collapsed the demand tail, MuJoCo success fell 100 -> 98% and leg-leg contact frames at mu 1.0 rose 750 -> 2,190 (worst rollout 177 -> 450). The one failure (prone seed 2) had the legs crossed, one foot on the other leg, trapped at 0.067 m - visible on video.
Context
Across the ten prone seeds, foot spacing and leg-leg contact frames were monotonically anti-correlated, and the failure was the extreme of the series. Pulling the legs toward the midline shortens the hip_roll lever arm and lowers torque demand. At the time the spec explained it as Isaac training without self-collisions ("a free lunch in a simulator without self-collision").
Change
Leg-leg contact frames and foot spacing were tracked in every MuJoCo gate; R3.2's candidates were "train with self-collision on" or "a minimum leg spacing term" - not stacked.
Outcome
The next rung's joint-velocity penalty incidentally erased the dependency (2,190 -> 86 frames). On 08-10 the runs' logged env.yaml showed enabled_self_collisions true in both r3_1 and v2_2 (inherited from walk v10): the tangle was physically learned bracing, visible to both simulators, and the Isaac/MuJoCo contact-count gap was mesh and contact fidelity. The "self-collision debt" narrative was withdrawn for the whole line.
Mechanism
A penalty on demand rewards any configuration that lowers demand; legs pressed together act as a mutual support that fails when contact geometry shifts slightly.
Conflicts
§24 attributes the dependency to self-collisions being disabled in training; §36 retracts that from the runs' env.yaml ("§24's mechanism explanation was wrong") and keeps the older sections unedited as history.
Applies when
- a torque, impact or energy penalty improves its metric and cross-sim success drops
- legs or links approach each other after a regularization change
- an explanation relies on a simulator setting nobody checked in the run config
“prone 十个 seed 逐条看,脚距与腿-腿接触帧数单调反相关, 而唯一失败的那条正是最极端的一条 … 机制上说得通:把腿收到身体中线附近能缩短 `hip_roll` 力臂、降低力矩需求”