It walks again — better tracking, less terrain

Three bugs found, all real

1. A division by zero only standing commands could reach. reward_tracking_goal_vel is min(proj_vel, cmd) / (cmd + 1e-5). With a zero command, any backward drift returns about −10⁵. Mean reward went to −1241 the moment standing commands were enabled. Table S3's r_lv carries exactly this branch, for exactly this reason.
2. The reward clip cut both ways. Earlier I found its harm — an unavoidable penalty above the positive ceiling zeroes the gradient — and removed it. That missed the other half: it also floors the cost of walking at zero. Unclipped, a policy exploring toward a gait pays every collision and torque penalty en route while standing avoids them all. The one arm that walked had the clip on; all six that stood had it off.
3. Three penalty sets, not two. Restoring the Extreme Parkour penalties while leaving in the Table S3 terms added to replace them meant paying three sets at once. Mean reward pinned at exactly 0.00 — the blackout, reintroduced by me, one iteration after writing the section that diagnosed it.

Where every configuration landed

A split, not a win. At matched commands (0.3, 0.8) the library arm tracks velocity far better — 98% of commanded speed against 56%, and 1.7× the absolute speed — while the baseline still climbs about 9% more terrain. That is consistent with what the prior is FOR: the library encodes periodic gaits at commanded speeds, so it should buy tracking; it encodes nothing about obstacles, so it should not buy terrain. An earlier version of this page claimed an outright win by comparing terrain across two different command ranges — terrain is scored relative to commanded speed, so it is not comparable that way, and the baseline itself reads 2.103 at (0.3, 0.8) versus 1.815 at (0.0, 1.5). The torque-split audit still has to run: these numbers say the policy works, not that the decoder is what drives it.

The three kept changes are genuine improvements — roughly an order of magnitude in sample efficiency over the original walking arm. The four reverted ones (retiring the EP penalties, widening commands, removing the clip, zeroing gait-logit entropy) are what broke locomotion. That is attributable now only because exactly one group differs between the top two rows.

Next, one variable at a time

Base task is …-MotionLibWalk-v0. Each reverted change goes back in separately, cheapest hypothesis first:

1. Gait-logit entropy zero and log(2) initial variance — pure exploration change, no reward rebalance, so it cannot reintroduce a blackout.
2. gait_diversity −1 → −15 — the actual fidelity question. Check mean reward ≠ 0.00 within 50 iterations.
3. Command range (0.0, 1.5) — needs the zero-command tracking branch, already in.
4. EP-penalty retirement paired with its Table S3 replacements — the largest rebalance, so last.