Two hardware complaints went in — a backward-command flip-over and curb hesitation on the deployed tall36 policy. Five single-variable training arms came out the other side with the causes separated, two instruments corrected, and two deployment candidates.
0.36 stance + sharpened tracking kernel + slow-command practice + gains covering the deployed 28/0.7. Best thermals ever measured here; no bottom-tier row in any test. Cost: tall36-level low swing.
Arm 4 + foot-clearance weight −2.5. Family-best tracking and shove tolerance — but it stalls at a 0.10 m step riser, so it is not the curb policy its swing numbers implied.
Deploy-contract observation path, DC-motor torque derating, 10 ms actuation delay. Four episodes per condition; zero falls anywhere, so distance is the discriminator. tall36 is the hardware-proven baseline.
MuJoCo thermal model at 1× (real time constants), motors from 42 °C, no perturbation. Rankings are meaningful; the absolute minutes are sim, not hardware duty predictions. The whole tall lineage is 3–5× cooler than the previously deployed PRD13d; foot-clearance work costs ~6–8 min at either stance height.
The original complaint. Cause: under the soft tracking kernel, standing still already pays ~73% of the tracking reward at a 0.3 m/s command — and this lineage grows up with a stand-still term, so doing nothing is in its repertoire from birth. Fix: kernel annealed 0.5 → 0.25 + half of training commands drawn slow (arm 2 onward).
| policy | MuJoCo −0.5 cmd achieved m/s | MuJoCo −1.0 cmd achieved m/s | Isaac −0.3 cmd tracked |
|---|---|---|---|
| tall36 (deployed) | −0.12 | −0.74 | 0% |
| arm 1 0.34 + gains | −0.25 | −0.78 | — |
| arm 2 + sharp kernel | −0.38 | −0.85 | 83% |
| arm 3 + clearance @0.34 | −0.35 | −0.80 | 80% |
| arm 4 | −0.38 | −0.82 | 73% |
| arm 5 s2 | −0.35 | — | — * |
| PRD13d (pre-thermal deploy) | −0.34 | −0.78 | — |
* seed 1 of the same config tracked 100% at −0.3 in Isaac; seed 2's backward render was not run. Its stop-from-speed behavior (the run0078 flip trigger) is clean in both engines for every sharp-kernel arm.
Foot-swing height percentiles over identical Isaac rough-terrain renders. The terrace result above is why this metric no longer decides curb capability.
| arm | single change vs parent | what it established |
|---|---|---|
| arm 1 tall34_widegains |
base_height 0.36→0.34; gains DR widened to cover deployed 28/0.7 | Height was not the deadband cause. Silent regression found: tall36 had deployed at gains outside its training randomization. |
| arm 2 tall34_sharp |
tracking kernel annealed 0.5→0.25; 50% slow commands; gains raised not stretched | The deadband was the reward's, not the gait's — fixed by intervention. Cost: swing collapsed; sharp tracking buys precision with low, quick steps. |
| arm 3 tall34_clear |
foot_clearance weight −1.0→−2.5 | The clearance term was priced too cheap, not shaped wrong. Swing recovered past tall36; deadband fix held. |
| arm 4 tall36_sharp |
base_height back to 0.36 | The tall stance is the thermal mechanism (straighter knee, shorter gravity arm on the calves): 8.5→16.4 min. Deadband fix independent of height. |
| arm 5 s1 tall36_clear_s1 |
arm 4 + clearance −2.5 | Family-best Isaac numbers, yet 10/10 forward-initiation stands under the deploy-path harness. A seed's basin can hide a deploy trap Isaac metrics never show. |
| arm 5 s2 tall36_clear_seed2 |
same config, seed 2 | Trap gone (3/3 initiates) — it was the basin, not the recipe. Deploy-loop initiation test is now a mandatory battery row. |