humanoid-brisk-handoff

`/app/work/controller.py` now passes the `brisk` rung on 100 of 100 fresh seeds (200–299) on the nominal plant. That is above the 0.9 gate. Mean speed is about 0.91 m/s, and it doesn't reach `fast`, which needs 1 m/s. I only ran the nominal `standard_v0` scenario. I haven't run the perturbation scenarios, so I can't say how it holds up on `push_v1` or `weak_motors_20`. The inherited walker stayed upright on every seed but averaged only 0.31 m/s, about 4.7 m in 15 s, so it failed `brisk`. I kept its structure and changed only the parameters. The tuner accepts at most 12, so I split `PARAMS` into a tuned set of 12 and a `FIXED` dict for the rest, and I added a line to the docstring saying so. - **First tune:** on 6 seeds it reached `brisk` on about 92% of seeds, with falls at roughly step 250. In the one case I traced, yaw drifted, the gait sped up into a run and the robot pitched forward. - **Second tune:** I swapped in a different 12 parameters, including the torso, yaw, roll and lateral-placement gains, and tuned on 40 seeds. It got 100% on a separate block of seeds 100–139 and then on the 100-seed check above. `v2.py` (the 92% version) and `v3.py` (a copy of the current controller) are scratch copies in `/app/work`, and `roll.py` is a debug rollout script. About 21 minutes of the 60 were used.

The final controller is in place at `/app/work/controller.py` and the budget is nearly spent (started 12:25, now 13:10). **Outcome.** The inherited SIMBICON walker went from 0.31 m/s (0 of 5 brisk) to about 0.65 m/s. The official evaluator gives 5 of 5 brisk on the visible seeds with a mean of 10.3 m and a return of about 5660. On 180 seeds in total the brisk rate is about 88%, which sits right at the 0.9 gate rather than clearly above it. | Seed set | Brisk | |---|---| | 0 to 19 (official evaluator) | 19 / 20 | | 20 to 179 (fresh, in-process) | 140 / 160 | **What changed in the controller.** The structure is the handed-over design, restructured to 12 tunable gains plus fixed constants, with a docstring explaining each gain. - Torso lean raised from 0.008 to 0.06 rad. This is the speed knob and it alone took the walker from 0.3 to 0.65 m/s. - Torso pitch damping raised to 45, the single most consistent robustness gain on fresh seeds. - Placement gains re-tuned by the official tuner on 16 seeds, then checked by hand on 120 held-out seeds. **What I found and why the gate is tight.** The gait is chaotic per seed: changing the speed target by 0.0004 m/s flips half the seeds, so every candidate had to be judged on 80 to 160 fresh seeds, and the tuner's own leaders failed those checks twice. The remaining failures are mid-episode forward falls. The hip rotation motors chatter between the command limits every step because their damping gain is far too high for the leg's inertia,…