16 foundation clips selected for keyboard driving: forward walk/jog (start / loop / stop), backward jog (start / loop / stop), sideways (45° / 90°), turning walk (270° / 360°), in-place turn, idle. Trained with the official reward recipe (zero overrides) at large batch (2048 env × 64 steps), resumed via checkpoint seeding.
Number on each chip = cadence ratio (policy steps / reference steps; 1.00 = perfectly in step). Green = mastered, red = not yet, grey = stationary clips excluded from the count.
| checkpoint | cadence pass | mean |cadence−1| |
|---|---|---|
| B4-7000 (modified recipe, 401-clip generalist) | 1/13 | 0.348 |
| basis-16 @2500 | 6/14 | 0.185 |
| basis-16 @3000 (finale) | 8/14 | 0.219 (median 0.13) |
Head-to-head at 2500: basis-16 wins 8 clips, ties 2, loses 3. The official recipe decisively beats the modified-recipe era on fidelity; remaining failures are per-family data gaps (backward, sideways, slow starts), not recipe problems.
Gates continue at 1500 / 2000 / 2500 / 3000. Cadence is measured with a Schmitt-trigger step counter on the left–right knee difference signal (hysteresis 0.1 rad), taken as the median over all evaluation envs — single-env readings proved unreliable (same checkpoint read 1.00 vs 0.75 on different env draws).