Basis-64: data diversity confirmed — 9/14, new champion
X2 SONIC · basis-64 · official recipe · 1536 env × 64 steps · 1×RTX 3090 · 2026-07-08
data diversitycadencechampionWASD
1 · Result
Core-16 exam: 9/14 pass (all-time high), mean |cadence−1| 0.151 (all-time best). Headline: jog_backward_start passes at exactly 1.00 — this clip never passed in any prior run; backward stop also exact 1.00. Data diversity is confirmed as the binding constraint.
Overnight continuation from the basis-16 finale checkpoint (iter 3000 → 5000), same official recipe, data widened 16 → 64 clips from the same corpus: +10 sideways, +17 start/stop variants, +8 turns, +6 loops, and the backward family oversampled ×2. Examined on the unchanged core-16 ruler, so every number is directly comparable to the basis-16 campaign.
gate
pass /14
mean |cadence−1|
start = basis-16 @3000
8
0.219
basis-64 @3500
7
0.225
basis-64 @4000
6
0.194
basis-64 @4500
8
0.151
basis-64 @5000
9
0.151
2 · Clips
jog_backward_start · cadence 1.00 (IQR 0.08) — first pass ever for this clip, in any run. Left = reference, right = policy.turn_walk_270 · cadence 1.00 — same clip as the basis-16 finale QA set, for direct visual comparison.
3 · What resisted, and notes
sideways-45° (0.54) did NOT respond to +10 sideways variants — suspect clip quality or embodiment, needs a dedicated look rather than more data.
jog_backward_loop oscillates (0.36 → 0.79 → 0.68); idle_turn 1.50; two start clips sit at the 1.25 line.
OOM note: 64 motions at 2048 env blew the 24 GB card at init; 1536 env × 64 steps runs at ~13 GB. Batch stayed ~16× the June-era runs.
Extrapolation: this is the single-GPU version of the scale-up thesis — official recipe + wider data keeps paying. The full-corpus multi-GPU run is the straight continuation.
SBS clips: left = reference motion data, right = policy rollout (Isaac Lab, full physics) · yellow ball = reference anchor · all runs single RTX 3090