| <!doctype html> |
| <html lang=en><meta charset=utf-8> |
| <meta name=viewport content="width=device-width,initial-scale=1"> |
| <title>Stage-1 gate — Cosmos</title> |
| <style> |
| body{font:15px/1.6 -apple-system,BlinkMacSystemFont,"Segoe UI",sans-serif; |
| max-width:820px;margin:40px auto;padding:0 20px;color:#1f2328} |
| h1{font-size:22px;margin-bottom:4px} .sub{color:#57606a;margin-bottom:24px} |
| .verdict{display:inline-block;padding:6px 14px;border-radius:6px;color:#fff; |
| font-weight:600;background:#bc4c00} |
| table{border-collapse:collapse;width:100%;margin:22px 0} |
| th,td{padding:9px 10px;border-bottom:1px solid #d8dee4;text-align:left;vertical-align:top} |
| th{font-size:12px;text-transform:uppercase;color:#57606a;letter-spacing:.04em} |
| .num{text-align:right;font-variant-numeric:tabular-nums} |
| .muted{color:#57606a;font-size:13px} |
| .win{color:#1a7f37;font-weight:600} .lose{color:#cf222e;font-weight:600} |
| .note{background:#f6f8fa;border-left:3px solid #d8dee4;padding:12px 16px; |
| margin:20px 0;font-size:14px} |
| </style> |
| <h1>Stage-1 gate — does the action-motion encoder transfer to Cosmos?</h1> |
| <div class=sub>tokenizer <code>cosmos-cv4x8x8</code> · generated 2026-07-31 07:55:33Z</div> |
| <p><span class=verdict>NO TRANSFER</span> the encoder learns well (D beats zero by 32.3%, C by 33.0%), but the action map adds NOTHING over tactile-only (D is 1.0% worse than C; Wan2.2 had D +3.3% over C). The conditioning signal Stages 2-3 inject is redundant here.</p> |
| <div class=note> |
| The encoder predicts per-latent-transition tactile flow. <b>Active-pixel EPE</b> is the |
| endpoint error over pixels whose ground-truth motion exceeds the active threshold; |
| lower is better. The zero-motion baseline on this val split is |
| <b>1.2195</b> (Wan2.2: 1.612). |
| An ablation is only meaningful if it beats that. |
| </div> |
| <table> |
| <tr><th>ablation</th><th>inputs</th><th class=num>EPE (Cosmos)</th> |
| <th class=num>EPE (Wan2.2)</th><th class=num>Δ</th><th class=num>trained</th></tr> |
| <tr><td><b>D</b><div class=muted>(z_ctx, a) -> F</div></td><td class=muted>action map + tactile state</td><td class='num win'>0.8254</td><td class='num muted'>1.130</td><td class='num muted'>-0.305</td><td class='num muted'>12 ep</td></tr><tr><td><b>B</b><div class=muted>action only</div></td><td class=muted>action map only</td><td class='num win'>1.0553</td><td class='num muted'>1.362</td><td class='num muted'>-0.307</td><td class='num muted'>12 ep</td></tr><tr><td><b>C</b><div class=muted>tactile only</div></td><td class=muted>tactile state only</td><td class='num win'>0.8170</td><td class='num muted'>1.168</td><td class='num muted'>-0.351</td><td class='num muted'>12 ep</td></tr> |
| </table> |
| <div class=note> |
| <b>Δ is context, not a result.</b> The two columns are not measured over the same |
| interval: a Wan2.2 latent transition spans 5 frame pairs at 6 fps (0.83 s), |
| a Cosmos one spans 4 pairs at 30 fps (0.13 s) — <b>6.2× less motion |
| time</b> — on top of a different tokenizer and dataset. The magnitudes happen to |
| land in a similar range because gel deformation saturates rather than accumulating, |
| but that is a coincidence of contact mechanics, not a controlled comparison. |
| <b>Judge the gate on the ordering</b> (D < C < B < zero) <b>and on D's |
| relative margin over zero</b>, both measured entirely within the Cosmos column. |
| </div> |
| <div class=note><b>Seeds.</b> 3 seeds per ablation. The D-vs-C increment is -0.9% ± 0.6% (std). The sign is stable across seeds.</div> |
| <h2 style="font-size:17px;margin-top:32px">What the two ablations actually see</h2> |
| <p>D's input is a strict superset of C's: the same tactile latent, plus the action |
| map. <b>D − C is therefore exactly what knowing the sensor's own motion buys |
| you beyond looking at the gel</b> — the signal Stages 2-3 inject.</p> |
| <img src="stage1_cosmos.png" style="width:100%;border:1px solid #d8dee4;border-radius:6px"> |
| <div class=note> |
| The action map (purple) is a near-<b>constant</b> field: 99.5% of its energy is |
| global translation, so it carries about one 2-D vector. The AllTracker target |
| (red) is only 33% global translation — two thirds is spatial structure a constant |
| field cannot express. That is why B (action only) is far behind, and why D adds |
| nothing C does not already have. |
| </div> |
| <h2 style="font-size:17px;margin-top:32px">Cosmos vs Wan2.2</h2> |
| <img src="stage1_cosmos_vs_wan.png" style="width:100%;border:1px solid #d8dee4;border-radius:6px"> |
| <div class=note> |
| The ordering broadly agrees — action-only weakest, tactile dominant. What differs |
| is the increment the recipe rests on. Note <b>D did not get worse</b> on Cosmos |
| (29.9 → 32.3): <b>C got better</b> (27.5 → 33.0). The Cosmos tactile |
| latent already carries more of the motion, leaving no headroom. And the increment |
| was only +3.3% on Wan2.2 to begin with. |
| </div> |
| <p class=muted>Raw metrics: <a href="data.json">data.json</a></p> |
| </html> |
