force recovery: methods, evaluation, debug log

results.html

@@ -69,11 +69,10 @@ better of the two. Nothing but the image→gradient step differs.</p>
 <figure><img src="assets/truncation.png" alt="truncated presses">
 <figcaption>What the row above excludes. A press is <b>truncated</b> when its
 contact core reaches an image border: part of the indentation is outside the
-frame, the free-boundary solve has a slope running off the edge and nothing on
-the far side to stop it, and the height ramps. 14.5% of truncated frames
-reconstruct deeper than the 4.25&nbsp;mm gel is thick, against 0%
-of frames imaged whole. Their depth is not identifiable from the image, so no
-reconstruction can fix them.</figcaption></figure>
+frame, so the free-boundary solve runs off the edge with nothing beyond it to
+stop the ramp. 14.5% of truncated frames reconstruct deeper than the
+4.25&nbsp;mm gel is thick, against 0% of frames imaged whole.
+Their depth is not identifiable, so no reconstruction can fix them.</figcaption></figure>
 
 <p>Excluding them is what the headline row buys: on GelSight Mini, CNC presses, 0-20 N
 calibration-free scores ρ&nbsp;0.996
@@ -84,15 +83,14 @@ evaluation does not exist</b>: no dataset here contains that many fully imaged
 presses.</p>
 
 <h3>All frames</h3>
-<p class="dim">The same protocol without that exclusion, for reference.</p>
+<p class="dim">The same protocol without that exclusion.</p>
 <div class="tablewrap"><table><thead><tr><th>dataset</th><th>n</th><th>LUT</th><th>calibration-free</th><th>shuffle floor</th><th>groups fitted</th></tr></thead><tbody>
 <tr><td>GelSight Mini, CNC presses, 0-20 N</td><td>2786</td><td>0.660</td><td><b>0.863</b></td><td class='dim'>+0.026 / +0.019</td><td class='dim'>6/6</td></tr><tr><td>FoTa cnc_Mini</td><td>2499</td><td>0.401</td><td><b>0.491</b></td><td class='dim'>+0.060 / +0.047</td><td class='dim'>6/6</td></tr><tr><td>FEATS (MARKER gel)</td><td>390</td><td><b>0.775</b></td><td>0.685</td><td class='dim'>-0.003 / +0.067</td><td class='dim'>4/6</td></tr><tr><td>Sparsh / Meta, 10 gel pads</td><td>960</td><td>0.456</td><td><b>0.588</b></td><td class='dim'>+0.092 / +0.099</td><td class='dim'>8/8</td></tr><tr><td>FeelAnyForce</td><td>5410</td><td>0.850</td><td><b>0.862</b></td><td class='dim'>+0.640 / +0.636</td><td class='dim'>42/42</td></tr></tbody></table></div>
 
 <p class="dim">The shuffle floor is an absolute ρ — what this protocol scores
 when the labels are permuted inside each group. Subtract it before reading a
-cell. React's own production number carries a fitted position gain field on top
-and is not in this table for that reason; it is on the
-<a href="method.html">method</a> page.</p>
+cell. React's production number adds a fitted position gain field and is
+therefore on the <a href="method.html">method</a> page instead.</p>
 
 <figure><img src="assets/pred_vs_gt.png" alt="predicted vs ground-truth force">
 <figcaption>Held-out prediction against ground truth, shared axes per row.
@@ -114,5 +112,9 @@ frames, mean difference 0.84&nbsp;N. <b>The published dataset still
 carries the LUT column</b>: switching it needs 36 episodes reprocessed.</p>
 
 <h2>Error analysis</h2>
+<p>The ten worst held-out frames of each dataset reconstruct as well as the
+five best — same gradient dipoles, same compact depth, no ramping. The residual
+is in the depth→force fit, not in image→depth, so a better reconstruction will
+not move these frames.</p>
 <figure><img src="assets/errors_cnc_mini_26.png" alt="cnc_mini_26 errors"><figcaption>GelSight Mini CNC — the ten worst held-out frames and, as a control, the five best. Relative error is |pred−true| over the dataset's force span (19.50&nbsp;N); median 1.1%, p90 3.7%, worst 11.6%.</figcaption></figure><figure><img src="assets/errors_cnc.png" alt="cnc errors"><figcaption>FoTa cnc_Mini — the ten worst held-out frames and, as a control, the five best. Relative error is |pred−true| over the dataset's force span (4.01&nbsp;N); median 3.8%, p90 10.9%, worst 48.3%.</figcaption></figure><figure><img src="assets/errors_feats.png" alt="feats errors"><figcaption>FEATS (marker) — the ten worst held-out frames and, as a control, the five best. Relative error is |pred−true| over the dataset's force span (34.11&nbsp;N); median 5.7%, p90 22.1%, worst 67.9%.</figcaption></figure><figure><img src="assets/errors_sparsh.png" alt="sparsh errors"><figcaption>Sparsh / Meta — the ten worst held-out frames and, as a control, the five best. Relative error is |pred−true| over the dataset's force span (0.98&nbsp;N); median 3.1%, p90 12.0%, worst 30.0%.</figcaption></figure><figure><img src="assets/errors_faf.png" alt="faf errors"><figcaption>FeelAnyForce — the ten worst held-out frames and, as a control, the five best. Relative error is |pred−true| over the dataset's force span (17.43&nbsp;N); median 1.6%, p90 12.7%, worst 48.1%.</figcaption></figure>
 </div></body></html>