Skip to content

Commit b245bd4

Browse files
FaizalKarim280280FaizalKarim280280
committed
Added table
1 parent 4d5cb2b commit b245bd4

1 file changed

Lines changed: 186 additions & 30 deletions

File tree

‎index.html‎

Lines changed: 186 additions & 30 deletions
Original file line numberDiff line numberDiff line change
@@ -216,7 +216,8 @@ <h2 class="title is-3 mt-3">Abstract</h2>
216216
<h2 class="title is-2">Model Architecture</h2>
217217
<img class="mt-4" src="./static/images/pipeline.svg">
218218
<div class="content has-text-justified my-4">
219-
<b>Overview of the proposed method</b>: <b>(a)</b> Given an object point cloud \(P\), our network
219+
<b>Overview of the proposed method</b>: <b>(a)</b> Given an object point cloud \(P\), our
220+
network
220221
encodes
221222
geometric features into dense feature maps. Next,
222223
randomly initialized dual-arm grasps \(H\) are used to transform a fixed query cloud into query
@@ -246,7 +247,8 @@ <h4 class="title is-4 has-text-centered">\(SE(3) \times SE(3) \longleftrightarro
246247
</div>
247248

248249
<div class="column is-full-width">
249-
<b>Denoising in the dual-arm grasp space:</b> Additionally, dual-arm grasp poses are represented as pairs of rigid-body transformations
250+
<b>Denoising in the dual-arm grasp space:</b> Additionally, dual-arm grasp poses are represented as
251+
pairs of rigid-body transformations
250252
in \(SE(3) \times SE(3)\), which are mapped into a \(12\text{D}\) Euclidean space for diffusion and
251253
back.
252254
Each \(SE(3)\) element is
@@ -267,54 +269,56 @@ <h4 class="title is-4 has-text-centered">\(SE(3) \times SE(3) \longleftrightarro
267269
<h4 class="title is-4 has-text-centered">Denoising using Classifier Guidance</h4>
268270

269271
<div class="columns has-text-centered">
270-
<div class="column is-full-width mt-2">
271-
<video autoplay loop muted poster="" preload="none" style="width:100%;">
272-
<source src="./static/videos/only_graph_cropped3.mp4">
273-
</video>
274-
</div>
272+
<div class="column is-full-width mt-2">
273+
<video autoplay loop muted poster="" preload="none" style="width:100%;">
274+
<source src="./static/videos/only_graph_cropped3.mp4">
275+
</video>
276+
</div>
275277
</div>
276-
278+
277279
<!-- Colormap bar -->
278280
<div class="columns has-text-centered my-5">
279-
<div class="column is-full-width">
280-
<div style="
281+
<div class="column is-full-width">
282+
<div style="
281283
background: linear-gradient(to right, rgb(255, 85, 85), rgb(63, 255, 63));
282284
height: 12px;
283285
border-radius: 30px;
284286
margin: 0 auto;
285287
width: 70%;
286288
position: relative;">
289+
</div>
290+
<div
291+
style="display: flex; justify-content: space-between; width: 70%; margin: 5px auto 0 auto; font-size: 0.9rem;">
292+
<span style="color: rgb(182, 1, 1); font-weight: 500;">Noisy Grasp Pairs</span>
293+
<span style="color: rgb(45, 150, 45); font-weight: 500;">Stable Grasp Pairs</span>
294+
</div>
287295
</div>
288-
<div style="display: flex; justify-content: space-between; width: 70%; margin: 5px auto 0 auto; font-size: 0.9rem;">
289-
<span style="color: rgb(182, 1, 1); font-weight: 500;">Noisy Grasp Pairs</span>
290-
<span style="color: rgb(45, 150, 45); font-weight: 500;">Stable Grasp Pairs</span>
291-
</div>
292-
</div>
293296
</div>
294-
297+
295298
<div class="content has-text-justified my-4">
296-
<b>Overview of the denoising process:</b> The above clip shows the joint denoising process step by step. As the
297-
time progresses, the <span style="color: rgb(182, 1, 1);">Energy \((E_\alpha)\)</span> gradually
298-
decreases, which means grasps are moving towards the object and
299-
not just floating in free space. At the same time, the <span
300-
style="color:rgb(11, 33, 158)">Force-Closure Probability \((C_{\beta}^{\text{fc}})\)</span> steadily
301-
increases,
302-
highlighting how the grasp becomes more stable and reliable over time. Finally, in the later stages of
303-
denoising, colliding grasps are
304-
refined for a small number of iterations using <span style="color:rgb(45, 150, 45)">Collision Classifier
305-
\((C_{\gamma}^{\text{col}})\)</span>, resulting in dual-arm grasps that are force-closure stable as
306-
well as collision-free.
299+
<b>Overview of the denoising process:</b> The above clip shows the joint denoising process step by step.
300+
As the
301+
time progresses, the <span style="color: rgb(182, 1, 1);">Energy \((E_\alpha)\)</span> gradually
302+
decreases, which means grasps are moving towards the object and
303+
not just floating in free space. At the same time, the <span
304+
style="color:rgb(11, 33, 158)">Force-Closure Probability \((C_{\beta}^{\text{fc}})\)</span> steadily
305+
increases,
306+
highlighting how the grasp becomes more stable and reliable over time. Finally, in the later stages of
307+
denoising, colliding grasps are
308+
refined for a small number of iterations using <span style="color:rgb(45, 150, 45)">Collision Classifier
309+
\((C_{\gamma}^{\text{col}})\)</span>, resulting in dual-arm grasps that are force-closure stable as
310+
well as collision-free.
307311
</div>
308-
309-
312+
313+
310314
</div>
311315
</section>
312316

313317
<section class="section" style="background-color: rgb(252, 252, 252);">
314318
<div class="container is-max-desktop">
315319
<div class="columns has-text-centered">
316320
<div class="column is-full-width">
317-
<h2 class="title is-3">
321+
<h2 class="title is-2">
318322
Real Life Results <sup style="font-size: 15px;">&dagger;</sup>
319323
</h2>
320324

@@ -370,6 +374,158 @@ <h4 class="title is-5">(e) Saucepan</h4>
370374
</div>
371375
</section>
372376

377+
<section class="section" style="background-color:#fff;">
378+
<div class="container is-max-desktop">
379+
<h2 class="title is-2 has-text-centered">Quantitative Results</h2>
380+
<p class="has-text-centered is-size-6 mb-4">
381+
Comparison on our evaluation set (<span class="icon"><i class="fas fa-arrow-up"></i></span>higher is
382+
better, <span class="icon"><i class="fas fa-arrow-down"></i></span> lower is better).
383+
</p>
384+
385+
<h3 class="title is-5 mt-5 pt-5">1. Comparison with Baselines</h3>
386+
387+
<!-- Main comparison -->
388+
<div class="table-container">
389+
<table class="table is-striped is-hoverable is-fullwidth">
390+
<thead>
391+
<tr>
392+
<th>Method</th>
393+
<th class="has-text-right">FCE&nbsp;(%) <span class="icon"><i
394+
class="fas fas-solid fa-arrow-up"></i></span></th>
395+
<th class="has-text-right">GSR&nbsp;(%) <span class="icon"><i
396+
class="fas fa-arrow-up"></i></span></th>
397+
<th class="has-text-right">GCR&nbsp;(%) <span class="icon"><i
398+
class="fas fa-arrow-down"></i></span></th>
399+
</tr>
400+
</thead>
401+
<tbody>
402+
<tr class="">
403+
<td><b>DAGDiff (ours)</b></td>
404+
<td class="has-text-right"><b>60.14</b></td>
405+
<td class="has-text-right"><b>72.50</b></td>
406+
<td class="has-text-right"><b>15.10</b></td>
407+
</tr>
408+
<tr>
409+
<td>CGDF</td>
410+
<td class="has-text-right">35.14</td>
411+
<td class="has-text-right">56.25</td>
412+
<td class="has-text-right">30.55</td>
413+
</tr>
414+
<tr>
415+
<td>VCGS</td>
416+
<td class="has-text-right">16.85</td>
417+
<td class="has-text-right">23.36</td>
418+
<td class="has-text-right">74.73</td>
419+
</tr>
420+
<tr>
421+
<td>UniDiffGrasp</td>
422+
<td class="has-text-right">10.10</td>
423+
<td class="has-text-right">31.68</td>
424+
<td class="has-text-right">59.90</td>
425+
</tr>
426+
<tr>
427+
<td>RoboBrainGrasp-KP</td>
428+
<td class="has-text-right">9.80</td>
429+
<td class="has-text-right">27.85</td>
430+
<td class="has-text-right">66.30</td>
431+
</tr>
432+
<tr>
433+
<td>RoboBrainGrasp-BB</td>
434+
<td class="has-text-right">7.12</td>
435+
<td class="has-text-right">27.81</td>
436+
<td class="has-text-right">70.26</td>
437+
</tr>
438+
</tbody>
439+
</table>
440+
</div>
441+
442+
<!-- Dual-Afford (zero-shot) block -->
443+
<h3 class="title is-5 mt-5 pt-5">2. Zero-Shot on Dual-Afford Objects<sup>†</sup></h3>
444+
<div class="table-container">
445+
<table class="table is-narrow is-striped is-fullwidth is-hoverable">
446+
<thead>
447+
<tr>
448+
<th>Method</th>
449+
<th class="has-text-right">FCE&nbsp;(%) <span class="icon is-small"><i
450+
class="fas fa-arrow-up"></i></span></th>
451+
<th class="has-text-right">GSR&nbsp;(%) <span class="icon is-small"><i
452+
class="fas fa-arrow-up"></i></span></th>
453+
<th class="has-text-right">GCR&nbsp;(%) <span class="icon is-small"><i
454+
class="fas fa-arrow-down"></i></span></th>
455+
</tr>
456+
</thead>
457+
<tbody>
458+
<tr>
459+
<td><b>Ours-DA</b><sup>†</sup></td>
460+
<td class="has-text-right"><b>56.45</b></td>
461+
<td class="has-text-right"><b>68.80</b></td>
462+
<td class="has-text-right"><b>18.59</b></td>
463+
</tr>
464+
<tr>
465+
<td>Dual-Afford<sup>††</sup></td>
466+
<td class="has-text-right">–</td>
467+
<td class="has-text-right">54.33</td>
468+
<td class="has-text-right">–</td>
469+
</tr>
470+
</tbody>
471+
</table>
472+
</div>
473+
474+
475+
<p class="is-size-7 has-text-grey mt-2">
476+
<sup>†</sup> Evaluated on Dual-Afford objects in a zero-shot setting. <br>
477+
<sup>††</sup> Values reported directly from the Dual-Afford paper.
478+
</p>
479+
480+
481+
<h3 class="title is-5 mt-5 pt-5">3. Real-World Dual-Arm Grasp Results</h3>
482+
483+
<div class="table-container">
484+
<table class="table is-striped is-hoverable is-fullwidth has-text-centered">
485+
<thead>
486+
<tr>
487+
<th>Object</th>
488+
<th>Tray</th>
489+
<th>Bucket</th>
490+
<th>Saucepan</th>
491+
<th>Frypan</th>
492+
<th>Drone</th>
493+
</tr>
494+
</thead>
495+
<tbody>
496+
<tr>
497+
<th>Success</th>
498+
<td>6/10</td>
499+
<td>8/10</td>
500+
<td>7/10</td>
501+
<td>6/10</td>
502+
<td>5/10</td>
503+
</tr>
504+
</tbody>
505+
</table>
506+
</div>
507+
508+
<div class="content has-text-justified my-4">
509+
<b>Quantitative Results:</b> DAGDiff consistently outperforms across all evaluation settings. It outperforms prior
510+
methods in <u>Force-Closure Evaluation (FCE)</u> and <u>Grasp Success Rate (GSR)</u> while maintaining the lowest <u>Grasp
511+
Collision Rate (GCR)</u>, indicating more physically valid and robust dual-arm grasps. In zero-shot transfer
512+
to Dual-Afford objects, DAGDiff continues to show strong generalization without task-specific
513+
retraining. Finally, real-world experiments on unseen objects such as trays, buckets, and pans
514+
demonstrate consistent success, confirming that DAGDiff’s classifier-guided diffusion produces grasps
515+
that are stable, collision-free, and transferable beyond simulation. Real-life failures occur mostly due to noisy point-cloud estimation and
516+
hance generated grasps are not always perfect.
517+
</div>
518+
</div>
519+
</section>
520+
521+
522+
<section class="section" id="BibTeX" style="margin-bottom: 1rem;">
523+
<div class="container is-max-desktop content">
524+
<h2 class="title">BibTeX</h2>
525+
<pre><code>Will be updated</code></pre>
526+
</div>
527+
</section>
528+
373529
<!-- <section class="section" id="BibTeX" style="margin-bottom: 1rem;">
374530
<div class="container is-max-desktop content">
375531
<h2 class="title">BibTeX</h2>

0 commit comments

Comments
 (0)