Skip to content

Commit 1a73862

Browse files
DAG-DiffDAG-Diff
authored
Update index.html
1 parent afff94b commit 1a73862

1 file changed

Lines changed: 114 additions & 6 deletions

File tree

‎index.html‎

Lines changed: 114 additions & 6 deletions
Original file line numberDiff line numberDiff line change
@@ -43,7 +43,8 @@
4343
<div class="container is-max-desktop">
4444
<div class="columns is-centered">
4545
<div class="column has-text-centered">
46-
<h1 class="title is-1 publication-title"><span class="gradient-text">DAGDiff</span>: Guiding Dual-Arm Grasp Diffusion to Stable and Collision-Free Grasps</h1>
46+
<h1 class="title is-1 publication-title"><span class="gradient-text">DAGDiff</span>: Guiding
47+
Dual-Arm Grasp Diffusion to Stable and Collision-Free Grasps</h1>
4748

4849
</div>
4950
</div>
@@ -52,29 +53,136 @@ <h1 class="title is-1 publication-title"><span class="gradient-text">DAGDiff</sp
5253
</section>
5354

5455

55-
<section class="hero is-light px-5 pb-3" style="background-color: rgb(255, 255, 255);">
56+
<!-- <section class="hero is-light px-5 pb-3" style="background-color: rgb(255, 255, 255);">
5657
<div class="container is-max-desktop">
5758
<div class="columns is-centered video-container">
58-
<video autoplay muted loop playsinline style="border-radius: 15px">
59-
<source src="./static/videos/intro_final0001-0500.mkv" type="video/mp4">
59+
<video autoplay muted style="border-radius: 15px">
60+
<source src="./static/videos/intro_video.mp4" type="video/mp4">
6061
</video>
6162
</div>
63+
</section> -->
64+
65+
<section class="section" style="background-color: rgb(252, 252, 252);">
66+
<div class="container is-max-desktop">
67+
<div class="columns has-text-centered">
68+
<div class="column is-full-width">
69+
<!-- <h2 class="title is-3">Model Architecture</h2> -->
70+
<img src="./static/images/teaser_icra2025.svg" width="70%">
71+
<!-- <div class="content has-text-justified my-4">
72+
<b>Overview of the proposed method</b>: <b>(a)</b> Given an object point cloud P , our network
73+
encodes
74+
geometric features into dense feature maps. Next,
75+
randomly initialized dual-arm grasps \(H\) are used to transform a fixed query cloud into query
76+
points, followed by feature sampling through bilinear interpolation. Conditioned on the noise
77+
step \(t\)
78+
these features are passed through \(F_{\theta}\), which predicts the SDF of the query points and
79+
a feature vector.
80+
This vector is then used by three output heads that predict energy \((E_\alpha)\), force-closure
81+
probability \((C_{\beta}^{\text{fc}})\) and
82+
collision probability \((C_{\gamma}^{\text{col}})\), jointly guiding the diffusion process. <b>(b)</b> At
83+
inference, denoising proceeds from random
84+
intialization \((t=250\)) to refine grasps \((t=0\)). The energy head drives the generative
85+
dynamics, while the force-closure and collision heads bias the generation until stable,
86+
collision-free dual-arm grasps emerge.
87+
88+
</div> -->
89+
</div>
90+
</div>
91+
</div>
6292
</section>
6393

64-
<section class="section mt-3 mb-0" style="background-color: rgb(252, 252, 252);">
94+
95+
<section class="section mt-3 mb-0" style="background-color: rgb(255, 255, 255);">
6596
<div class="container is-max-desktop">
6697
<div class="columns is-centered has-text-centered">
6798
<div class="column is-four-fifths">
6899
<h2 class="title is-3 mt-3">Abstract</h2>
69100
<div class="content has-text-justified">
70-
Reliable dual-arm grasping is essential for manipulating large and complex objects but remains a challenging problem due to stability, collision, and generalization requirements. Prior methods typically decompose the task into two independent grasp proposals, relying on region priors or heuristics that limit generalization and provide no principled guarantee of stability. We propose DAGDiff, an end-to-end framework that directly denoises to grasp pairs in the $SE(3) \times SE(3)$ space. Our key insight is that stability and collision can be enforced more effectively by guiding the diffusion process with classifier signals, rather than relying on explicit region detection or object priors. To this end, DAGDiff integrates geometry-, stability-, and collision-aware guidance terms that steer the generative process toward grasps that are physically valid and force-closure compliant. We comprehensively evaluate DAGDiff through analytical force-closure checks, collision analysis, and large-scale physics-based simulations, showing consistent improvements over previous work on these metrics. Finally, we demonstrate that our framework generates dual-arm grasps directly from real-world point clouds of previously unseen objects, which are executed on a heterogeneous dual-arm setup where two manipulators reliably grasp and lift them.
101+
Reliable dual-arm grasping is essential for manipulating large and complex objects but remains a
102+
challenging problem due to stability, collision, and generalization requirements. Prior methods
103+
typically decompose the task into two independent grasp proposals, relying on region priors or
104+
heuristics that limit generalization and provide no principled guarantee of stability. We
105+
propose DAGDiff, an end-to-end framework that directly denoises to grasp pairs in the $SE(3)
106+
\times SE(3)$ space. Our key insight is that stability and collision can be enforced more
107+
effectively by guiding the diffusion process with classifier signals, rather than relying on
108+
explicit region detection or object priors. To this end, DAGDiff integrates geometry-,
109+
stability-, and collision-aware guidance terms that steer the generative process toward grasps
110+
that are physically valid and force-closure compliant. We comprehensively evaluate DAGDiff
111+
through analytical force-closure checks, collision analysis, and large-scale physics-based
112+
simulations, showing consistent improvements over previous work on these metrics. Finally, we
113+
demonstrate that our framework generates dual-arm grasps directly from real-world point clouds
114+
of previously unseen objects, which are executed on a heterogeneous dual-arm setup where two
115+
manipulators reliably grasp and lift them.
71116
</div>
72117
</div>
73118
</div>
74119
</div>
75120
</div>
76121
</section>
77122

123+
<!-- <section class="hero is-light px-5 pb-3" style="background-color: rgb(255, 255, 255);">
124+
<div class="container is-max-desktop">
125+
<h2 class="title is-3 mt-3">Abstract</h2>
126+
<div class="columns is-centered video-container">
127+
<video autoplay muted style="border-radius: 15px">
128+
<source src="./static/videos/intro_video.mp4" type="video/mp4">
129+
</video>
130+
</div>
131+
</section> -->
132+
133+
<section class="section" style="background-color: rgb(252, 252, 252);">
134+
<div class="container is-max-desktop">
135+
<div class="columns has-text-centered">
136+
<div class="column is-full-width">
137+
<h2 class="title is-3">Video Explanation</h2>
138+
<div class="columns is-centered video-container">
139+
<video controls muted poster="./static/images/video_thumbnail.png" preload="none"
140+
src="./static/videos/intro_video.mp4">
141+
</video>
142+
</div>
143+
</div>
144+
</div>
145+
</div>
146+
</section>
147+
148+
<section class="section" style="background-color: rgb(255, 255, 255);">
149+
<div class="container is-max-desktop">
150+
<div class="columns has-text-centered">
151+
<div class="column is-full-width">
152+
<h2 class="title is-3">Model Architecture</h2>
153+
<img src="./static/images/pipeline.svg">
154+
<div class="content has-text-justified my-4">
155+
<b>Overview of the proposed method</b>: <b>(a)</b> Given an object point cloud P , our network
156+
encodes
157+
geometric features into dense feature maps. Next,
158+
randomly initialized dual-arm grasps \(H\) are used to transform a fixed query cloud into query
159+
points, followed by feature sampling through bilinear interpolation. Conditioned on the noise
160+
step \(t\)
161+
these features are passed through \(F_{\theta}\), which predicts the SDF of the query points and
162+
a feature vector.
163+
This vector is then used by three output heads that predict energy \((E_\alpha)\), force-closure
164+
probability \((C_{\beta}^{\text{fc}})\) and
165+
collision probability \((C_{\gamma}^{\text{col}})\), jointly guiding the diffusion process. <b>(b)</b> At
166+
inference, denoising proceeds from random
167+
intialization \((t=250\)) to refine grasps \((t=0\)). The energy head drives the generative
168+
dynamics, while the force-closure and collision heads bias the generation until stable,
169+
collision-free dual-arm grasps emerge.
170+
171+
</div>
172+
</div>
173+
</div>
174+
</div>
175+
</section>
176+
177+
<section class="section" style="background-color: rgb(252, 252, 252);">
178+
<div class="container is-max-desktop">
179+
<div class="columns has-text-centered">
180+
<div class="column is-full-width">
181+
<h2 class="title is-3">Results</h2>
182+
</div>
183+
</div>
184+
</div>
185+
</section>
78186

79187

80188
<footer class="footer">

0 commit comments

Comments
 (0)