Repository navigation
Expand file tree
/
Copy pathindex.html
More file actions
282 lines (250 loc) · 16.5 KB
/
Copy pathindex.html
File metadata and controls
282 lines (250 loc) · 16.5 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation</title>
<meta name="description" content="BEACON jointly learns a diffusion-based visuomotor policy and per-sample source weights, guided by the Best-Effort Adaptation bound, to transfer robot policies from abundant source demonstrations to a new domain with only a handful of target demonstrations.">
<meta property="og:title" content="BEACON: Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation">
<meta property="og:description" content="Learn which source demonstrations should supervise the target policy, and by how much. Sim-to-sim, sim-to-real and multi-source results on a Franka Panda.">
<meta property="og:image" content="assets/video/beacon-promo-poster.jpg">
<meta property="og:type" content="website">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Instrument+Sans:ital,wght@0,400;0,500;0,600;0,700;1,400&display=swap" rel="stylesheet">
<link rel="icon" href="assets/favicon.svg" type="image/svg+xml">
<link rel="stylesheet" href="style.css">
</head>
<body>
<a class="skip-link" href="#method">Skip to content</a>
<nav class="topbar" aria-label="Page">
<a class="topbar-brand" href="#top">BEACON</a>
<div class="topbar-links">
<a href="#method">How it works</a>
<a href="#rollouts">Rollouts</a>
<a href="#citation">Cite</a>
</div>
<div class="topbar-actions">
<a class="btn btn-small" href="https://arxiv.org/abs/2605.08571" target="_blank" rel="noopener">Paper</a>
<a class="btn btn-small btn-primary" href="https://github.com/zatpds/BEACON" target="_blank" rel="noopener">Code</a>
</div>
</nav>
<main id="top">
<!-- ───────────────────────── HERO ───────────────────────── -->
<header class="hero">
<a class="hero-lab" href="https://computationalrobotics.seas.harvard.edu/" target="_blank" rel="noopener">
<img src="assets/group-logo.png" alt="Computational Robotics Group, Harvard" width="240">
</a>
<h1 class="hero-title">
<span class="hero-acronym">BEACON</span>
Cross-Domain Co-Training of Generative Robot Policies via Best-Effort Adaptation
</h1>
<div class="hero-meta">
<p class="authors">
<span class="author">Antong Zhang<sup>1</sup></span>
<span class="author">Han Qi<sup>2</sup></span>
<span class="author">Heng Yang<sup>2</sup></span>
</p>
<p class="affiliations">
<span><sup>1</sup>Brown University</span>
<span><sup>2</sup>Harvard University</span>
</p>
</div>
<div class="hero-actions">
<a class="btn" href="https://arxiv.org/abs/2605.08571" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M6 3h8l5 5v13a1 1 0 0 1-1 1H6a1 1 0 0 1-1-1V4a1 1 0 0 1 1-1z"/><path d="M14 3v5h5M9 13h7M9 17h7"/></svg>
arXiv
</a>
<a class="btn" href="https://github.com/zatpds/BEACON" target="_blank" rel="noopener">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M9 19c-4.3 1.4-4.3-2.5-6-3m12 5v-3.5c0-1 .1-1.4-.5-2 2.8-.3 5.5-1.4 5.5-6a4.6 4.6 0 0 0-1.3-3.2 4.2 4.2 0 0 0-.1-3.2s-1.1-.3-3.5 1.3a12.3 12.3 0 0 0-6.2 0C6.5 2.8 5.4 3.1 5.4 3.1a4.2 4.2 0 0 0-.1 3.2A4.6 4.6 0 0 0 4 9.5c0 4.6 2.7 5.7 5.5 6-.6.6-.6 1.2-.5 2V21"/></svg>
Code
</a>
<a class="btn" href="#citation">
<svg viewBox="0 0 24 24" aria-hidden="true"><path d="M7 8h10M7 12h10M7 16h6"/><rect x="4" y="4" width="16" height="16" rx="2"/></svg>
BibTeX
</a>
</div>
<figure class="hero-video" id="video">
<video id="promo" autoplay muted loop controls playsinline preload="auto" poster="assets/video/beacon-promo-poster.jpg">
<source src="assets/video/beacon-promo.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</figure>
<p class="hero-lede">
BEACON co-trains a diffusion policy on hundreds of simulated demonstrations and a few
dozen from the target domain, and learns how much each source demonstration should count.
</p>
</header>
<!-- ───────────────────────── METHOD / SCROLLYTELLING ───────────────────────── -->
<section class="method" id="method" aria-labelledby="method-title">
<div class="method-head">
<h2 id="method-title">How it works</h2>
</div>
<div class="method-body">
<div class="method-steps">
<article class="step" data-task="stack" data-title="Block stacking">
<h3>Setup</h3>
<p>
500 simulated source demonstrations per task; 20–40 target demonstrations in a
shifted simulator or on the real robot. Fixed-ratio co-training counts every source
trajectory the same, whether or not it resembles the target.
</p>
</article>
<article class="step" data-task="mug" data-title="Mug cleanup">
<h3>Score each source sample</h3>
<p>
Every source window gets a discrepancy score <span class="math">d<sub>i</sub></span>:
the distance from its encoder embedding to its <span class="math">k</span> nearest
target embeddings. Near 1 means in-distribution; large means mismatched.
</p>
</article>
<article class="step" data-task="thread" data-title="Threading">
<h3>Learn weights and policy together</h3>
<p>
A surrogate of the Best-Effort Adaptation bound: <span class="math">q</span>-weighted
imitation loss, a discrepancy penalty, and a regularizer. Alternating updates of the
source weights and the diffusion policy; target samples always stay in.
</p>
</article>
<article class="step" data-task="ball" data-title="Ball rollover">
<h3>Several sources</h3>
<p>
MS-BEACON adds a domain-level weight on top of the per-sample weights, so simulators
closer to the target get more say.
</p>
</article>
<article class="step step-last" data-task="stack" data-title="Block stacking">
<h3>What BEACON achieves</h3>
<p>
Stronger sim-to-sim, sim-to-real and multi-source transfer than target-only,
fixed-ratio and feature-alignment baselines, and it holds up on out-of-distribution
initial states.
</p>
</article>
</div>
<div class="method-stage">
<div class="stage-sticky">
<canvas id="robot-canvas" aria-label="Animated Franka Panda arm performing the four manipulation tasks" role="img"></canvas>
<div class="stage-hud" aria-hidden="true">
<span class="hud-task" id="hud-task">Block stacking</span>
<span class="hud-progress"><span id="hud-bar"></span></span>
</div>
<noscript><img src="assets/video/beacon-promo-poster.jpg" alt="BEACON title card"></noscript>
</div>
</div>
</div>
</section>
<!-- ───────────────────────── ABSTRACT ───────────────────────── -->
<section class="abstract" id="abstract" aria-labelledby="abstract-title">
<h2 id="abstract-title">Abstract</h2>
<p>
We introduce <strong>BEACON</strong>, Best-Effort Adaptation for Cross-Domain Co-Training, a
theory-guided framework for training generative robot policies with abundant source
demonstrations and limited target demonstrations. BEACON casts cross-domain co-training
as a discrepancy-aware importance-reweighting problem, jointly learning a diffusion-based
visuomotor policy and per-sample source weights through an objective inspired by the
Best-Effort Adaptation (BEA) bound. To make this approach practical for high-dimensional
sequence policies, we develop scalable instance-level discrepancy proxies, stochastic
alternating updates for policy and weights, and a multi-source extension that balances
heterogeneous source domains. Across sim-to-sim, sim-to-real, and multi-source manipulation
settings, BEACON improves robustness and data efficiency over target-only, fixed-ratio
co-training, and feature-alignment baselines. Without an explicit alignment objective,
BEACON also brings source and target representations closer together as an implicit
result of discrepancy-aware co-training.
</p>
</section>
<!-- ───────────────────────── ROLLOUTS ───────────────────────── -->
<section class="rollouts" id="rollouts" aria-labelledby="rollouts-title">
<div class="section-head">
<h2 id="rollouts-title">BEACON rollouts</h2>
<p>
Real-robot trials are recorded by the robot’s camera. OOD trials start from object
positions outside the range of the target demonstrations.
</p>
</div>
<!-- Real robot -->
<div class="gallery" id="gallery-real">
<div class="gallery-head">
<h3>Real robot</h3>
<div class="segmented" role="tablist" aria-label="Real-robot task">
<button role="tab" aria-selected="true" data-group="mug">Mug cleanup</button>
<button role="tab" aria-selected="false" data-group="stack">Block stacking</button>
<button role="tab" aria-selected="false" data-group="ball">Ball rollover</button>
</div>
</div>
<div class="gallery-group is-active" data-group="mug" role="tabpanel">
<div class="tiles tiles-6">
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/mug/beacon.jpg" data-src="assets/video/real/mug/beacon.mp4"></video><figcaption>Continuous rollouts</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/mug/trial-1.jpg" data-src="assets/video/real/mug/trial-1.mp4"></video><figcaption>Trial 1</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/mug/trial-2.jpg" data-src="assets/video/real/mug/trial-2.mp4"></video><figcaption>Trial 2</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/mug/trial-3.jpg" data-src="assets/video/real/mug/trial-3.mp4"></video><figcaption>Trial 3</figcaption></figure>
<figure class="tile tile-ood"><video muted loop playsinline preload="none" poster="assets/video/real/mug/ood-1.jpg" data-src="assets/video/real/mug/ood-1.mp4"></video><figcaption>OOD trial 1</figcaption></figure>
<figure class="tile tile-ood"><video muted loop playsinline preload="none" poster="assets/video/real/mug/ood-2.jpg" data-src="assets/video/real/mug/ood-2.mp4"></video><figcaption>OOD trial 2</figcaption></figure>
</div>
</div>
<div class="gallery-group" data-group="stack" role="tabpanel" hidden>
<div class="tiles tiles-6">
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/stack/beacon.jpg" data-src="assets/video/real/stack/beacon.mp4"></video><figcaption>Continuous rollouts</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/stack/trial-1.jpg" data-src="assets/video/real/stack/trial-1.mp4"></video><figcaption>Trial 1</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/stack/trial-2.jpg" data-src="assets/video/real/stack/trial-2.mp4"></video><figcaption>Trial 2</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/stack/trial-3.jpg" data-src="assets/video/real/stack/trial-3.mp4"></video><figcaption>Trial 3</figcaption></figure>
<figure class="tile tile-ood"><video muted loop playsinline preload="none" poster="assets/video/real/stack/ood-1.jpg" data-src="assets/video/real/stack/ood-1.mp4"></video><figcaption>OOD trial 1</figcaption></figure>
<figure class="tile tile-ood"><video muted loop playsinline preload="none" poster="assets/video/real/stack/ood-2.jpg" data-src="assets/video/real/stack/ood-2.mp4"></video><figcaption>OOD trial 2</figcaption></figure>
</div>
</div>
<div class="gallery-group" data-group="ball" role="tabpanel" hidden>
<div class="tiles tiles-3">
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/ball/beacon.jpg" data-src="assets/video/real/ball/beacon.mp4"></video><figcaption>Continuous rollout</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/ball/trial-1.jpg" data-src="assets/video/real/ball/trial-1.mp4"></video><figcaption>Trial 1</figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/real/ball/trial-2.jpg" data-src="assets/video/real/ball/trial-2.mp4"></video><figcaption>Trial 2</figcaption></figure>
</div>
</div>
</div>
<!-- Simulation -->
<div class="gallery" id="gallery-sim">
<div class="gallery-head">
<h3>Simulation</h3>
<p class="gallery-note">Target-T: texture gap. Target-T+V: texture and viewpoint gap.</p>
</div>
<div class="tiles tiles-sim">
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/stack-t.jpg" data-src="assets/video/sim/stack-t.mp4"></video><figcaption>Block stacking <span>Target-T</span></figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/stack-tv.jpg" data-src="assets/video/sim/stack-tv.mp4"></video><figcaption>Block stacking <span>Target-T+V</span></figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/mug-t.jpg" data-src="assets/video/sim/mug-t.mp4"></video><figcaption>Mug cleanup <span>Target-T</span></figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/mug-tv.jpg" data-src="assets/video/sim/mug-tv.mp4"></video><figcaption>Mug cleanup <span>Target-T+V</span></figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/thread-t.jpg" data-src="assets/video/sim/thread-t.mp4"></video><figcaption>Threading <span>Target-T</span></figcaption></figure>
<figure class="tile"><video muted loop playsinline preload="none" poster="assets/video/sim/thread-tv.jpg" data-src="assets/video/sim/thread-tv.mp4"></video><figcaption>Threading <span>Target-T+V</span></figcaption></figure>
</div>
</div>
</section>
<!-- ───────────────────────── CITATION ───────────────────────── -->
<section class="citation" id="citation" aria-labelledby="citation-title">
<h2 id="citation-title">Citation</h2>
<p class="citation-venue">Published at NeurIPS 2026.</p>
<div class="bibtex-wrap">
<button type="button" class="bibtex-copy" data-copy-target="#bibtex">Copy BibTeX</button>
<pre id="bibtex"><code>@inproceedings{zhang2026beacon,
title = {{BEACON}: Cross-Domain Co-Training of Generative Robot Policies
via Best-Effort Adaptation},
author = {Zhang, Antong and Qi, Han and Yang, Heng},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
year = {2026},
url = {https://arxiv.org/abs/2605.08571}
}</code></pre>
</div>
</section>
</main>
<footer class="site-footer">
<div class="footer-inner">
<p>
Robot meshes from the robosuite Panda model (MIT). Built with
<a href="https://threejs.org" target="_blank" rel="noopener">three.js</a>; layout adapted from
<a href="https://github.com/gated-memory-policy/gated-memory-policy.github.io" target="_blank" rel="noopener">this template</a>
(CC BY-SA 4.0).
</p>
</div>
</footer>
<div id="toast" role="status" aria-live="polite"></div>
<script src="main.js" defer></script>
<script type="module" src="robot.js"></script>
</body>
</html>