-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
459 lines (419 loc) · 22.1 KB
/
Copy pathindex.html
File metadata and controls
459 lines (419 loc) · 22.1 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
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="title" content="Improving Robotic Imitation Learning via Trajectory Standardization">
<meta name="description"
content="Information-Standardized Trajectory Resampling (ISR) is an offline preprocessing method for improving robotic imitation learning from noisy and temporally irregular demonstrations.">
<meta name="keywords"
content="ISR, imitation learning, robot learning, trajectory resampling, trajectory standardization, manipulation">
<meta name="author"
content="Licheng Yang, Lingfeng Qian, Fei Zheng, Yonghao He, Wei Sui, Shuangshuang Li, Hu Su">
<meta property="og:type" content="article">
<meta property="og:site_name" content="D-Robotics AI Lab">
<meta property="og:url" content="https://d-robotics-ai-lab.github.io/isr.page/">
<meta property="og:title" content="Improving Robotic Imitation Learning via Trajectory Standardization">
<meta property="og:description"
content="ISR resamples robotic demonstration trajectories by enforcing approximately equal information distance between adjacent points.">
<meta property="og:image" content="https://d-robotics-ai-lab.github.io/isr.page/static/images/video-poster.jpg">
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:title" content="Improving Robotic Imitation Learning via Trajectory Standardization">
<meta name="twitter:description"
content="Information-Standardized Trajectory Resampling (ISR) improves robotic imitation learning from noisy and temporally irregular demonstrations.">
<meta name="twitter:image" content="https://d-robotics-ai-lab.github.io/isr.page/static/images/video-poster.jpg">
<meta name="citation_title" content="Improving Robotic Imitation Learning via Trajectory Standardization">
<meta name="citation_author" content="Yang, Licheng">
<meta name="citation_author" content="Qian, Lingfeng">
<meta name="citation_author" content="Zheng, Fei">
<meta name="citation_author" content="He, Yonghao">
<meta name="citation_author" content="Sui, Wei">
<meta name="citation_author" content="Li, Shuangshuang">
<meta name="citation_author" content="Su, Hu">
<meta name="citation_publication_date" content="2026">
<meta name="citation_conference_title" content="IROS 2026">
<meta name="citation_pdf_url" content="https://arxiv.org/pdf/2606.22907">
<meta name="citation_arxiv_id" content="2606.22907">
<meta name="theme-color" content="#2563eb">
<title>ISR: Improving Robotic Imitation Learning via Trajectory Standardization</title>
<link rel="canonical" href="https://d-robotics-ai-lab.github.io/isr.page/">
<link rel="sitemap" type="application/xml" title="Sitemap" href="sitemap.xml">
<link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
<link rel="apple-touch-icon" href="static/images/favicon.ico">
<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=Inter:wght@400;500;600;700;800&display=swap"
rel="stylesheet">
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
<link rel="stylesheet" href="static/css/index.css">
<link rel="stylesheet" href="static/css/isr.css?v=20260623-4">
<script defer src="static/js/fontawesome.all.min.js"></script>
<script defer src="static/js/isr-video.js?v=20260620-19"></script>
</head>
<body>
<section class="hero publication-header">
<div class="hero-body">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title isr-title">
<span>Improving Robotic Imitation Learning</span>
<span>via Trajectory Standardization</span>
</h1>
<div class="is-size-5 publication-authors isr-authors">
<span class="author-block">
<a onclick="return false">Licheng Yang</a><sup>1,2</sup>,
</span>
<span class="author-block">
<a onclick="return false">Lingfeng Qian</a><sup>2</sup>,
</span>
<span class="author-block">
<a onclick="return false">Fei Zheng</a><sup>2</sup>,
</span>
<span class="author-block">
<a onclick="return false">Yonghao He</a><sup>2†</sup>,
</span>
<span class="author-block">
<a onclick="return false">Wei Sui</a><sup>2</sup>,
</span>
<span class="author-block">
<a onclick="return false">Shuangshuang Li</a><sup>1</sup>,
</span>
<span class="author-block">
<a onclick="return false">Hu Su</a><sup>1*</sup>
</span>
</div>
<div class="is-size-5 publication-authors isr-affiliations">
<span class="author-block">
<sup>1</sup>State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of
Automation, Chinese Academy of Sciences <br>
<sup>2</sup>D-Robotics
</span>
</div>
<div class="isr-notes">
<sup>*</sup>Corresponding author
<sup>†</sup>Project lead
</div>
<div class="column has-text-centered">
<div class="publication-links">
<span class="link-block">
<a href="https://arxiv.org/abs/2606.22907" target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/pdf/2606.22907" target="_blank" rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>PDF</span>
</a>
</span>
<span class="link-block">
<a href="#video" class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-video"></i>
</span>
<span>Video</span>
</a>
</span>
<span class="link-block">
<a href="https://github.com/D-Robotics-AI-Lab/ISR" target="_blank"
rel="noopener"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column custom-width-95">
<h2 class="title is-3">Abstract</h2>
<div class="content has-text-justified">
<p>
Imitation learning for robotic manipulation relies on large sets of human demonstration trajectories,
which are often noisy and temporally irregular due to variable operator speed, intermittent pauses, and
inconsistent action density. A common preprocessing strategy is time-uniform downsampling to shorten
sequences, but it cannot effectively remove speed-induced non-uniformity or redundant pauses. This
mismatch degrades data quality and hinders policy learning. To address this issue, we propose
<strong>Information-Standardized Trajectory Resampling (ISR)</strong>, an offline preprocessing method
for effective imitation learning. ISR resamples each trajectory by enforcing approximately equal
information distance between adjacent points. Specifically, we map trajectories onto an
information-modulated Riemannian manifold and perform geodesic-equidistant parameterization. We construct
an information-intensity field from velocity and acceleration norms: the velocity term removes
small-motion redundancy, while the acceleration term preserves high-curvature and fine-manipulation
phases. We evaluate ISR on three real-world manipulation tasks with mainstream imitation learning
policies. Compared with the baseline time-uniform 3x downsampling, ISR improves task success rates by
about 25%, remains robust across datasets collected from different operators, and reduces both dataset
size and training cost.
</p>
</div>
<div class="isr-video-frame isr-video-after-abstract" id="video">
<video controls preload="auto" poster="static/images/video-poster.jpg?v=20260620-19" playsinline>
<source src="static/videos/ISR.mp4?v=20260620-19" type="video/mp4">
</video>
</div>
</div>
</div>
</div>
</section>
<section class="section isr-section" id="method">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column custom-width-95">
<h2 class="title is-3 has-text-centered">Method</h2>
<div class="content has-text-justified isr-copy">
<p>
<strong>Information-Standardized Trajectory Resampling (ISR)</strong> is an offline preprocessing method
that resamples raw teleoperated demonstration trajectories so that consecutive retained points carry
approximately equal kinematic and dynamic information. ISR comprises three steps, as illustrated in the
figure below.
</p>
<p>
<strong>Step 1 — Information Intensity Field.</strong> From the end-effector position sequence, ISR
computes per-frame velocity and acceleration via finite differences, then combines them into a scalar
information-intensity field weighted by two complementary terms. The velocity term captures the
instantaneous displacement rate and identifies low-value segments such as unintentional pauses and slow
transits. The acceleration term encodes curvature-induced centripetal acceleration during arc-like
movements and tangential deceleration before contact-rich manipulation phases, serving as a kinematic
proxy for force-sensitive operator intent.
</p>
<p>
<strong>Step 2 — Riemannian Manifold Mapping.</strong> Using the information-intensity field as a
conformal factor, ISR warps each trajectory onto a one-dimensional Riemannian manifold. Under this
metric, high-activity regions are stretched while low-activity regions are compressed, so that equal
geodesic arc lengths correspond to equal amounts of kinematic–dynamic information. The geodesic distance
between two trajectory points is approximated in discrete form using the secant distance for the velocity
term and the accumulated acceleration sum for the acceleration term.
</p>
<p>
<strong>Step 3 — Parametric Standardization.</strong> ISR selects the resampled index set by minimizing
the total squared deviation of each segment's geodesic distance from a prescribed target distance. This
geodesic-equidistant optimization is solved exactly via dynamic programming with prefix-sum acceleration,
producing a compact, standardized trajectory that serves as a drop-in replacement for any downstream
imitation learning pipeline without architectural changes.
</p>
</div>
<figure class="isr-figure isr-pipeline-figure">
<img src="static/images/isr_pipeline.png?v=20260623-4" alt="ISR pipeline overview" loading="lazy">
</figure>
</div>
</div>
</div>
</section>
<section class="section hero is-light isr-section" id="effectiveness">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column custom-width-95">
<h2 class="title is-3 has-text-centered">Effectiveness</h2>
<div class="content has-text-justified isr-copy">
<p>
ISR redistributes samples from low-value motion to task-critical phases. Its velocity term compresses
pauses, hesitation, and slow low-displacement transit into larger, more uniform steps, reducing
operator-speed artifacts and redundant frames; its acceleration term keeps denser coverage around sharp
turns, deceleration-to-contact, and fine manipulation, preventing compact but essential actions such as
grasping, alignment, and placement from being skipped. This information-balanced resampling improves
downstream success while retaining only about 28-40% of the original action points and remaining more
robust when demonstrations come from multiple operators.
</p>
</div>
<figure class="isr-figure isr-effect-figure">
<img src="static/images/isr_effect.png" alt="Visualization of ISR resampling effects" loading="lazy">
</figure>
</div>
</div>
</div>
</section>
<section class="section isr-section" id="evaluation">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column custom-width-95">
<h2 class="title is-3 has-text-centered">Evaluation</h2>
<div class="content has-text-justified isr-copy">
<p>
We evaluate ISR on three real-world manipulation tasks: Place&Cover, Place&Stack, and Push-T.
The evaluation contains three studies. <strong>Baseline Comparison</strong> compares ISR with
time-uniform 3x downsampling under two imitation learning policies, π<sub>0.5</sub> and VO-DP.
<strong>Ablation Study</strong> varies the acceleration weight while keeping the velocity term fixed to
analyze how dynamic-information preservation affects downstream policy learning.
<strong>Cross-Operator Robustness</strong> tests whether ISR can absorb operator-specific differences in
speed, pauses, and teleoperation style when demonstrations are collected from multiple people.
</p>
</div>
<div class="content has-text-justified isr-copy isr-eval-summary">
<p>
Across the three studies, ISR consistently improves downstream policy learning by redistributing samples
according to task-relevant motion information rather than timestamp spacing. The baseline comparison shows
broad gains across tasks and policy backbones; the ablation study shows that the acceleration term is
essential for contact-sensitive manipulation; and the cross-operator study shows that information
standardization reduces style-induced conflicts when demonstrations are collected from multiple operators.
</p>
</div>
<div class="isr-table-scroll" aria-label="Baseline comparison results">
<div class="isr-panel-title">Baseline Comparison</div>
<div class="isr-table-overflow">
<table class="isr-result-table">
<thead>
<tr>
<th>Method</th>
<th>Place&Cover</th>
<th>Place&Stack</th>
<th>Push-T</th>
<th>AVG. (↑)</th>
</tr>
</thead>
<tbody>
<tr>
<td>3x + π<sub>0.5</sub></td>
<td>45.8 <span>(33.3)</span></td>
<td>48.6 <span>(33.3)</span></td>
<td>48.9 <span>(33.3)</span></td>
<td>47.8</td>
</tr>
<tr class="isr-highlight-row">
<td>ISR + π<sub>0.5</sub></td>
<td>72.2 <span>(28.9)</span></td>
<td>72.2 <span>(28.2)</span></td>
<td>71.1 <span>(40.2)</span></td>
<td>71.8</td>
</tr>
<tr>
<td>3x + VO-DP</td>
<td>56.9 <span>(33.3)</span></td>
<td>66.7 <span>(33.3)</span></td>
<td>64.4 <span>(33.3)</span></td>
<td>62.7</td>
</tr>
<tr class="isr-highlight-row">
<td>ISR + VO-DP</td>
<td>77.8 <span>(28.9)</span></td>
<td>91.6 <span>(28.2)</span></td>
<td>71.1 <span>(40.2)</span></td>
<td>80.2</td>
</tr>
</tbody>
</table>
</div>
</div>
<p class="isr-table-note">Success rate is shown in percent; resampling ratio is shown in parentheses.</p>
<div class="isr-eval-grid">
<figure class="isr-figure isr-eval-figure">
<figcaption class="isr-panel-title">Ablation Study (Acceleration Weight in Parentheses)</figcaption>
<img src="static/images/isr_ablation_cover.png" alt="Acceleration weight ablation on Place and Cover"
loading="lazy">
<img src="static/images/isr_ablation_stack.png" alt="Acceleration weight ablation on Place and Stack"
loading="lazy">
</figure>
<figure class="isr-figure isr-eval-figure">
<figcaption class="isr-panel-title">Cross-Operator Robustness</figcaption>
<img src="static/images/isr_cross_operator_single.png" alt="Single-operator Place and Stack comparison"
loading="lazy">
<img src="static/images/isr_cross_operator_mixed.png" alt="Mixed-operator Place and Stack comparison"
loading="lazy">
</figure>
</div>
</div>
</div>
</div>
</section>
<section class="section hero is-light isr-section" id="limitations">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column custom-width-95">
<h2 class="title is-3 has-text-centered">Limitations</h2>
<div class="content has-text-justified isr-copy">
<p>
ISR currently operates on end-effector position trajectories. Extending the information-intensity field to
joint-space representations could broaden its applicability to robots and tasks where joint-level motion
carries important information beyond the end-effector path. In addition, the acceleration weight
λ<sub>acc</sub> is manually selected for each task, since different manipulation behaviors can have
different sensitivity to acceleration-level features. An adaptive mechanism that infers this weight from
trajectory statistics would make ISR more general and easier to deploy across new tasks.
</p>
</div>
</div>
</div>
</div>
</section>
<section class="section isr-section" id="BibTeX">
<div class="container is-max-desktop">
<div class="isr-bibtex-block">
<div class="bibtex-header">
<h2 class="title">BibTeX</h2>
<button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
<i class="fas fa-copy"></i>
<span class="copy-text">Copy</span>
</button>
</div>
<pre id="bibtex-code"><code>@inproceedings{yang2026isr,
title={Improving Robotic Imitation Learning via Trajectory Standardization},
author={Licheng Yang and Lingfeng Qian and Fei Zheng and Yonghao He and Wei Sui and Shuangshuang Li and Hu Su},
booktitle={arXiv:2606.22907},
year={2026}
}</code></pre>
</div>
</div>
</section>
<footer class="footer isr-footer">
<div class="container is-max-desktop">
<div class="content has-text-centered">
<p>
This page was built using the
<a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank"
rel="noopener">Academic Project Page Template</a>.
</p>
</div>
</div>
</footer>
<script>
function copyBibTeX() {
const code = document.getElementById('bibtex-code').innerText;
const button = document.querySelector('.copy-bibtex-btn');
const text = button.querySelector('.copy-text');
const originalText = text.textContent;
const setCopied = () => {
button.classList.add('copied');
text.textContent = 'Copied';
window.setTimeout(() => {
button.classList.remove('copied');
text.textContent = originalText;
}, 1600);
};
const fallbackCopy = () => {
const textarea = document.createElement('textarea');
textarea.value = code;
textarea.setAttribute('readonly', '');
textarea.style.position = 'fixed';
textarea.style.opacity = '0';
document.body.appendChild(textarea);
textarea.select();
document.execCommand('copy');
document.body.removeChild(textarea);
setCopied();
};
if (navigator.clipboard && window.isSecureContext) {
navigator.clipboard.writeText(code).then(setCopied).catch(fallbackCopy);
} else {
fallbackCopy();
}
}
</script>
</body>
</html>