1<!DOCTYPE html> 2<html lang="en"> 3 <!-- Head --> 4 <head> 5<meta http-equiv="Content-Type" content="text/html; charset=UTF-8"> 6 7 <!-- Metadata, OpenGraph and Schema.org --> 8 9 10 11 12<!-- Standard metadata --> 13<meta charset="utf-8"> 14<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no"> 15<meta http-equiv="X-UA-Compatible" content="IE=edge"> 16<title> 17 18 19 20 Publications | Yuhong Cao 21 22 23</title> 24<meta name="author" content="Yuhong Cao"> 25<meta name="description" content="I hope you enjoy reading my publications!"> 26 27 <meta name="keywords" content="deep reinforcement learning, robot learning, view planning, path planning"> 28 29 30 31 32 33 34 35 36 37 38<!-- Bootstrap & MDB --> 39<link rel="stylesheet" href="/assets/css/bootstrap.min.css?a4b3f509e79c54a512b890d73235ef04"> 40<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/[email protected]/css/mdb.min.css" integrity="sha256-jpjYvU3G3N6nrrBwXJoVEYI/0zw8htfFnhT9ljN3JJw=" crossorigin="anonymous"> 41 42<!-- Bootstrap Table --> 43 44 45<!-- Fonts & Icons --> 46<link defer rel="stylesheet" href="/assets/css/academicons.min.css?f0b7046b84e425c55f3463ac249818f5"> 47<link defer rel="stylesheet" type="text/css" href="https://fonts.googleapis.com/css?family=Roboto:300,400,500,700|Roboto+Slab:100,300,400,500,700|Material+Icons&display=swap"> 48 49<!-- Code Syntax Highlighting --> 50<link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-github.css?591dab5a4e56573bf4ef7fd332894c99" media="" id="highlight_theme_light"> 51 52 53 54<!-- Styles --> 55 56<!-- pseudocode --> 57 58 59 60 <link rel="shortcut icon" href="/assets/img/logo.svg?7b8f72542d151aa3a3fff62f21cbd55d"> 61 62<link rel="stylesheet" href="/assets/css/main.css?d41d8cd98f00b204e9800998ecf8427e"> 63<link rel="canonical" href="https://www.yuhongcao.online/publications/"> 64 65<!-- Dark Mode -->
66<script src="/assets/js/theme.js?9a0c749ec5240d9cda97bc72359a72c0"></script>
66 67 68 <link defer rel="stylesheet" href="/assets/css/jekyll-pygments-themes-native.css?5847e5ed4a4568527aa6cfab446049ca" media="none" id="highlight_theme_dark"> 69
69<script> 70 initTheme(); 71 </script>
71 72 73 74<!-- GeoJSON support via Leaflet --> 75 76 77<!-- diff2html --> 78 79 80 81 82 83 84 </head> 85 86 <!-- Body --> 87 <body class="fixed-top-nav "> 88 <!-- Header --> 89 <header> 90 <!-- Nav Bar --> 91 <nav id="navbar" class="navbar navbar-light navbar-expand-sm fixed-top" role="navigation"> 92 <div class="container"> 93 94 <a class="navbar-brand title font-weight-lighter" href="/"> 95 Yuhong Cao 96 </a> 97 98 <!-- Navbar Toggle --> 99 <button class="navbar-toggler collapsed ml-auto" type="button" data-toggle="collapse" data-target="#navbarNav" aria-controls="navbarNav" aria-expanded="false" aria-label="Toggle navigation"> 100 <span class="sr-only">Toggle navigation</span> 101 <span class="icon-bar top-bar"></span> 102 <span class="icon-bar middle-bar"></span> 103 <span class="icon-bar bottom-bar"></span> 104 </button> 105 106 <div class="collapse navbar-collapse text-right" id="navbarNav"> 107 <ul class="navbar-nav ml-auto flex-nowrap"> 108 109 110 <!-- About --> 111 <li class="nav-item "> 112 <a class="nav-link" href="/">About 113 114 </a> 115 </li> 116 117 <!-- Other pages --> 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 <li class="nav-item active"> 147 148 <a class="nav-link" href="/publications/">Publications 149 150 <span class="sr-only">(current)</span> 151 152 </a> 153 </li> 154 155 156 157 158 159 160 161 162 163 <li class="nav-item "> 164 165 <a class="nav-link" href="/cv/">CV 166 167 </a> 168 </li> 169 170 171 172 173 174 175 176 177 178 179 180 <!-- Toogle theme mode --> 181 <li class="toggle-container"> 182 <button id="light-toggle" title="Change theme"> 183 <i class="ti ti-sun-moon" id="light-toggle-system"></i> 184 <i class="ti ti-moon-filled" id="light-toggle-dark"></i> 185 <i class="ti ti-sun-filled" id="light-toggle-light"></i> 186 </button> 187 </li> 188 189 </ul> 190 </div> 191 </div> 192 </nav> 193 194 <!-- Scrolling Progress Bar --> 195 <progress id="progress" value="0"> 196 <div class="progress-container"> 197 <span class="progress-bar"></span> 198 </div> 199 </progress> 200 201</header> 202 203 204 <!-- Content --> 205 <div class="container mt-5" role="main"> 206 207 <div class="post"> 208 <header class="post-header"> 209 <h1 class="post-title">Publications</h1> 210 <p class="post-description">I hope you enjoy reading my publications!</p> 211 </header> 212 213 <article> 214 <!-- _pages/publications.md --> 215<div class="publications"> 216 217<h2 class="bibliography">2026</h2> 218<ol class="bibliography"> 219<li> 220<div class="row"> 221 222 <div class="col col-sm-2 abbr"> 223 224 225 226 227 228<figure> 229 <picture> 230 <!-- Auto scaling with imagemagick --> 231 <!-- 232 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 233 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 234 --> 235 236 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/orion-480.webp 480w,/assets/img/publication_preview/orion-800.webp 800w,/assets/img/publication_preview/orion-1400.webp 1400w," sizes="200px" type="image/webp"></source> 237 238 <img src="/assets/img/publication_preview/orion.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="orion.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 239 </picture> 240 241 242</figure> 243 244 245 246 </div> 247 248 249 <!-- Entry bib key --> 250 <div id="zhang2026orion" class="col-sm-8"> 251 <!-- Title --> 252 <div class="title">ORION: Option-Regularized Deep Reinforcement Learning for Cooperative Multi-Agent Online Navigation</div> 253 <!-- Author --> 254 <div class="author"> 255 256 257 258 Shizhe 259 Zhang, Jingsong 260 Liang, Zhitao 261 Zhou, Shuhan 262 Ye, Yizhuo 263 Wang, Derek Ming Siang 264 Tan, Jimmy
265 Chiun, and 266 <span class="more-authors" title="click to view 2 more authors" onclick=" 267 var element = $(this); 268 element.attr('title', ''); 269 var more_authors_text = element.text() == '2 more authors' ? 'Yuhong Cao, Guillaume Sartoretti' : '2 more authors'; 270 var cursorPosition = 0; 271 var textAdder = setInterval(function(){ 272 element.text(more_authors_text.substring(0, cursorPosition + 1)); 273 if (++cursorPosition == more_authors_text.length){ 274 clearInterval(textAdder); 275 } 276 }, '10'); 277 ">2 more authors</span> 278 279 </div> 280 281 <!-- Journal/Book title and date --> 282 283 284 285 286 287 288 289 290 291 292 <div class="periodical"> 293 <em>IEEE Robotics and Automation Letters</em>, 2026 294 </div> 295 <div class="periodical"> 296 297 </div> 298 299 <!-- Links/Buttons --> 300 <div class="links"> 301 302 303 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 304 305 306 307 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 308 309 310 311 312 <a href="https://arxiv.org/pdf/2601.01155" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 313 314 315 316 317 <a href="https://drive.google.com/file/d/1Qp9IXjEZurAvdBfSAm2AeNoiBVhM_EFI/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 318 319 320 321 <a href="https://github.com/marmotlab/ORION" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 322 323 324 325 326 </div> 327 328 329 330 331 332 333 334 335 336 337 <!-- Hidden abstract block --> 338 <div class="abstract hidden"> 339 <p>Existing methods for multi-agent navigation typically assume fully known environments, offering limited support for partially known scenarios with outdated or imperfect prior maps, such as warehouses or factory floors. There, agents need to balance path optimality with collecting and sharing environmental information to help teammates reach their own targets. To these ends, we propose ORION, a novel deep reinforcement learning framework for cooperative multi-agent online navigation in partially known environments. Starting from an imperfect prior map, ORION trains agents to make decentralized decisions, coordinate toward individual targets, and actively reduce task-relevant map uncertainty through online observation sharing in a closed perceptionâaction loop. We first design a shared graph encoder that fuses prior map with online perception into a unified representation, providing robust state embeddings under environmental discrepancies. At the core of ORION is an option-critic framework that learns high-level cooperative modes translated into sequences of low-level actions, enabling adaptive switching between individual navigation and team-level exploration. We further introduce a dual-stage cooperation strategy that allows agents to assist teammates under map uncertainty, thereby reducing the overall makespan. Across extensive maze-like maps and large-scale warehouse environments, ORION achieves high-quality real-time decentralized cooperation while scaling to up to 10 robots, outperforming state-of-the-art classical and learning-based baselines. Finally, we validate ORION on physical robot teams, demonstrating its robustness and practicality for real-world cooperative navigation.</p> 340 </div> 341 342 343 344 <!-- Hidden bibtex block --> 345 <div class="bibtex hidden"> 346 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">zhang2026orion</span><span class="p">,</span> 347 <span class="na">title</span> <span class="p">=</span> <span class="s">{ORION: Option-Regularized Deep Reinforcement Learning for Cooperative Multi-Agent Online Navigation}</span><span class="p">,</span> 348 <span class="na">author</span> <span class="p">=</span> <span class="s">{Zhang, Shizhe and Liang, Jingsong and Zhou, Zhitao and Ye, Shuhan and Wang, Yizhuo and Tan, Derek Ming Siang and Chiun, Jimmy and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 349 <span class="na">
349journal</span> <span class="p">=</span> <span class="s">{IEEE Robotics and Automation Letters}</span><span class="p">,</span> 350 <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span> 351 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span><span class="p">,</span> 352<span class="p">}</span></code></pre></figure> 353 </div> 354 355 356 357 </div> 358</div> 359</li> 360<li> 361<div class="row"> 362 363 <div class="col col-sm-2 abbr"> 364 365 366 367 368 369<figure> 370 <picture> 371 <!-- Auto scaling with imagemagick --> 372 <!-- 373 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 374 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 375 --> 376 377 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/imaginav-480.webp 480w,/assets/img/publication_preview/imaginav-800.webp 800w,/assets/img/publication_preview/imaginav-1400.webp 1400w," sizes="200px" type="image/webp"></source> 378 379 <img src="/assets/img/publication_preview/imaginav.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="imaginav.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 380 </picture> 381 382 383</figure> 384 385 386 387 </div> 388 389 390 <!-- Entry bib key --> 391 <div id="chen2026imaginav" class="col-sm-8"> 392 <!-- Title --> 393 <div class="title">ImagiNav: Scalable Embodied Navigation via Generative Visual Prediction and Inverse Dynamics</div> 394 <!-- Author --> 395 <div class="author"> 396 397 398 399 Jie 400 Chen, Yuxin 401 Cai, Yizhuo 402 Wang, Ruofei 403 Bai, <em>Yuhong 404 Cao</em>, Jun 405 Li, Yau Wei 406 Yun, and 407 <span class="more-authors" title="click to view 1 more author" onclick=" 408 var element = $(this); 409 element.attr('title', ''); 410 var more_authors_text = element.text() == '1 more author' ? 'Guillaume Sartoretti' : '1 more author'; 411 var cursorPosition = 0; 412 var textAdder = setInterval(function(){ 413 element.text(more_authors_text.substring(0, cursorPosition + 1)); 414 if (++cursorPosition == more_authors_text.length){ 415 clearInterval(textAdder); 416 } 417 }, '10'); 418 ">1 more author</span> 419 420 </div> 421 422 <!-- Journal/Book title and date --> 423 424 425 426 427 428 429 430 431 432 433 <div class="periodical"> 434 <em>2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)</em>, 2026 435 </div> 436 <div class="periodical"> 437 438 </div> 439 440 <!-- Links/Buttons --> 441 <div class="links"> 442 443 444 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 445 446 447 448 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 449 450 451 452 453 <a href="https://arxiv.org/pdf/2603.13833" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 454 455 456 457 458 <a href="https://drive.google.com/file/d/1Q6WXRKYvgH_OBwIZXz0rde4BnagzBclz/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 459 460 461 462 463 464 465 </div> 466 467 468 469 470 471 472 473 474 475 476 <!-- Hidden abstract block --> 477 <div class="abstract hidden"> 478 <p>Enabling robots to navigate open-world environments via natural language is critical for general-purpose autonomy. Yet, Vision-Language Navigation has relied on end-to-end policies trained on expensive, embodiment-specific robot data. While recent foundation models trained on vast simulation data show promise, the challenge of scaling and generalizing due to the limited scene diversity and visual fidelity in simulation persists. To address this gap, we propose ImagiNav, a novel modular paradigm that decouples visual planning from robot actuation, enabling the direct utilization of diverse in-the-wild navigation videos. Our framework operates as a hierarchy: a Vision-Language Model first decomposes instructions into textual subgoals; a finetuned generative video model then imagines the future video trajectory towards that subgoal; finally, an inverse dynamics model extracts the trajectory from the imagined video, which can then be tracked via a low-level controller. We additionally develop a scalable data pipeline of in-the-wild navigation videos auto-labeled via inverse dynamics and a pretrained Vision-Language Model. ImagiNav demonstrates strong zero-shot transfer to robot navigation without requiring robot demonstrations, paving the way for generalist robots that learn navigation directly from unlabeled, open-world data.</p> 479 </div> 480 481 482 483 <!-- Hidden bibtex block --> 484 <div class="bibtex hidden"> 485 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">chen2026imaginav</span><span class="p">,</span> 486 <span class="na">title</span> <span class="p">=</span> <span class="s">{ImagiNav: Scalable Embodied Navigation via Generative Visual Prediction and Inverse Dynamics}</span><span class="p">,</span> 487 <span class="na">author</span> <span class="p">=</span> <span class="s">{Chen, Jie and Cai, Yuxin and Wang, Yizhuo and Bai, Ruofei and Cao, Yuhong and Li, Jun and Yun, Yau Wei and Sartoretti, Guillaume}</span><span class="p">,</span> 488 <span class="na">
488journal</span> <span class="p">=</span> <span class="s">{2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}</span><span class="p">,</span> 489 <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span><span class="p">,</span> 490 <span class="na">webiste</span> <span class="p">=</span> <span class="s">{https://j1dan.github.io/ImagiNav/}</span> 491<span class="p">}</span></code></pre></figure> 492 </div> 493 494 495 496 </div> 497</div> 498</li> 499<li> 500<div class="row"> 501 502 <div class="col col-sm-2 abbr"> 503 504 505 506 507 508<figure> 509 <picture> 510 <!-- Auto scaling with imagemagick --> 511 <!-- 512 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 513 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 514 --> 515 516 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/kivi-480.webp 480w,/assets/img/publication_preview/kivi-800.webp 800w,/assets/img/publication_preview/kivi-1400.webp 1400w," sizes="200px" type="image/webp"></source> 517 518 <img src="/assets/img/publication_preview/kivi.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="kivi.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 519 </picture> 520 521 522</figure> 523 524 525 526 </div> 527 528 529 <!-- Entry bib key --> 530 <div id="li2025kivi" class="col-sm-8"> 531 <!-- Title --> 532 <div class="title">KiVi: Kinesthetic-Visuospatial Integration for Dynamic and Safe Egocentric
532Legged Locomotion</div> 533 <!-- Author --> 534 <div class="author"> 535 536 537 538 Peizhuo 539 Li, Hongyi 540 Li, Yuxuan 541 Ma, Linnan 542 Chang, Xinrong 543 Yang, Ruiqi 544 Yu, Yifeng 545 Zhang, and 546 <span class="more-authors" title="click to view 3 more authors" onclick=" 547 var element = $(this); 548 element.attr('title', ''); 549 var more_authors_text = element.text() == '3 more authors' ? 'Yuhong Cao, Qiuguo Zhu, Guillaume Sartoretti' : '3 more authors'; 550 var cursorPosition = 0; 551 var textAdder = setInterval(function(){ 552 element.text(more_authors_text.substring(0, cursorPosition + 1)); 553 if (++cursorPosition == more_authors_text.length){ 554 clearInterval(textAdder); 555 } 556 }, '10'); 557 ">3 more authors</span> 558 559 </div> 560 561 <!-- Journal/Book title and date --> 562 563 564 565 566 567 568 569 570 571 572 <div class="periodical"> 573 <em>2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)</em>, 2026 574 </div> 575 <div class="periodical"> 576 577 </div> 578 579 <!-- Links/Buttons --> 580 <div class="links"> 581 582 583 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 584 585 586 587 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 588 589 590 591 592 <a href="https://www.arxiv.org/pdf/2509.23650" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 593 594 595 596 597 <a href="https://drive.google.com/file/d/1Oh5L6opqgko-5RTwq9Yz6vBt7WTxIxF6/view?usp=drivesdk" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 598 599 600 601 602 603 604 </div> 605 606 607 608 609 610 611 612 613 614 615 <!-- Hidden abstract block --> 616 <div class="abstract hidden"> 617 <p>Vision-based locomotion has shown great promise in enabling legged robots to perceive and adapt to complex environments. However, visual information is inherently fragile, being vulnerable to occlusions, reflections, and lighting changes, which often cause instability in locomotion. Inspired by animal sensorimotor integration, we propose KiVi, a Kinesthetic-Visuospatial integration framework, where kinesthetics encodes proprioceptive sensing of body motion and visuospatial reasoning captures visual perception of surrounding terrain. Specifically, KiVi separates these pathways, leveraging proprioception as a stable backbone while selectively incorporating vision for terrain awareness and obstacle avoidance. This modality-balanced, yet integrative design, combined with memory-enhanced attention, allows the robot to robustly interpret visual cues while maintaining fallback stability through proprioception. Extensive experiments show that our method enables quadruped robots to stably traverse diverse terrains and operate reliably in unstructured outdoor environments, remaining robust to out-of-distribution (OOD) visual noise and occlusion unseen during training, thereby highlighting its effectiveness and applicability to real-world legged locomotion.</p> 618 </div> 619 620 621 622 <!-- Hidden bibtex block --> 623 <div class="bibtex hidden"> 624 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">li2025kivi</span><span class="p">,</span> 625 <span class="na">title</span> <span class="p">=</span> <span class="s">{KiVi: Kinesthetic-Visuospatial Integration for Dynamic and Safe Egocentric
625Legged Locomotion}</span><span class="p">,</span> 626 <span class="na">author</span> <span class="p">=</span> <span class="s">{Li, Peizhuo and Li, Hongyi and Ma, Yuxuan and Chang, Linnan and Yang, Xinrong and Yu, Ruiqi and Zhang, Yifeng and Cao, Yuhong and Zhu, Qiuguo and Sartoretti, Guillaume}</span><span class="p">,</span> 627 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}</span><span class="p">,</span> 628 <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span> 629<span class="p">}</span></code></pre></figure> 630 </div> 631 632 633 634 </div> 635</div> 636</li> 637<li> 638<div class="row"> 639 640 <div class="col col-sm-2 abbr"> 641 642 643 644 645 646<figure> 647 <picture> 648 <!-- Auto scaling with imagemagick --> 649 <!-- 650 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 651 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 652 --> 653 654 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/apex-480.webp 480w,/assets/img/publication_preview/apex-800.webp 800w,/assets/img/publication_preview/apex-1400.webp 1400w," sizes="200px" type="image/webp"></source> 655 656 <img src="/assets/img/publication_preview/apex.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="apex.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 657 </picture> 658 659 660</figure> 661 662 663 664 </div> 665 666 667 <!-- Entry bib key --> 668 <div id="sood2025apex" class="col-sm-8"> 669 <!-- Title --> 670 <div class="title">APEX: Action Priors Enable Efficient Exploration for Skill Imitation on Articulated Robots</div> 671 <!-- Author --> 672 <div class="author"> 673 674 675 676 Shivam 677 Sood, Laukik 678 Nakhwa, Ge 679 Sun, <em>Yuhong 680 Cao</em>, Jin 681 Cheng, Fatemah 682 Zargarbashi, Taerim 683 Yoon, and 684 <span class="more-authors" title="click to view 3 more authors" onclick=" 685 var element = $(this); 686 element.attr('title', ''); 687 var more_authors_text = element.text() == '3 more authors' ? 'Sungjoon Choi, Stelian Coros, Guillaume Sartoretti' : '3 more authors'; 688 var cursorPosition = 0; 689 var textAdder = setInterval(function(){ 690 element.text(more_authors_text.substring(0, cursorPosition + 1)); 691 if (++cursorPosition == more_authors_text.length){ 692 clearInterval(textAdder); 693 } 694 }, '10'); 695 ">3 more authors</span> 696 697 </div> 698 699 <!-- Journal/Book title and date --> 700 701 702 703 704 705 706 707 708 709 710 <div class="periodical"> 711 <em>2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)</em>, 2026 712 </div> 713 <div class="periodical"> 714 715 </div> 716 717 <!-- Links/Buttons --> 718 <div class="links"> 719 720 721 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 722 723 724 725 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 726 727 728 729 730 <a href="https://www.arxiv.org/pdf/2505.10022" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 731 732 733 734 735 <a href="https://drive.google.com/file/d/1vY5M1HS7L5etdv09ygXzTuavgFsa1Qzm/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 736 737 738 739 740 741 742 <a href="https://marmotlab.github.io/kakashi/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Website</a> 743 744 </div> 745 746 747 748 749 750 751 752 753 754 755 <!-- Hidden abstract block --> 756 <div class="abstract hidden"> 757 <p>Learning by imitation provides an effective way for robots to develop well-regulated complex behaviors and directly benefit from natural demonstrations. State-of-the-art imitation learning (IL) approaches typically leverage Adversarial Motion Priors (AMP), which, despite their impressive results, suffer from two key limitations. They are prone to mode collapse, which often leads to overfitting to the simulation environment and thus increased sim-to-real gap, and they struggle to learn diverse behaviors effectively. To overcome these limitations, we introduce APEX (Action Priors enable Efficient eXploration): a simple yet versatile imitation learning framework that integrates demonstrations directly into reinforcement learning (RL), maintaining high exploration while grounding behavior with expert-informed priors. We achieve this through a combination of decaying action priors, which initially bias exploration toward expert demonstrations but gradually allow the policy to explore independently. This is complemented by a multi-critic RL framework that effectively balances stylistic consistency with task performance. Our approach achieves sample-efficient imitation learning and enables the acquisition of diverse skills within a single policy. APEX generalizes to varying velocities and preserves reference-like styles across complex tasks such as navigating rough terrain and climbing stairs, utilizing only flat-terrain kinematic motion data as a prior. We val
757idate our framework through extensive hardware experiments on the Unitree Go2 quadruped. There, APEX yields diverse and agile locomotion gaits, inherent gait transitions, and the highest reported speed for the platform to the best of our knowledge (peak velocity of  3.3 m/s on hardware). Our results establish APEX as a compelling alternative to existing IL methods, offering better efficiency, adaptability, and real-world performance.</p> 758 </div> 759 760 761 762 <!-- Hidden bibtex block --> 763 <div class="bibtex hidden"> 764 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">sood2025apex</span><span class="p">,</span> 765 <span class="na">title</span> <span class="p">=</span> <span class="s">{APEX: Action Priors Enable Efficient Exploration for Skill Imitation on Articulated Robots}</span><span class="p">,</span> 766 <span class="na">author</span> <span class="p">=</span> <span class="s">{Sood, Shivam and Nakhwa, Laukik and Sun, Ge and Cao, Yuhong and Cheng, Jin and Zargarbashi, Fatemah and Yoon, Taerim and Choi, Sungjoon and Coros, Stelian and Sartoretti, Guillaume}</span><span class="p">,</span> 767 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}</span><span class="p">,</span> 768 <span class="na">year</span> <span class="p">=</span> <span class="s">{2026}</span> 769<span class="p">}</span></code></pre></figure> 770 </div> 771 772 773 774 </div> 775</div> 776</li> 777</ol> 778<h2 class="bibliography">2025</h2> 779<ol class="bibliography"> 780<li> 781<div class="row"> 782 783 <div class="col col-sm-2 abbr"> 784 785 786 787 788 789<figure> 790 <picture> 791 <!-- Auto scaling with imagemagick --> 792 <!-- 793 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 794 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 795 --> 796 797 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/header-480.webp 480w,/assets/img/publication_preview/header-800.webp 800w,/assets/img/publication_preview/header-1400.webp 1400w," sizes="200px" type="image/webp"></source> 798 799 <img src="/assets/img/publication_preview/header.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="header.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 800 </picture> 801 802 803</figure> 804 805 806 807 </div> 808 809 810 <!-- Entry bib key --> 811 <div id="cao2025header" class="col-sm-8"> 812 <!-- Title --> 813 <div class="title">HEADER: Hierarchical Robot Exploration via Attention-Based Deep Reinforcement Learning with Expert-Guided Reward</div> 814 <!-- Author --> 815 <div class="author"> 816 817 818 819 <em>Yuhong 820 Cao</em>, Yizhuo 821 Wang, Jingsong 822 Liang, Shuhao 823 Liao, Yifeng 824 Zhang, Peizhuo 825 Li, and Guillaume 826 Sartoretti 827 </div> 828 829 <!-- Journal/Book title and date --> 830 831 832 833 834 835 836 837 838 839 840 <div class="periodical"> 841 <em>Under review</em>, 2025 842 </div> 843 <div class="periodical"> 844 845 </div> 846 847 <!-- Links/Buttons --> 848 <div class="links"> 849 850 851 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 852 853 854 855 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 856 857 858 859 860 <a href="https://arxiv.org/pdf/2510.15679" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 861 862 863 864 865 <a href="https://drive.google.com/file/d/1aG3a5452QwMNHUftXHqKqVALj2yP-7KC/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 866 867 868 869 <a href="https://github.com/marmotlab/ARiADNE2-ROS-Planner" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 870 871 872 873 874 </div> 875 876 877 878 879 880 881 882 883 884 885 <!-- Hidden abstract block --> 886 <div class="abstract hidden"> 887 <p>This work pushes the boundaries of learning-based methods in autonomous robot exploration in terms of environmental scale and exploration efficie
887ncy. We present HEADER, an attention-based reinforcement learning approach with hierarchical graphs for efficient exploration in large-scale environments. HEADER follows existing conventional methods to construct hierarchical representations for the robot belief/map, but further designs a novel community-based algorithm to construct and update a global graph, which remains fully incremental, shape-adaptive, and operates with linear complexity. Building upon attention-based networks, our planner finely reasons about the nearby belief within the local range while coarsely leveraging distant information at the global scale, enabling next-best-viewpoint decisions that consider multi-scale spatial dependencies. Beyond novel map representation, we introduce a parameter-free privileged reward that significantly improves model performance and produces near-optimal exploration behaviors, by avoiding training objective bias caused by handcrafted reward shaping. In simulated challenging, large-scale exploration scenarios, HEADER demonstrates better scalability than most existing learning and non-learning methods, while achieving a significant improvement in exploration efficiency (up to 20%) over state-of-the-art baselines. We also deploy HEADER on hardware and validate it in complex, large-scale real-life scenarios, including a 300m*230m campus environment.</p> 888 </div> 889 890 891 892 <!-- Hidden bibtex block --> 893 <div class="bibtex hidden"> 894 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">cao2025header</span><span class="p">,</span> 895 <span class="na">title</span> <span class="p">=</span> <span class="s">{HEADER: Hierarchical Robot Exploration via Attention-Based Deep Reinforcement Learning with Expert-Guided Reward}</span><span class="p">,</span> 896 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Wang, Yizhuo and Liang, Jingsong and Liao, Shuhao and Zhang, Yifeng and Li, Peizhuo and Sartoretti, Guillaume}</span><span class="p">,</span> 897 <span class="na">journal</span> <span class="p">=</span> <span class="s">{Under review}</span><span class="p">,</span> 898 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span> 899<span class="p">}</span></code></pre></figure> 900 </div> 901 902 903 904 </div> 905</div> 906</li> 907<li> 908<div class="row"> 909 910 <div class="col col-sm-2 abbr"> 911 912 913 914 915 916<figure> 917 <picture> 918 <!-- Auto scaling with imagemagick --> 919 <!-- 920 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 921 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 922 --> 923 924 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/helom-480.webp 480w,/assets/img/publication_preview/helom-800.webp 800w,/assets/img/publication_preview/helom-1400.webp 1400w," sizes="200px" type="image/webp"></source> 925 926 <img src="/assets/img/publication_preview/helom.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="helom.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 927 </picture> 928 929 930</figure> 931 932 933 934 </div> 935 936 937 <!-- Entry bib key --> 938 <div id="yang2025helom" class="col-sm-8"> 939 <!-- Title --> 940 <div class="title">HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation in Hexapod Robot</div> 941 <!-- Author --> 942 <div class="author"> 943 944 945 946 Xinrong 947 Yang, Peizhuo 948 Li, Hongyi 949 Li, Junkai 950 Lu, Linnan 951 Chang, <em>Yuhong 952 Cao</em>, Yifeng 953 Zhang, and 954 <span class="more-authors" title="click to view 2 more authors" onclick=" 955 var element = $(this); 956 element.attr('title', ''); 957 var more_authors_text = element.text() == '2 more authors' ? 'Ge Sun, Guillaume Sartoretti' : '2 more authors'; 958 var cursorPosition = 0; 959 var textAdder = setInterval(function(){ 960 element.text(more_authors_text.substring(0, cursorPosition + 1)); 961 if (++cursorPosition == more_authors_text.length){ 962 clearInterval(textAdder); 963 } 964 }, '10'); 965 ">2 more authors</span> 966 967 </div> 968 969 <!-- Journal/Book title and date --> 970 971 972 973 974 975 976 977 978 979 980 <div class="periodical"> 981 <em>Under review</em>, 2025 982 </div> 983 <div class="periodical"> 984 985 </div> 986 987 <!-- Links/Buttons --> 988 <div class="links"> 989 990 991 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 992 993 994 995 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 996 997 998 999 1000 <a href="https://arxiv.org/abs/2509.23651" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1001 1002 1003 1004 1005 <a href="https://drive.google.com/file/d/1jDrqZTLF5Ub1jWsbQDRULE2UeKcGldxu/view?usp=drivesdk" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1006 1007 1008 1009 1010 1011 1012 </div> 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 <!-- Hidden abstract block --> 1024 <div class="abstract hidden"> 1025 <p>Robots in real-world environments are often required to move/manipulate objects comparable in weight to their own bodies. Compared to grasping and carrying, pushing provides a more straightforward and efficie
1025nt non-prehensile manipulation strategy, avoiding complex grasp design while leveraging direct contact to regulate an objectâs pose. Achieving effective pushing, however, demands both sufficient manipulation forces and the ability to maintain stability, which is particularly challenging when dealing with heavy or irregular objects. To address these challenges, we propose HeLoM, a learning-based hierarchical whole-body manipulation framework for a hexapod robot that exploits coordinated multi-limb control. Inspired by the cooperative strategies of multi-legged insects, our framework leverages redundant contact points and high degrees of freedom to enable dynamic redistribution of contact forces. HeLoMâs high-level planner plans pushing behaviors and target object poses, while its low-level controller maintains locomotion stability and generates dynamically consistent joint actions. Our policies trained in simulation are directly deployed on real robots without additional fine-tuning. This design allows the robot to maintain balance while exerting continuous and controllable pushing forces through coordinated foreleg interaction and supportive hind-leg propulsion. We validate the effectiveness of HeLoM through both simulation and real-world experiments. Results show that our framework can stably push boxes of varying sizes and unknown physical properties to designated goal poses in the real world.</p> 1026 </div> 1027 1028 1029 1030 <!-- Hidden bibtex block --> 1031 <div class="bibtex hidden"> 1032 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">yang2025helom</span><span class="p">,</span> 1033 <span class="na">journal</span> <span class="p">=</span> <span class="s">{Under review}</span><span class="p">,</span> 1034 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1035 <span class="na">title</span> <span class="p">=</span> <span class="s">{HeLoM: Hierarchical Learning for Whole-Body Loco-Manipulation in Hexapod Robot}</span><span class="p">,</span> 1036 <span class="na">author</span> <span class="p">=</span> <span class="s">{Yang, Xinrong and Li, Peizhuo and Li, Hongyi and Lu, Junkai and Chang, Linnan and Cao, Yuhong and Zhang, Yifeng and Sun, Ge and Sartoretti, Guillaume}</span> 1037<span class="p">}</span></code></pre></figure> 1038 </div> 1039 1040 1041 1042 </div> 1043</div> 1044</li> 1045<li> 1046<div class="row"> 1047 1048 <div class="col col-sm-2 abbr"> 1049 1050 1051 1052 1053 1054<figure> 1055 <picture> 1056 <!-- Auto scaling with imagemagick --> 1057 <!-- 1058 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1059 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1060 --> 1061 1062 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/cogniplan.webp-480.webp 480w,/assets/img/publication_preview/cogniplan.webp-800.webp 800w,/assets/img/publication_preview/cogniplan.webp-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1063 1064 <img src="/assets/img/publication_preview/cogniplan.webp" class="preview z-depth-1 rounded" width="100%" height="auto" alt="cogniplan.webp" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1065 </picture> 1066 1067 1068</figure> 1069 1070 1071 1072 </div> 1073 1074 1075 <!-- Entry bib key --> 1076 <div id="wang2025cogniplan" class="col-sm-8"> 1077 <!-- Title --> 1078 <div class="title">CogniPlan: Uncertainty-Guided Path Planning with Conditional Generative Layout Prediction</div> 1079 <!-- Author --> 1080 <div class="author"> 1081 1082 1083 1084 Yizhuo 1085 Wang, Haodong 1086 He, Jingsong 1087 Liang, <em>Yuhong 1088 Cao</em>, Ritabrata 1089 Chakraborty, and Guillaume 1090 Sartoretti 1091 </div> 1092 1093 <!-- Journal/Book title and date --> 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 <div class="periodical"> 1105 <em>Conference on Robot Learning</em>, 2025 1106 </div> 1107 <div class="periodical"> 1108 1109 </div> 1110 1111 <!-- Links/Buttons --> 1112 <div class="links"> 1113 1114 1115 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1116 1117 1118 1119 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1120 1121 1122 1123 1124 <a href="https://www.arxiv.org/pdf/2508.03027" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1125 1126 1127 1128
1129 <a href="https://drive.google.com/file/d/1j-d2frHjUxDTm14VGeuwWLqowxG82SpT/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1130 1131 1132 1133 1134 1135 1136 <a href="https://yizhuo-wang.com/cogniplan/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Website</a> 1137 1138 </div> 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 <!-- Hidden abstract block --> 1150 <div class="abstract hidden"> 1151 <p>Path planning in unknown environments is a crucial yet inherently challenging capability for mobile robots, which primarily encompasses two coupled tasks: autonomous exploration and point-goal navigation. In both cases, the robot must perceive the environment, update its belief, and accurately estimate potential information gain on-the-fly to guide planning. In this work, we propose CogniPlan, a novel path planning framework that leverages multiple plausible layouts predicted by a COnditional GeNerative Inpainting model, mirroring how humans rely on cognitive maps during navigation. These predictions, based on the partially observed map and a set of layout conditioning vectors, enable our planner to reason effectively under uncertainty. We demonstrate strong synergy between generative image-based layout prediction and graph-attention-based path planning, allowing CogniPlan to combine the scalability of graph representations with the fidelity and predictiveness of occupancy maps, yielding notable performance gains in both exploration and navigation. We extensively evaluate CogniPlan on two datasets (hundreds of maps and realistic floor plans), consistently outperforming state-of-the-art planners. We further deploy it in a high-fidelity simulator and on hardware, showcasing its high-quality path planning and real-world applicability.</p> 1152 </div> 1153 1154 1155 1156 <!-- Hidden bibtex block --> 1157 <div class="bibtex hidden"> 1158 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">wang2025cogniplan</span><span class="p">,</span> 1159 <span class="na">title</span> <span class="p">=</span> <span class="s">{CogniPlan: Uncertainty-Guided Path Planning with Conditional Generative Layout Prediction}</span><span class="p">,</span> 1160 <span class="na">author</span> <span class="p">=</span> <span class="s">{Wang, Yizhuo and He, Haodong and Liang, Jingsong and Cao, Yuhong and Chakraborty, Ritabrata and Sartoretti, Guillaume}</span><span class="p">,</span> 1161 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1162 <span class="na">journal</span> <span class="p">=</span> <span class="s">{Conference on Robot Learning}</span> 1163<span class="p">}</span></code></pre></figure> 1164 </div> 1165 1166 1167 1168 </div> 1169</div> 1170</li> 1171<li> 1172<div class="row"> 1173 1174 <div class="col col-sm-2 abbr"> 1175 1176 1177 1178 1179 1180<figure> 1181 <picture> 1182 <!-- Auto scaling with imagemagick --> 1183 <!-- 1184 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1185 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1186 --> 1187 1188 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/search_tta-480.webp 480w,/assets/img/publication_preview/search_tta-800.webp 800w,/assets/img/publication_preview/search_tta-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1189 1190 <img src="/assets/img/publication_preview/search_tta.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="search_tta.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1191 </picture> 1192 1193 1194</figure> 1195 1196 1197 1198 </div> 1199 1200 1201 <!-- Entry bib key --> 1202 <div id="tan2025searchtta" class="col-sm-8"> 1203 <!-- Title --> 1204 <div class="title">Search-TTA: A Multi-Modal Test-Time Adaptation Framework for Visual Search in the Wild</div> 1205 <!-- Author --> 1206 <div class="author"> 1207 1208 1209 1210 Derek Ming Siang 1211 Tan, Boyang 1212 Liu, Alok 1213 Raj, Qi Xuan 1214 Ang, Weiheng 1215 Dai, Tanishq 1216 Duhan, Jimmy
1217 Chiun, and 1218 <span class="more-authors" title="click to view 3 more authors" onclick=" 1219 var element = $(this); 1220 element.attr('title', ''); 1221 var more_authors_text = element.text() == '3 more authors' ? 'Yuhong Cao, Florian Shkurti, Guillaume Sartoretti' : '3 more authors'; 1222 var cursorPosition = 0; 1223 var textAdder = setInterval(function(){ 1224 element.text(more_authors_text.substring(0, cursorPosition + 1)); 1225 if (++cursorPosition == more_authors_text.length){ 1226 clearInterval(textAdder); 1227 } 1228 }, '10'); 1229 ">3 more authors</span> 1230 1231 </div> 1232 1233 <!-- Journal/Book title and date --> 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 <div class="periodical"> 1245 <em>Conference on Robot Learning</em>, 2025 1246 </div> 1247 <div class="periodical"> 1248 1249 </div> 1250 1251 <!-- Links/Buttons --> 1252 <div class="links"> 1253 1254 1255 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1256 1257 1258 1259 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1260 1261 1262 1263 1264 <a href="https://arxiv.org/pdf/2505.11350" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1265 1266 1267 1268 1269 <a href="https://drive.google.com/file/d/1X4U7iSIDV07qIgsWtn4i6fNr-KKJ2_8X/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1270 1271 1272 1273 1274 1275 1276 <a href="https://search-tta.github.io/" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Website</a> 1277 1278 </div> 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 <!-- Hidden abstract block --> 1290 <div class="abstract hidden"> 1291 <p>To perform autonomous visual search for environmental monitoring, a robot may leverage satellite imagery as a prior map. This can help inform coarse, high level search and exploration strategies, even when such images lack sufficient resolution to allow fine-grained, explicit visual recognition of targets. However, there are some challenges to overcome with using satellite images to direct visual search. For one, targets that are unseen in satellite images are underrepresented (compared to real life) in most existing datasets, and thus vision models trained on these datasets fail to reason effectively based on indirect visual cues. Furthermore, approaches which leverage large Vision Language Models (VLMs) for generalization may yield inaccurate outputs due to hallucination, leading to inefficient search. To address these challenges, we introduce Search-TTA, a multimodal test-time adaptation framework that can accept text and/or image input. First, we pretrain a remote sensing image encoder to align with CLIPâs visual encoder to output probability distributions of target presence used for visual search. Second, our framework dynamically refines CLIPâs predictions during search using a test-time adaptation mechanism. Through a feedback loop inspired by Spatial Poisson Point Processes, gradient updates (weighted by uncertainty) are used to correct (potentially inaccurate) predictions and improve search performance. To validate Search-TTAâs performance, we curate a visual search dataset based on internet-scale ecological data. We find that Search-TTA improves planner performance by up to 9.7%, particularly in cases with poor initial CLIP predictions. It also achieves comparable performance to state-of-the-art VLMs. Finally, we deploy Search-TTA on a real UAV via hardware-in-the-loop testing, by simulating its operation within a large-scale simulation that provides onboard sensing.</p> 1292 </div> 1293 1294 1295 1296 <!-- Hidden bibtex block --> 1297 <div class="bibtex hidden"> 1298 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">tan2025searchtta</span><span class="p">,</span> 1299 <span class="na">title</span> <span class="p">=</span> <span class="s">{Search-TTA: A Multi-Modal Test-Time Adaptation Framework for Visual Search in the Wild}</span><span class="p">,</span> 1300 <span class="na">author</span> <span class="p">=</span> <span class="s">{Tan, Derek Ming Siang and Liu, Boyang and Raj, Alok and Ang, Qi Xuan and Dai, Weiheng and Duhan, Tanishq and Chiun, Jimmy and Cao, Yuhong and Shkurti, Florian and Sartoretti, Guillaume}</span><span class="p">,</span> 1301 <span class="na">
1301journal</span> <span class="p">=</span> <span class="s">{Conference on Robot Learning}</span><span class="p">,</span> 1302 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span> 1303<span class="p">}</span></code></pre></figure> 1304 </div> 1305 1306 1307 1308 </div> 1309</div> 1310</li> 1311<li> 1312<div class="row"> 1313 1314 <div class="col col-sm-2 abbr"> 1315 1316 1317 1318 1319 1320<figure> 1321 <picture> 1322 <!-- Auto scaling with imagemagick --> 1323 <!-- 1324 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1325 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1326 --> 1327 1328 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/sata-480.webp 480w,/assets/img/publication_preview/sata-800.webp 800w,/assets/img/publication_preview/sata-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1329 1330 <img src="/assets/img/publication_preview/sata.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="sata.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1331 </picture> 1332 1333 1334</figure> 1335 1336 1337 1338 </div> 1339 1340 1341 <!-- Entry bib key --> 1342 <div id="li2025sata" class="col-sm-8"> 1343 <!-- Title --> 1344 <div class="title">SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal Learning</div> 1345 <!-- Author --> 1346 <div class="author"> 1347 1348 1349 1350 Peizhuo 1351 Li, Hongyi 1352 Li, Ge 1353 Sun, Jin 1354 Cheng, Xinrong 1355 Yang, Guillaume 1356 Bellegarda, Milad 1357 Shafiee, and 1358 <span class="more-authors" title="click to view 3 more authors" onclick=" 1359 var element = $(this); 1360 element.attr('title', ''); 1361 var more_authors_text = element.text() == '3 more authors' ? 'Yuhong Cao, Auke Ijspeert, Guillaume Sartoretti' : '3 more authors'; 1362 var cursorPosition = 0; 1363 var textAdder = setInterval(function(){ 1364 element.text(more_authors_text.substring(0, cursorPosition + 1)); 1365 if (++cursorPosition == more_authors_text.length){ 1366 clearInterval(textAdder); 1367 } 1368 }, '10'); 1369 ">3 more authors</span> 1370 1371 </div> 1372 1373 <!-- Journal/Book title and date --> 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 <div class="periodical"> 1385 <em>2025 Robotics: Science and Systems (RSS)</em>, 2025 1386 </div> 1387 <div class="periodical"> 1388 1389 </div> 1390 1391 <!-- Links/Buttons --> 1392 <div class="links"> 1393 1394 1395 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1396 1397 1398 1399 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1400 1401 1402 1403 1404 <a href="https://www.arxiv.org/pdf/2502.12674" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1405 1406 1407 1408 1409 <a href="https://drive.google.com/file/d/1PLxFbyujH_4cdFEy0wmO5gsHDFZASIUV/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1410 1411 1412 1413 <a href="https://github.com/marmotlab/SATA" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 1414 1415 1416 1417 1418 </div> 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 <!-- Hidden abstract block --> 1430 <div class="abstract hidden"> 1431 <p>Despite recent advances in learning-based controllers for legged robots, deployments in human-centric environments remain limited by safety concerns. Most of these approaches use position-based control, where policies output target joint angles that must be processed by a low-level controller (e.g., PD or impedance controllers) to compute joint torques. Although impressive results have been achieved in controlled real-world scenarios, these methods often struggle with compliance and adaptability when encountering environments or disturbances unseen during training, potentially resulting in extreme or unsafe behaviors. Inspired by how animals achieve smooth and adaptive movements by controlling muscle extension and contraction, torque-based policies offer a promising alternative by enabling precise and direct control of the actuators in torque space. In principle, this approach facilitates more effective interactions with the environment, resulting in safer and more adaptable behaviors. However, challenges such as a highly nonlinear state space and inefficient exploration during training have hindered their broader adoption. To address these limitations, we propose SATA, a bio-inspired framework that mimics key biomechanical principles and adaptive learning mechanisms observed in animal locomotion. Our approach effectively addresses the inherent challenges of learning torque-based policies by significantly improving early-stage exploration, leading to high-performance final policies. Remarkably, our method achieves zero-shot sim-to-real transfer. Our experimental results indicate that SATA demonstrates remarkable compliance and safety, even in challenging environments such as soft/slippery terrain or narrow passages, and under significant external disturbances, highlighting its potential for practical deployments in human-centric and safety-critical scenarios.</p> 1432 </div> 1433 1434 1435 1436 <!-- Hidden bibtex block --> 1437 <div class="bibtex hidden"> 1438 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">li2025sata</span><span class="p">,</span> 1439 <span class="na">title</span> <span class="p">=</span> <span class="s">
1439{SATA: Safe and Adaptive Torque-Based Locomotion Policies Inspired by Animal Learning}</span><span class="p">,</span> 1440 <span class="na">author</span> <span class="p">=</span> <span class="s">{Li, Peizhuo and Li, Hongyi and Sun, Ge and Cheng, Jin and Yang, Xinrong and Bellegarda, Guillaume and Shafiee, Milad and Cao, Yuhong and Ijspeert, Auke and Sartoretti, Guillaume}</span><span class="p">,</span> 1441 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2025 Robotics: Science and Systems (RSS)}</span><span class="p">,</span> 1442 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1443<span class="p">}</span></code></pre></figure> 1444 </div> 1445 1446 1447 1448 </div> 1449</div> 1450</li> 1451<li> 1452<div class="row"> 1453 1454 <div class="col col-sm-2 abbr"> 1455 1456 1457 1458 1459 1460<figure> 1461 <picture> 1462 <!-- Auto scaling with imagemagick --> 1463 <!-- 1464 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1465 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1466 --> 1467 1468 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/marvel-480.webp 480w,/assets/img/publication_preview/marvel-800.webp 800w,/assets/img/publication_preview/marvel-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1469 1470 <img src="/assets/img/publication_preview/marvel.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="marvel.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1471 </picture> 1472 1473 1474</figure> 1475 1476 1477 1478 </div> 1479 1480 1481 <!-- Entry bib key --> 1482 <div id="chiun2025marvel" class="col-sm-8"> 1483 <!-- Title --> 1484 <div class="title">MARVEL: Multi-Agent Reinforcement Learning for constrained field-of-View multi-robot Exploration in Large-scale environments</div> 1485 <!-- Author --> 1486 <div class="author"> 1487 1488 1489 1490 Jimmy 1491 Chiun, Shizhe 1492 Zhang, Yizhuo 1493 Wang, <em>Yuhong 1494 Cao</em>, and Guillaume 1495 Sartoretti 1496 </div> 1497 1498 <!-- Journal/Book title and date --> 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 <div class="periodical"> 1510 <em>2025 IEEE International Conference on Robotics and Automation (ICRA)</em>, 2025 1511 </div> 1512 <div class="periodical"> 1513 1514 </div> 1515 1516 <!-- Links/Buttons --> 1517 <div class="links"> 1518 1519 1520 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1521 1522 1523 1524 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1525 1526 1527 1528 1529 <a href="https://marmotlab.org/publications/77-ICRA2025-MARVEL.pdf" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1530 1531 1532 1533 1534 <a href="https://www.youtube.com/watch?v=uwQqkruBvwY" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1535 1536 1537 1538 <a href="https://github.com/marmotlab/MARVEL" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 1539 1540 1541 1542 1543 </div> 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 <!-- Hidden abstract block --> 1555 <div class="abstract hidden"> 1556 <p>In multi-robot exploration, a team of mobile robot is tasked with efficiently mapping an unknown environments. While most exploration planners assume omnidirectional sensors like LiDAR, this is impractical for small robots such as drones, where lightweight, directional sensors like cameras may be the only option due to payload constraints. These sensors have a constrained field-of-view (FoV), which adds complexity to the exploration problem, requiring not only optimal robot positioning but also sensor orientation during movement. In this work, we propose MARVEL, a neural framework that leverages graph attention networks, together with novel frontiers and orientation features fusion technique, to develop a collaborative, decentralized policy using multi-agent reinfor
1556cement learning (MARL) for robots with constrained FoV. To handle the large action space of viewpoints planning, we further introduce a novel information-driven action pruning strategy. MARVEL improves multi-robot coordination and decision-making in challenging large-scale indoor environments, while adapting to various team sizes and sensor configurations (ie, FoV and sensor range) without additional training. Our extensive evaluation shows that MARVELâs learned policies exhibit effective coordinated behaviors, outperforming state-ofthe-art exploration planners across multiple metrics. We experimentally demonstrate MARVELâs generalizability in large-scale environments, of up to 90m by 90m, and validate its practical applicability through successful deployment on a team of real drone hardware.</p> 1557 </div> 1558 1559 1560 1561 <!-- Hidden bibtex block --> 1562 <div class="bibtex hidden"> 1563 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">chiun2025marvel</span><span class="p">,</span> 1564 <span class="na">title</span> <span class="p">=</span> <span class="s">{MARVEL: Multi-Agent Reinforcement Learning for constrained field-of-View multi-robot Exploration in Large-scale environments}</span><span class="p">,</span> 1565 <span class="na">author</span> <span class="p">=</span> <span class="s">{Chiun, Jimmy and Zhang, Shizhe and Wang, Yizhuo and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 1566 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2025 IEEE International Conference on Robotics and Automation (ICRA)}</span><span class="p">,</span> 1567 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1568 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span> 1569<span class="p">}</span></code></pre></figure> 1570 </div> 1571 1572 1573 1574 </div> 1575</div> 1576</li> 1577<li> 1578<div class="row"> 1579 1580 <div class="col col-sm-2 abbr"> 1581 1582 1583 1584 1585 1586<figure> 1587 <picture> 1588 <!-- Auto scaling with imagemagick --> 1589 <!-- 1590 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1591 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1592 --> 1593 1594 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/sigma-480.webp 480w,/assets/img/publication_preview/sigma-800.webp 800w,/assets/img/publication_preview/sigma-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1595 1596 <img src="/assets/img/publication_preview/sigma.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="sigma.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1597 </picture> 1598 1599 1600</figure> 1601 1602 1603 1604 </div> 1605 1606 1607 <!-- Entry bib key --> 1608 <div id="liao2025sigma" class="col-sm-8"> 1609 <!-- Title --> 1610 <div class="title">SIGMA: Sheaf-Informed Geometric Multi-Agent Pathfinding</div> 1611 <!-- Author --> 1612 <div class="author"> 1613 1614 1615 1616 Shuhao 1617 Liao, Weihang 1618 Xia, <em>Yuhong 1619 Cao</em>, Weiheng 1620 Dai, Chengyang 1621 He, Wenjun 1622 Wu, and Guillaume 1623 Sartoretti 1624 </div> 1625 1626 <!-- Journal/Book title and date --> 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 <div class="periodical"> 1638 <em>2025 IEEE International Conference on Robotics and Automation (ICRA)</em>, 2025 1639 </div> 1640 <div class="periodical"> 1641 1642 </div> 1643 1644 <!-- Links/Buttons --> 1645 <div class="links"> 1646 1647 1648 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1649 1650 1651 1652 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1653 1654 1655 1656 1657 <a href="https://arxiv.org/pdf/2502.06440" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1658 1659 1660 1661 1662 1663 1664 <a href="https://github.com/marmotlab/SIGMA" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 1665 1666 1667 1668 1669 </div> 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 <!-- Hidden abstract block --> 1681 <div class="abstract hidden"> 1682 <p>The Multi-Agent Path Finding (MAPF) problem aims to determine the shortest and collision-free paths for multiple agents in a known, potentially obstacle-ridden environment. It is the core challenge for robotic deployments in large-scale logistics and transportation. Decentralized learning-based approaches have shown great potential for addressing the MAPF problems, offering more reactive and scalable solutions. However, existing learning-based MAPF methods usually rely on agents making decisions based on a limited field of view (FOV), resulting in short-sighted policies and inefficient cooperation in complex scenarios. There, a critical challenge is to achieve consensus on potential movements between agents based on limited observations and communications. To tackle this challenge, we introduce a new framework that applies sheaf theory to decentralized deep reinforcement learning, enabling agents to learn geometric cross-dependencies between each other through local consensus and utilize them for tightly cooperative decision-making. In particular, sheaf theory provides a mathematical proof of conditions for achieving global consensus through local observation. Inspired by this, we incorporate a neural network to approximately model the consensus in latent space based on sheaf theory and train it through self-supervised learning. During the task, in addition to normal features for MAPF as in previous works, each agent distributedly reasons about a learned consensus feature, leading to efficie
1682nt cooperation on pathfinding and collision avoidance. As a result, our proposed method demonstrates significant improvements over state-of-the-art learning-based MAPF planners, especially in relatively large and complex scenarios, demonstrating its superiority over baselines in various simulations and real-world robot experiments.</p> 1683 </div> 1684 1685 1686 1687 <!-- Hidden bibtex block --> 1688 <div class="bibtex hidden"> 1689 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">liao2025sigma</span><span class="p">,</span> 1690 <span class="na">title</span> <span class="p">=</span> <span class="s">{SIGMA: Sheaf-Informed Geometric Multi-Agent Pathfinding}</span><span class="p">,</span> 1691 <span class="na">author</span> <span class="p">=</span> <span class="s">{Liao, Shuhao and Xia, Weihang and Cao, Yuhong and Dai, Weiheng and He, Chengyang and Wu, Wenjun and Sartoretti, Guillaume}</span><span class="p">,</span> 1692 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2025 IEEE International Conference on Robotics and Automation (ICRA)}</span><span class="p">,</span> 1693 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1694 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span> 1695<span class="p">}</span></code></pre></figure> 1696 </div> 1697 1698 1699 1700 </div> 1701</div> 1702</li> 1703<li> 1704<div class="row"> 1705 1706 <div class="col col-sm-2 abbr"> 1707 1708 1709 1710 1711 1712<figure> 1713 <picture> 1714 <!-- Auto scaling with imagemagick --> 1715 <!-- 1716 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1717 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1718 --> 1719 1720 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/dare-480.webp 480w,/assets/img/publication_preview/dare-800.webp 800w,/assets/img/publication_preview/dare-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1721 1722 <img src="/assets/img/publication_preview/dare.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="dare.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1723 </picture> 1724 1725 1726</figure> 1727 1728 1729 1730 </div> 1731 1732 1733 <!-- Entry bib key --> 1734 <div id="cao2025dare" class="col-sm-8"> 1735 <!-- Title --> 1736 <div class="title">DARE: Diffusion Policy for Autonomous Robot Exploration</div> 1737 <!-- Author --> 1738 <div class="author"> 1739 1740 1741 1742 <em>Yuhong 1743 Cao</em>, Jeric 1744 Lew, Jingsong 1745 Liang, Jin 1746 Cheng, and Guillaume 1747 Sartoretti 1748 </div> 1749 1750 <!-- Journal/Book title and date --> 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 <div class="periodical"> 1762 <em>2025 IEEE International Conference on Robotics and Automation (ICRA)</em>, 2025 1763 </div> 1764 <div class="periodical"> 1765 1766 </div> 1767 1768 <!-- Links/Buttons --> 1769 <div class="links"> 1770 1771 1772 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1773 1774 1775 1776 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1777 1778 1779 1780 1781 <a href="https://arxiv.org/pdf/2410.16687" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1782 1783 1784 1785 1786 <a href="https://drive.google.com/file/d/1fMzJoOQKK_nfPjQDta6mOIjewsQhN4rB/view?usp=drive_link" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1787 1788 1789 1790 <a href="https://github.com/marmotlab/DARE" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 1791 1792 1793 1794 1795 </div> 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 <!-- Hidden abstract block --> 1807 <div class="abstract hidden"> 1808 <p>
1808Autonomous robot exploration requires a robot to efficiently explore and map unknown environments. Compared to conventional methods that can only optimize paths based on the current robot belief, learning-based methods show the potential to achieve improved performance by drawing on past experiences to reason about unknown areas. In this paper, we propose DARE, a novel generative approach that leverages diffusion models trained on expert demonstrations, which can explicitly generate an exploration path through one-time inference. We build DARE upon an attention-based encoder and a diffusion policy model, and introduce ground truth optimal demonstrations for training to learn better patterns for exploration. The trained planner can reason about the partial belief to recognize the potential structure in unknown areas and consider these areas during path planning. Our experiments demonstrate that DARE achieves on-par performance with both conventional and learning-based state-of-the-art exploration planners, as well as good generalizability in both simulations and real-life scenarios.</p> 1809 </div> 1810 1811 1812 1813 <!-- Hidden bibtex block --> 1814 <div class="bibtex hidden"> 1815 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">cao2025dare</span><span class="p">,</span> 1816 <span class="na">title</span> <span class="p">=</span> <span class="s">{DARE: Diffusion Policy for Autonomous Robot Exploration}</span><span class="p">,</span> 1817 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Lew, Jeric and Liang, Jingsong and Cheng, Jin and Sartoretti, Guillaume}</span><span class="p">,</span> 1818 <span class="na">journal</span> <span class="p">=</span> <span class="s">{2025 IEEE International Conference on Robotics and Automation (ICRA)}</span><span class="p">,</span> 1819 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1820 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span> 1821<span class="p">}</span></code></pre></figure> 1822 </div> 1823 1824 1825 1826 </div> 1827</div> 1828</li> 1829<li> 1830<div class="row"> 1831 1832 <div class="col col-sm-2 abbr"> 1833 1834 1835 1836 1837 1838<figure> 1839 <picture> 1840 <!-- Auto scaling with imagemagick --> 1841 <!-- 1842 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1843 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1844 --> 1845 1846 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/hetmrta-480.webp 480w,/assets/img/publication_preview/hetmrta-800.webp 800w,/assets/img/publication_preview/hetmrta-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1847 1848 <img src="/assets/img/publication_preview/hetmrta.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="hetmrta.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1849 </picture> 1850 1851 1852</figure> 1853 1854 1855 1856 </div> 1857 1858 1859 <!-- Entry bib key --> 1860 <div id="dai2025heterogeneous" class="col-sm-8"> 1861 <!-- Title --> 1862 <div class="title">Heterogeneous Multi-robot Task Allocation and Scheduling via Reinforcement Learning</div> 1863 <!-- Author --> 1864 <div class="author"> 1865 1866 1867 1868 Weiheng 1869 Dai, Utkarsh 1870 Rai, Jimmy 1871 Chiun, <em>Yuhong 1872 Cao</em>, and Guillaume 1873 Sartoretti 1874 </div> 1875 1876 <!-- Journal/Book title and date --> 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 <div class="periodical"> 1888 <em>IEEE Robotics and Automation Letters</em>, 2025 1889 </div> 1890 <div class="periodical"> 1891 1892 </div> 1893 1894 <!-- Links/Buttons --> 1895 <div class="links"> 1896 1897 1898 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 1899 1900 1901 1902 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 1903 1904 1905 1906 1907 <a href="https://marmotlab.org/publications/73-RAL2025-HetMRTA.pdf" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 1908 1909 1910 1911 1912 <a href="https://drive.google.com/file/d/1nJze4E4Unl7p0mQH0YpJVnVVVzxU5ruR/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 1913 1914 1915 1916 <a href="https://github.com/marmotlab/HeteroMRTA" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 1917 1918 1919 1920 1921 </div> 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 <!-- Hidden abstract block --> 1933 <div class="abstract hidden"> 1934 <p>Many multi-robot applications require allocating a team of heterogeneous agents (robots) with different abilities to cooperatively complete a given set of spatially distributed tasks as quickly as possible. We focus on tasks that can only be initiated when all required agents are present otherwise arrived agents would be waiting idly. Agents need to not only execute a sequence of tasks by dynamically forming and disbanding teams to satisfy/match diverse ability requirements of each task but also account for the schedules of other agents to minimize unnecessary idle time. Conventional methods such as mix-integer programming generally require centralized scheduling and a long optimization time, which limits their potential for real-world applications. In this work, we propose a reinforcement learning framework to train a decentralized policy applicable to heterogeneous agents. To address the challenge of complex cooperation learning, we further introduce a constrained flashforward mechanism to guide/constrain the agentsâ exploration and help them make better predictions. Through an attention mechanism that reasons about both
1934short-term cooperation and long-term scheduling dependency, agents learn to reactively choose their next tasks (and subsequent coalitions) to avoid wasting abilities and to shorten the overall task completion time (makespan). We compare our method with state-of-the-art heuristic and mixed-integer programming methods, demonstrating its generalization ability and showing it closely matches or outperforms these baselines while remaining at least two orders of magnitude faster.</p> 1935 </div> 1936 1937 1938 1939 <!-- Hidden bibtex block --> 1940 <div class="bibtex hidden"> 1941 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">dai2025heterogeneous</span><span class="p">,</span> 1942 <span class="na">title</span> <span class="p">=</span> <span class="s">{Heterogeneous Multi-robot Task Allocation and Scheduling via Reinforcement Learning}</span><span class="p">,</span> 1943 <span class="na">author</span> <span class="p">=</span> <span class="s">{Dai, Weiheng and Rai, Utkarsh and Chiun, Jimmy and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 1944 <span class="na">journal</span> <span class="p">=</span> <span class="s">{IEEE Robotics and Automation Letters}</span><span class="p">,</span> 1945 <span class="na">year</span> <span class="p">=</span> <span class="s">{2025}</span><span class="p">,</span> 1946 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span> 1947<span class="p">}</span></code></pre></figure> 1948 </div> 1949 1950 1951 1952 </div> 1953</div> 1954</li> 1955</ol> 1956<h2 class="bibliography">2024</h2> 1957<ol class="bibliography"> 1958<li> 1959<div class="row"> 1960 1961 <div class="col col-sm-2 abbr"> 1962 1963 1964 1965 1966 1967<figure> 1968 <picture> 1969 <!-- Auto scaling with imagemagick --> 1970 <!-- 1971 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 1972 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 1973 --> 1974 1975 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/hdplanner-480.webp 480w,/assets/img/publication_preview/hdplanner-800.webp 800w,/assets/img/publication_preview/hdplanner-1400.webp 1400w," sizes="200px" type="image/webp"></source> 1976 1977 <img src="/assets/img/publication_preview/hdplanner.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="hdplanner.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 1978 </picture> 1979 1980 1981</figure> 1982 1983 1984 1985 </div> 1986 1987 1988 <!-- Entry bib key --> 1989 <div id="liang2024hdplanner" class="col-sm-8"> 1990 <!-- Title --> 1991 <div class="title">HDPlanner: Advancing Autonomous Deployments in Unknown Environments through Hierarchical Decision Networks</div> 1992 <!-- Author --> 1993 <div class="author"> 1994 1995 1996 1997 Jingsong 1998 Liang, <em>Yuhong 1999 Cao</em>, Yixiao 2000 Ma, Hanqi 2001 Zhao, and Guillaume 2002 Sartoretti 2003 </div> 2004 2005 <!-- Journal/Book title and date --> 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 <div class="periodical"> 2017 <em>IEEE Robotics and Automation Letters</em>, 2024 2018 </div> 2019 <div class="periodical"> 2020 2021 </div> 2022 2023 <!-- Links/Buttons --> 2024 <div class="links"> 2025 2026 2027 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2028 2029 2030 2031 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2032 2033 2034 2035 2036 <a href="https://arxiv.org/pdf/2408.03768" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2037 2038 2039 2040 2041 <a href="https://drive.google.com/file/d/1FWyKCfvqxUaYldor781FxEAHeChyUR4z/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2042 2043 2044 2045 <a href="https://github.com/marmotlab/HDPlanner_Exp_and_Nav" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2046 2047 2048 2049 2050 </div> 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 <!-- Hidden abstract block --> 2062 <div class="abstract hidden"> 2063 <p>In this paper, we introduce HDPlanner, a deep reinforcement learning (DRL) based framework designed to tackle two core and challenging tasks for mobile robots: autonomous exploration and navigation, where the robot must optimize its trajectory adaptively to achieve the task objective through continuous interactions in unknown environments. Specifically, HDPlanner relies on novel hierarchical attention networks to empower the robot to reason about its belief across multiple spatial scales and sequence collaborative decisions, where our networks decompose long-term objectives into short-term informative task assignments and informative path plannings. We further propose a contrastive learning-based joint optimization to enhance the robustness of HDPlanner. We empirically demonstrate that HDPlanner significantly outperforms state-of-the-art conventional and learning-based baselines on an extensive set of simulations, including hundreds of test maps and large-scale, complex Gazebo environments. Notably, HDPlanner achieves real-time planning with travel distances reduced by up to 35.7% compared to exploration benchmarks and by up to 16.5% than navigation benchmarks. Furthermore, we validate our approach on hardware, where it generates high-quality, adaptive trajectories in both indoor and outdoor environments, highlighting its real-world applicability without additional training.</p> 2064 </div> 2065 2066 2067 2068 <!-- Hidden bibtex block --> 2069 <div class="bibtex hidden"> 2070 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">liang2024hdplanner</span><span class="p">,</span> 2071 <span class="na">title</span> <span class="p">=</span> <span class="s">{HDPlanner: Advancing Autonomous Deployments in Unknown Environments through Hierarchical Decision Networks}</span><span class="p">,</span> 2072 <span class="na">author</span> <span class="p">=</span> <span class="s">{Liang, Jingsong and Cao, Yuhong and Ma, Yixiao and Zhao, Hanqi and Sartoretti, Guillaume}</span><span class="p">,</span> 2073 <span class="na">
2073journal</span> <span class="p">=</span> <span class="s">{IEEE Robotics and Automation Letters}</span><span class="p">,</span> 2074 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span> 2075 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span><span class="p">,</span> 2076<span class="p">}</span></code></pre></figure> 2077 </div> 2078 2079 2080 2081 </div> 2082</div> 2083</li> 2084<li> 2085<div class="row"> 2086 2087 <div class="col col-sm-2 abbr"> 2088 2089 2090 2091 2092 2093<figure> 2094 <picture> 2095 <!-- Auto scaling with imagemagick --> 2096 <!-- 2097 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2098 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2099 --> 2100 2101 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/viper-480.webp 480w,/assets/img/publication_preview/viper-800.webp 800w,/assets/img/publication_preview/viper-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2102 2103 <img src="/assets/img/publication_preview/viper.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="viper.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2104 </picture> 2105 2106 2107</figure> 2108 2109 2110 2111 </div> 2112 2113 2114 <!-- Entry bib key --> 2115 <div id="wang2024viper" class="col-sm-8"> 2116 <!-- Title --> 2117 <div class="title">ViPER: Visibility-based Pursuit-Evasion via Reinforcement Learning</div> 2118 <!-- Author --> 2119 <div class="author"> 2120 2121 2122 2123 Yizhuo 2124 Wang*, <em>Yuhong 2125 Cao*</em>, Jimmy 2126 Chiun, Koley 2127 Subhadeep, Pham 2128 Mandy, and Guillaume 2129 Sartoretti 2130 </div> 2131 2132 <!-- Journal/Book title and date --> 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 <div class="periodical"> 2144 <em>In Conference on Robot Learning</em> , 2024 2145 </div> 2146 <div class="periodical"> 2147 2148 </div> 2149 2150 <!-- Links/Buttons --> 2151 <div class="links"> 2152 2153 2154 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2155 2156 2157 2158 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2159 2160 2161 2162 2163 <a href="https://openreview.net/pdf?id=EPujQZWemk" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2164 2165 2166 2167 2168 <a href="https://drive.google.com/file/d/189UFUa1BALAe8GTF2efyd195l6D2GqVH/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2169 2170 2171 2172 <a href="https://github.com/marmotlab/ViPER" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2173 2174 2175 2176 2177 </div> 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 <!-- Hidden abstract block --> 2189 <div class="abstract hidden"> 2190 <p>In visibility-based pursuit-evasion tasks, a team of mobile pursuer robots with limited sensing capabilities is tasked with detecting all evaders in a multiply-connected planar environment, whose map may or may not be known to pursuers beforehand. 2191This requires tight coordination among multiple agents to ensure that the omniscient and potentially arbitrarily fast evaders are guaranteed to be detected by the pursuers. 2192Whereas existing methods typically rely on a relatively large team of agents to clear the environment, we propose ViPER, a neural solution that leverages a graph attention network to learn a coordinated yet distributed policy via multi-agent reinforcement learning (MARL). 2193We experimentally demonstrate that ViPER significantly outperforms other state-of-the-art non-learning planners, showcasing its emergent coordinated behaviors and adaptability to more challenging scenarios and various team sizes, and finally deploy its learned policies on hardware in an aerial search task.</p> 2194 </div> 2195 2196 2197 2198 <!-- Hidden bibtex block --> 2199 <div class="bibtex hidden"> 2200 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">wang2024viper</span><span class="p">,</span> 2201 <span class="na">title</span> <span class="p">=</span> <span class="s">{ViPER: Visibility-based Pursuit-Evasion via Reinforcement Learning}</span><span class="p">,</span> 2202 <span class="na">author</span> <span class="p">=</span> <span class="s">{Wang*, Yizhuo and Cao*, Yuhong and Chiun, Jimmy and Subhadeep, Koley and Mandy, Pham and Sartoretti, Guillaume}</span><span class="p">,</span> 2203 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Conference on Robot Learning}</span><span class="p">,</span> 2204 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span> 2205<span class="p">}</span></code></pre></figure> 2206 </div> 2207 2208 2209 2210 </div> 2211</div> 2212</li> 2213<li> 2214<div class="row"> 2215 2216 <div class="col col-sm-2 abbr"> 2217 2218 2219 2220 2221 2222<figure> 2223 <picture> 2224 <!-- Auto scaling with imagemagick --> 2225 <!-- 2226 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2227 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2228 --> 2229 2230 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/bandwidth-limited-exploration-480.webp 480w,/assets/img/publication_preview/bandwidth-limited-exploration-800.webp 800w,/assets/img/publication_preview/bandwidth-limited-exploration-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2231 2232 <img src="/assets/img/publication_preview/bandwidth-limited-exploration.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="bandwidth-limited-exploration.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2233 </picture> 2234 2235 2236</figure> 2237 2238 2239 2240 </div> 2241 2242 2243 <!-- Entry bib key --> 2244 <div id="ma2024privileged" class="col-sm-8"> 2245 <!-- Title --> 2246 <div class="title">Privileged Reinforcement and Communication Learning for Distributed, Bandwidth-limited Multi-robot Exploration</div> 2247 <!-- Author --> 2248 <div class="author"> 2249 2250 2251 2252 Yixiao 2253 Ma, Jingsong 2254 Liang, <em>Yuhong 2255 Cao</em>, Derek Ming Siang 2256 Tan, and Guillaume 2257 Sartoretti 2258 </div> 2259 2260 <!-- Journal/Book title and date --> 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 <div class="periodical"> 2272 <em>In International Symposium on Distributed Autonomous Robotic Systems (DARS)</em> , 2024 2273 </div> 2274 <div class="periodical"> 2275 2276 </div> 2277 2278 <!-- Links/Buttons --> 2279 <div class="links"> 2280 2281 2282 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2283 2284 2285 2286 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2287 2288 2289 2290 2291 <a href="https://arxiv.org/pdf/2407.20203" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2292 2293 2294 2295 2296 2297 2298 <a href="https://github.com/marmotlab/Bandwidth-Limited-Multi-Robot-Exploration" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2299 2300 2301 2302 2303 </div> 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 <!-- Hidden abstract block --> 2315 <div class="abstract hidden"> 2316 <p>Communication bandwidth is an important consideration in multi-robot exploration, where information exchange among robots is critical. While existing methods typically aim to reduce communication throughput, they either require significant computation or significantly compromise exploration efficie
2316ncy. In this work, we propose a deep reinforcement learning framework based on communication and privileged reinforcement learning to achieve a significant reduction in bandwidth consumption, while minimally sacrificing exploration efficiency. Specifically, our approach allows robots to learn to embed the most salient information from their individual belief (partial map) over the environment into fixed-sized messages. Robots then reason about their own belief as well as received messages to distributedly explore the environment while avoiding redundant work. In doing so, we employ privileged learning and learned attention mechanisms to endow the critic (i.e., teacher) network with ground truth map knowledge to effectively guide the policy (i.e., student) network during training. Compared to relevant baselines, our model allows the team to reduce communication by up to two orders of magnitude, while only sacrificing a marginal 2.4% in total travel distance, paving the way for efficient, distributed multi-robot exploration in bandwidth-limited scenarios.</p> 2317 </div> 2318 2319 2320 2321 <!-- Hidden bibtex block --> 2322 <div class="bibtex hidden"> 2323 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">ma2024privileged</span><span class="p">,</span> 2324 <span class="na">title</span> <span class="p">=</span> <span class="s">{Privileged Reinforcement and Communication Learning for Distributed, Bandwidth-limited Multi-robot Exploration}</span><span class="p">,</span> 2325 <span class="na">author</span> <span class="p">=</span> <span class="s">{Ma, Yixiao and Liang, Jingsong and Cao, Yuhong and Tan, Derek Ming Siang and Sartoretti, Guillaume}</span><span class="p">,</span> 2326 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{International Symposium on Distributed Autonomous Robotic Systems (DARS)}</span><span class="p">,</span> 2327 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span> 2328<span class="p">}</span></code></pre></figure> 2329 </div> 2330 2331 2332 2333 </div> 2334</div> 2335</li> 2336<li> 2337<div class="row"> 2338 2339 <div class="col col-sm-2 abbr"> 2340 2341 2342 2343 2344 2345<figure> 2346 <picture> 2347 <!-- Auto scaling with imagemagick --> 2348 <!-- 2349 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2350 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2351 --> 2352 2353 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/i2r-480.webp 480w,/assets/img/publication_preview/i2r-800.webp 800w,/assets/img/publication_preview/i2r-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2354 2355 <img src="/assets/img/publication_preview/i2r.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="i2r.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2356 </picture> 2357 2358 2359</figure> 2360 2361 2362 2363 </div> 2364 2365 2366 <!-- Entry bib key --> 2367 <div id="tan2024IR2" class="col-sm-8"> 2368 <!-- Title --> 2369 <div class="title">IR2: Implicit Rendezvous for Robotic Exploration Teams under Sparse Intermittent Connectivity</div> 2370 <!-- Author --> 2371 <div class="author"> 2372 2373 2374 2375 Derek Ming Siang 2376 Tan, Yixiao 2377 Ma, Jingsong 2378 Liang, Yi Cheng 2379 Chng, <em>Yuhong 2380 Cao</em>, and Guillaume 2381 Sartoretti 2382 </div> 2383 2384 <!-- Journal/Book title and date --> 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 <div class="periodical"> 2396 <em>In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)</em> , 2024 2397 </div> 2398 <div class="periodical"> 2399 2400 </div> 2401 2402 <!-- Links/Buttons --> 2403 <div class="links"> 2404 2405 2406 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2407 2408 2409 2410 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2411 2412 2413 2414 2415 <a href="https://arxiv.org/pdf/2409.04730" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2416 2417 2418 2419 2420 <a href="https://drive.google.com/file/d/1vh3jUCLaxYTk40VNZW1dhtokaL1AfrnX/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2421 2422 2423 2424 <a href="https://github.com/marmotlab/IR2-Multi-Robot-RL-Exploration" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2425 2426 2427 2428 2429 <a href="https://ir2-explore.github.io" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Website</a> 2430 2431 </div> 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 <!-- Hidden abstract block --> 2443 <div class="abstract hidden"> 2444 <p>Information sharing is critical in time-sensitive and realistic multi-robot exploration, especially for smaller robotic teams in large-scale environments where connectivity may be sparse and intermittent. Existing methods often overlook such communication constraints by assuming unrealistic global connectivity. Other works account for communication constraints (by maintaining close proximity or line of sight during information exchange), but are often inefficient. For instance, preplanned rendezvous approaches typically involve unnecessary detours resulting from poorly timed rendezvous, while pursuit-based approaches often result in short-sighted decisions due to their greedy
2444nature. We present IR2, a deep reinforcement learning approach to information sharing for multi-robot exploration. Leveraging attention-based neural networks trained via reinforcement and curriculum learning, IR2 allows robots to effectively reason about the longer-term trade-offs between disconnecting for solo exploration and reconnecting for information sharing. In addition, we propose a hierarchical graph formulation to maintain a sparse yet informative graph, enabling our approach to scale to large-scale environments. We present simulation results in three large-scale Gazebo environments, which show that our approach yields 6.6-34.1% shorter exploration paths and significantly improved mapped area consistency among robots when compared to state-of-the-art baselines.</p> 2445 </div> 2446 2447 2448 2449 <!-- Hidden bibtex block --> 2450 <div class="bibtex hidden"> 2451 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">tan2024IR2</span><span class="p">,</span> 2452 <span class="na">title</span> <span class="p">=</span> <span class="s">{IR2: Implicit Rendezvous for Robotic Exploration Teams under Sparse Intermittent Connectivity}</span><span class="p">,</span> 2453 <span class="na">author</span> <span class="p">=</span> <span class="s">{Tan, Derek Ming Siang and Ma, Yixiao and Liang, Jingsong and Chng, Yi Cheng and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 2454 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}</span><span class="p">,</span> 2455 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span> 2456<span class="p">}</span></code></pre></figure> 2457 </div> 2458 2459 2460 2461 </div> 2462</div> 2463</li> 2464<li> 2465<div class="row"> 2466 2467 <div class="col col-sm-2 abbr"> 2468 2469 2470 2471 2472 2473<figure> 2474 <picture> 2475 <!-- Auto scaling with imagemagick --> 2476 <!-- 2477 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2478 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2479 --> 2480 2481 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/star-480.webp 480w,/assets/img/publication_preview/star-800.webp 800w,/assets/img/publication_preview/star-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2482 2483 <img src="/assets/img/publication_preview/star.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="star.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2484 </picture> 2485 2486 2487</figure> 2488 2489 2490 2491 </div> 2492 2493 2494 <!-- Entry bib key --> 2495 <div id="chiun2024star" class="col-sm-8"> 2496 <!-- Title --> 2497 <div class="title">STAR: Swarm Technology for Aerial Robotics Research</div> 2498 <!-- Author --> 2499 <div class="author"> 2500 2501 2502 2503 Jimmy 2504 Chiun, Yan Rui 2505 Tan, <em>Yuhong 2506 Cao</em>, John 2507 Tan, and Guillaume 2508 Sartoretti 2509 </div> 2510 2511 <!-- Journal/Book title and date --> 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 <div class="periodical"> 2523 <em>The International Conference on Control, Automation, and Systems</em>, 2024 2524 </div> 2525 <div class="periodical"> 2526 2527 </div> 2528 2529 <!-- Links/Buttons --> 2530 <div class="links"> 2531 2532 2533 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2534 2535 2536 2537 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2538 2539 2540 2541 2542 <a href="https://arxiv.org/pdf/2406.16671" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2543 2544 2545 2546 2547 <a href="https://youtu.be/8NVm2VyeYkQ?feature=shared" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2548 2549 2550 2551 2552 2553 2554 </div> 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 <!-- Hidden abstract block --> 2566 <div class="abstract hidden"> 2567 <p>In recent years, the field of aerial robotics has witnessed significant progress, finding applications in diverse domains, including post-disaster search and rescue operations. Despite these strides, the prohibitive acquisition costs associated with deploying physical multi-UAV systems have posed challenges, impeding their widespread utilization in research endeavors. To overcome these challenges, we present STAR (Swarm Technology for Aerial Robotics Research), a framework developed explicitly to improve the accessibility of aerial swarm research experiments. Our framework introduces a swarm architecture based on the Crazyflie, a low-cost, open-source, palm-sized aerial platform, well suited for experimental swarm algorithms. To augment cost-effectiveness and mitigate the limitations of employing low-cost robots in experiments, we propose a landmark-based localization module leveraging fiducial markers. This module, also serving as a target detection module, enhances the adaptability and versatility of the framework. Additionally, collision and obstacle avoidance are implemented through velocity obstacles. The presented work strives to bridge the gap between theoretical advances and tangible implementations, thu
2567s fostering progress in the field.</p> 2568 </div> 2569 2570 2571 2572 <!-- Hidden bibtex block --> 2573 <div class="bibtex hidden"> 2574 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">chiun2024star</span><span class="p">,</span> 2575 <span class="na">title</span> <span class="p">=</span> <span class="s">{STAR: Swarm Technology for Aerial Robotics Research}</span><span class="p">,</span> 2576 <span class="na">author</span> <span class="p">=</span> <span class="s">{Chiun, Jimmy and Tan, Yan Rui and Cao, Yuhong and Tan, John and Sartoretti, Guillaume}</span><span class="p">,</span> 2577 <span class="na">journal</span> <span class="p">=</span> <span class="s">{The International Conference on Control, Automation, and Systems}</span><span class="p">,</span> 2578 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span> 2579<span class="p">}</span></code></pre></figure> 2580 </div> 2581 2582 2583 2584 </div> 2585</div> 2586</li> 2587<li> 2588<div class="row"> 2589 2590 <div class="col col-sm-2 abbr"> 2591 2592 2593 2594 2595 2596<figure> 2597 <picture> 2598 <!-- Auto scaling with imagemagick --> 2599 <!-- 2600 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2601 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2602 --> 2603 2604 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/drl_exploration-480.webp 480w,/assets/img/publication_preview/drl_exploration-800.webp 800w,/assets/img/publication_preview/drl_exploration-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2605 2606 <img src="/assets/img/publication_preview/drl_exploration.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="drl_exploration.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2607 </picture> 2608 2609 2610</figure> 2611 2612 2613 2614 </div> 2615 2616 2617 <!-- Entry bib key --> 2618 <div id="cao2024deep" class="col-sm-8"> 2619 <!-- Title --> 2620 <div class="title">Deep Reinforcement Learning-based Large-scale Robot Exploration</div> 2621 <!-- Author --> 2622 <div class="author"> 2623 2624 2625 2626 <em>Yuhong 2627 Cao</em>, Rui 2628 Zhao, Yizhuo 2629 Wang, Bairan 2630 Xiang, and Guillaume 2631 Sartoretti 2632 </div> 2633 2634 <!-- Journal/Book title and date --> 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 <div class="periodical"> 2646 <em>IEEE Robotics and Automation Letters</em>, 2024 2647 </div> 2648 <div class="periodical"> 2649 2650 </div> 2651 2652 <!-- Links/Buttons --> 2653 <div class="links"> 2654 2655 2656 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2657 2658 2659 2660 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2661 2662 2663 2664 <a href="https://ieeexplore.ieee.org/abstract/document/10476684" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 2665 2666 2667 2668 2669 <a href="https://arxiv.org/pdf/2403.10833" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2670 2671 2672 2673 2674 <a href="https://drive.google.com/file/d/1nyVJnGOX2kv8Wsm2hLWIDGGKkhXAI4Tu/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2675 2676 2677 2678 <a href="https://github.com/marmotlab/large-scale-DRL-exploration" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2679 2680 2681 2682 2683 </div> 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 <!-- Hidden abstract block --> 2695 <div class="abstract hidden"> 2696 <p>In this work, we propose a deep reinforcement learning (DRL) based reactive planner to solve large-scale Lidar-based autonomous robot exploration problems in 2D action space. Our DRL-based planner allows the agent to reactively plan its exploration path by making implicit predictions about unknown areas, based on a learned estimation of the underlying transition model of the environment. To this end, our approach relies on learned attention mechanisms for their powerful ability to capture long-term dependencies at different spatial scales to reason about the robotâs entire belief over known areas. Our approach relies on ground truth information (i.e., privileged learning) to guide the environment estimation during training, as well as on a graph rarefaction algorithm, which allows models trained in small-scale environments to scale to large-scale ones. Simulation results show that our model exhibits better exploration efficie
2696ncy (12% in path length, 6% in makespan) and lower planning time (60%) than the state-of-the-art planners in a 130mx100m benchmark scenario. We also validate our learned model on hardware.</p> 2697 </div> 2698 2699 2700 2701 <!-- Hidden bibtex block --> 2702 <div class="bibtex hidden"> 2703 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">cao2024deep</span><span class="p">,</span> 2704 <span class="na">title</span> <span class="p">=</span> <span class="s">{Deep Reinforcement Learning-based Large-scale Robot Exploration}</span><span class="p">,</span> 2705 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Zhao, Rui and Wang, Yizhuo and Xiang, Bairan and Sartoretti, Guillaume}</span><span class="p">,</span> 2706 <span class="na">journal</span> <span class="p">=</span> <span class="s">{IEEE Robotics and Automation Letters}</span><span class="p">,</span> 2707 <span class="na">year</span> <span class="p">=</span> <span class="s">{2024}</span><span class="p">,</span> 2708 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{IEEE}</span> 2709<span class="p">}</span></code></pre></figure> 2710 </div> 2711 2712 2713 2714 </div> 2715</div> 2716</li> 2717</ol> 2718<h2 class="bibliography">2023</h2> 2719<ol class="bibliography"> 2720<li> 2721<div class="row"> 2722 2723 <div class="col col-sm-2 abbr"> 2724 2725 2726 2727 2728 2729<figure> 2730 <picture> 2731 <!-- Auto scaling with imagemagick --> 2732 <!-- 2733 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2734 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2735 --> 2736 2737 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/mapless_nav-480.webp 480w,/assets/img/publication_preview/mapless_nav-800.webp 800w,/assets/img/publication_preview/mapless_nav-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2738 2739 <img src="/assets/img/publication_preview/mapless_nav.png" class="preview z-depth-1 rounded" width="100%" height="auto" alt="mapless_nav.png" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2740 </picture> 2741 2742 2743</figure> 2744 2745 2746 2747 </div> 2748 2749 2750 <!-- Entry bib key --> 2751 <div id="liang2023context" class="col-sm-8"> 2752 <!-- Title --> 2753 <div class="title">Context-Aware Deep Reinforcement Learning for Autonomous Robotic Navigation in Unknown Area</div> 2754 <!-- Author --> 2755 <div class="author"> 2756 2757 2758 2759 Jingsong 2760 Liang*, Zhichen 2761 Wang*, <em>Yuhong 2762 Cao*</em>, Jimmy 2763 Chiun, Mengqi 2764 Zhang, and Guillaume Adrien 2765 Sartoretti 2766 </div> 2767 2768 <!-- Journal/Book title and date --> 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 <div class="periodical"> 2780 <em>In Conference on Robot Learning</em> , 2023 2781 </div> 2782 <div class="periodical"> 2783 2784 </div> 2785 2786 <!-- Links/Buttons --> 2787 <div class="links"> 2788 2789 2790 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2791 2792 2793 2794 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2795 2796 2797 2798 <a href="https://proceedings.mlr.press/v229/liang23a.html" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 2799 2800 2801 2802 2803 <a href="https://proceedings.mlr.press/v229/liang23a/liang23a.pdf" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2804 2805 2806 2807 2808 <a href="https://drive.google.com/file/d/1r-5TXA6zOAJUfI9E8I_48SoT3WXgRSzu/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2809 2810 2811 2812 <a href="https://github.com/marmotlab/Context_Aware_Navigation" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2813 2814 2815 2816 2817 </div> 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 <!-- Hidden abstract block --> 2829 <div class="abstract hidden"> 2830 <p>Mapless navigation refers to a challenging task where a mobile robot must rapidly navigate to a predefined destination using its partial knowledge of the environment, which is updated online along the way, instead of a prior map of the environment. Inspired by the recent developments in deep reinforcement learning (DRL), we propose a learning-based framework for mapless navigation, which employs a context-aware policy network to achieve efficient decision-making (ie, maximize the likelihood of finding the shortest route towards the target destination), especially in complex and large-scale environments. Specifically, our robot learns to form a context of its belief over the entire known area, which it uses to reason about long-term efficiency and sequence show-term movements. Additionally, we propose a graph rarefaction algorithm to enable more efficient decision-making in large-scale applications. We empirically demonstrate that our approach reduces average travel time by up to 61.4% and average planning time by up to 88.2% compared to benchmark planners (D* lite and BIT) on hundreds of test scenarios. We also validate our approach both in high-fidelity Gazebo simulations as well as on hardware, highlighting its promising applicability in the real world without further training/tuning.</p> 2831 </div> 2832 2833 2834 2835 <!-- Hidden bibtex block --> 2836 <div class="bibtex hidden"> 2837 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">liang2023context</span><span class="p">,</span> 2838 <span class="na">title</span> <span class="p">=</span> <span class="s">{Context-Aware Deep Reinforcement Learning for Autonomous Robotic Navigation in Unknown Area}</span><span class="p">,</span> 2839 <span class="na">author</span> <span class="p">=</span> <span class="s">{Liang*, Jingsong and Wang*, Zhichen and Cao*, Yuhong and Chiun, Jimmy and Zhang, Mengqi and Sartoretti, Guillaume Adrien}</span><span class="p">,</span> 2840 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Conference on Robot Learning}</span><span class="p">,</span> 2841 <span class="na">pages</span> <span class="p">=</span> <span class="s">{1425--1436}</span><span class="p">,</span> 2842 <span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span> 2843 <span class="na">organization</span> <span class="p">=</span> <span class="s">{PMLR}</span> 2844<span class="p">}</span></code></pre></figure> 2845 </div> 2846 2847 2848 2849 </div> 2850</div> 2851</li> 2852<li> 2853<div class="row"> 2854 2855 <div class="col col-sm-2 abbr"> 2856 2857 2858 2859 2860 2861<figure> 2862 <picture> 2863 <!-- Auto scaling with imagemagick --> 2864 <!-- 2865 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2866 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2867 --> 2868 2869 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/maipp-480.webp 480w,/assets/img/publication_preview/maipp-800.webp 800w,/assets/img/publication_preview/maipp-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2870 2871 <img src="/assets/img/publication_preview/maipp.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="maipp.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 2872 </picture> 2873 2874 2875</figure> 2876 2877 2878 2879 </div> 2880 2881 2882 <!-- Entry bib key -->
2883 <div id="yang2023intent" class="col-sm-8"> 2884 <!-- Title --> 2885 <div class="title">Intent-based Deep Reinforcement Learning for Multi-agent Informative Path Planning</div> 2886 <!-- Author --> 2887 <div class="author"> 2888 2889 2890 2891 Tianze 2892 Yang, <em>Yuhong 2893 Cao</em>, and Guillaume 2894 Sartoretti 2895 </div> 2896 2897 <!-- Journal/Book title and date --> 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 <div class="periodical"> 2909 <em>In 2023 International Symposium on Multi-Robot and Multi-Agent Systems (MRS)</em> , 2023 2910 </div> 2911 <div class="periodical"> 2912 2913 </div> 2914 2915 <!-- Links/Buttons --> 2916 <div class="links"> 2917 2918 2919 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 2920 2921 2922 2923 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 2924 2925 2926 2927 <a href="https://ieeexplore.ieee.org/abstract/document/10416797" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 2928 2929 2930 2931 2932 <a href="https://arxiv.org/pdf/2303.05351" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 2933 2934 2935 2936 2937 <a href="https://drive.google.com/file/d/1pUg6oZQGTELy4IIGfedvMd04r85GSme6/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 2938 2939 2940 2941 <a href="https://github.com/marmotlab/MAIPP/tree/trajectory_intent" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 2942 2943 2944 2945 2946 </div> 2947 2948 2949 2950 2951 2952 2953 2954 2955 2956 2957 <!-- Hidden abstract block --> 2958 <div class="abstract hidden"> 2959 <p>In multi-agent informative path planning (MAIPP), agents must collectively construct a global belief map of an underlying distribution of interest (e.g., gas concentration, light intensity, or pollution levels) over a given domain, based on measurements taken along their trajectory. They must frequently replan their path to balance the distributed exploration of new areas and the collective, meticulous exploitation of known high-interest areas, to maximize the information gained within a predefined budget (e.g., path length or working time). A common approach to achieving such cooperation relies on planning the agentsâ paths reactively, conditioned on other agentsâ future actions. However, as the agentâs belief is updated continuously, these predicted future actions may not end up being the ones executed by agents, introducing a form of noise/inaccuracy in the system and often decreasing performance. In this work, we propose a decentralized deep reinforcement learning (DRL) approach to MAIPP, which relies on an attention-based neural network, where agents optimize long-term individual and cooperative objectives by explicitly sharing their intent (i.e., medium-/long-term future positions distribution, obtained from their individual policy) in a reactive, asynchronous manner. That is, in our work, intent sharing allows agents to learn to claim/avoid broader areas of the world. Moreover, since our approach relies on learned attention over these shared intents, agents are able to learn to recognize the useful portion(s) of these (imperfect) predictions to maximize cooperation even in the presence of imperfect information. Our comparison experiments demonstrate the performance of our approach compared to its variants and high-quality baselines over a large set of MAIPP simulations.</p> 2960 </div> 2961 2962 2963 2964 <!-- Hidden bibtex block --> 2965 <div class="bibtex hidden"> 2966 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">yang2023intent</span><span class="p">,</span> 2967 <span class="na">title</span> <span class="p">=</span> <span class="s">{Intent-based Deep Reinforcement Learning for Multi-agent Informative Path Planning}</span><span class="p">,</span> 2968 <span class="na">author</span> <span class="p">=</span> <span class="s">{Yang, Tianze and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 2969 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{2023 International Symposium on Multi-Robot and Multi-Agent Systems (MRS)}</span><span class="p">,</span> 2970 <span class="na">pages</span> <span class="p">=</span> <span class="s">{71--77}</span><span class="p">,</span> 2971 <span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span> 2972 <span class="na">organization</span> <span class="p">=</span> <span class="s">{IEEE}</span> 2973<span class="p">}</span></code></pre></figure> 2974 </div> 2975 2976 2977 2978 </div> 2979</div> 2980</li> 2981<li> 2982<div class="row"> 2983 2984 <div class="col col-sm-2 abbr"> 2985 2986 2987 2988 2989 2990<figure> 2991 <picture> 2992 <!-- Auto scaling with imagemagick --> 2993 <!-- 2994 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 2995 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 2996 --> 2997 2998 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/stamp-480.webp 480w,/assets/img/publication_preview/stamp-800.webp 800w,/assets/img/publication_preview/stamp-1400.webp 1400w," sizes="200px" type="image/webp"></source> 2999 3000 <img src="/assets/img/publication_preview/stamp.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="stamp.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 3001 </picture> 3002 3003 3004</figure> 3005 3006 3007 3008 </div> 3009 3010 3011 <!-- Entry bib key --> 3012 <div id="wang2023spatio" class="col-sm-8"> 3013 <!-- Title --> 3014 <div class="title">Spatio-temporal attention network for persistent monitoring of multiple mobile targets</div> 3015 <!-- Author --> 3016 <div class="author"> 3017 3018 3019 3020 Yizhuo 3021 Wang, Yutong 3022 Wang, <em>Yuhong 3023 Cao</em>, and Guillaume 3024 Sartoretti 3025 </div> 3026 3027 <!-- Journal/Book title and date --> 3028 3029 3030 3031 3032 3033 3034 3035 3036 3037 3038 <div class="periodical"> 3039 <em>In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (I
3039ROS)</em> , 2023 3040 </div> 3041 <div class="periodical"> 3042 3043 </div> 3044 3045 <!-- Links/Buttons --> 3046 <div class="links"> 3047 3048 3049 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 3050 3051 3052 3053 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 3054 3055 3056 3057 <a href="https://ieeexplore.ieee.org/abstract/document/10341674" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 3058 3059 3060 3061 3062 <a href="https://arxiv.org/pdf/2303.06350" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 3063 3064 3065 3066 3067 <a href="https://drive.google.com/file/d/1OiOg9guiJnos-IaYwK5cLVLadCmx741T/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 3068 3069 3070 3071 <a href="https://github.com/marmotlab/STAMP" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 3072 3073 3074 3075 3076 </div> 3077 3078 3079 3080 3081 3082 3083 3084 3085 3086 3087 <!-- Hidden abstract block --> 3088 <div class="abstract hidden"> 3089 <p>This work focuses on the persistent monitoring problem, where a set of targets moving based on an unknown model must be monitored by an autonomous mobile robot with a limited sensing range. To keep each targetâs position estimate as accurate as possible, the robot needs to adaptively plan its path to (re-)visit all the targets and update its belief from measurements collected along the way. In doing so, the main challenge is to strike a balance between exploitation, i.e., re-visiting previously-located targets, and exploration, i.e., finding new targets or re-acquiring lost ones. 3090Encouraged by recent advances in deep reinforcement learning, we introduce an attention-based neural solution to the persistent monitoring problem, where the agent can learn the inter-dependencies between targets, i.e., their spatial and temporal correlations, conditioned on past measurements. 3091This endows the agent with the ability to determine which target, time, and location to attend to across multiple scales, which we show also helps relax the usual limitations of a finite target set with prior positional information. 3092We experimentally demonstrate that our method outperforms other baselines in terms of number of targets visits and average estimation error in complex environments. 3093Finally, we implement and validate our model in a drone-based simulation experiment to monitor mobile ground targets in a high-fidelity simulator.</p> 3094 </div> 3095 3096 3097 3098 <!-- Hidden bibtex block --> 3099 <div class="bibtex hidden"> 3100 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">wang2023spatio</span><span class="p">,</span> 3101 <span class="na">title</span> <span class="p">=</span> <span class="s">{Spatio-temporal attention network for persistent monitoring of multiple mobile targets}</span><span class="p">,</span> 3102 <span class="na">author</span> <span class="p">=</span> <span class="s">{Wang, Yizhuo and Wang, Yutong and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 3103 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)}</span><span class="p">,</span> 3104 <span class="na">pages</span> <span class="p">=</span> <span class="s">{3903--3910}</span><span class="p">,</span> 3105 <span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span> 3106 <span class="na">organization</span> <span class="p">=</span> <span class="s">{IEEE}</span> 3107<span class="p">}</span></code></pre></figure> 3108 </div> 3109 3110 3111 3112 </div> 3113</div> 3114</li> 3115<li> 3116<div class="row"> 3117 3118 <div class="col col-sm-2 abbr"> 3119 3120 3121 3122 3123 3124<figure> 3125 <picture> 3126 <!-- Auto scaling with imagemagick --> 3127 <!-- 3128 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 3129 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 3130 --> 3131 3132 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/ariadne-480.webp 480w,/assets/img/publication_preview/ariadne-800.webp 800w,/assets/img/publication_preview/ariadne-1400.webp 1400w," sizes="200px" type="image/webp"></source> 3133 3134 <img src="/assets/img/publication_preview/ariadne.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="ariadne.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 3135 </picture> 3136 3137 3138</figure> 3139 3140 3141 3142 </div> 3143 3144 3145 <!-- Entry bib key --> 3146 <div id="cao2023ariadne" class="col-sm-8"> 3147 <!-- Title --> 3148 <div class="title">
3148Ariadne: A reinforcement learning approach using attention-based deep networks for exploration</div> 3149 <!-- Author --> 3150 <div class="author"> 3151 3152 3153 3154 <em>Yuhong 3155 Cao</em>, Tianxiang 3156 Hou, Yizhuo 3157 Wang, Xian 3158 Yi, and Guillaume 3159 Sartoretti 3160 </div> 3161 3162 <!-- Journal/Book title and date --> 3163 3164 3165 3166 3167 3168 3169 3170 3171 3172 3173 <div class="periodical"> 3174 <em>In 2023 IEEE International Conference on Robotics and Automation (ICRA)</em> , 2023 3175 </div> 3176 <div class="periodical"> 3177 3178 </div> 3179 3180 <!-- Links/Buttons --> 3181 <div class="links"> 3182 3183 3184 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 3185 3186 3187 3188 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 3189 3190 3191 3192 <a href="https://ieeexplore.ieee.org/abstract/document/10160565" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 3193 3194 3195 3196 3197 <a href="https://arxiv.org/pdf/2301.11575" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 3198 3199 3200 3201 3202 <a href="https://drive.google.com/file/d/1h0p8d9Ry96FJGYI0mK1aNYicDCS7UPzs/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 3203 3204 3205 3206 <a href="https://github.com/marmotlab/ARiADNE" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 3207 3208 3209 3210 3211 </div> 3212 3213 3214 3215 3216 3217 3218 3219 3220 3221 3222 <!-- Hidden abstract block --> 3223 <div class="abstract hidden"> 3224 <p>In autonomous robot exploration tasks, a mobile robot needs to actively explore and map an unknown environment as fast as possible. Since the environment is being revealed during exploration, the robot needs to frequently re-plan its path online, as new information is acquired by onboard sensors and used to update its partial map. While state-of-the-art exploration planners are frontier- and sampling-based, encouraged by the recent development in deep reinforcement learning (DRL), we propose ARiADNE, an attention-based neural approach to obtain real-time, non-myopic path planning for autonomous exploration. ARiADNE is able to learn dependencies at multiple spatial scales between areas of the agentâs partial map, and implicitly predict potential gains associated with exploring those areas. This allows the agent to sequence movement actions that balance the natural trade-off between exploitation/refinement of the map in known areas and exploration of new areas. We experimentally demonstrate that our method outperforms both learning and non-learning state-of-the-art baselines in terms of average trajectory length to complete exploration in hundreds of simplified 2D indoor scenarios. We further validate our approach in high-fidelity Robot Operating System (ROS) simulations, where we consider a real sensor model and a realistic low-level motion controller, toward deployment on real robots.</p> 3225 </div> 3226 3227 3228 3229 <!-- Hidden bibtex block --> 3230 <div class="bibtex hidden"> 3231 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">cao2023ariadne</span><span class="p">,</span> 3232 <span class="na">title</span> <span class="p">=</span> <span class="s">{Ariadne: A reinforcement learning approach using attention-based deep networks for exploration}</span><span class="p">,</span> 3233 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Hou, Tianxiang and Wang, Yizhuo and Yi, Xian and Sartoretti, Guillaume}</span><span class="p">,</span> 3234 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{2023 IEEE International Conference on Robotics and Automation (ICRA)}</span><span class="p">,</span> 3235 <span class="na">pages</span> <span class="p">=</span> <span class="s">{10219--10225}</span><span class="p">,</span> 3236 <span class="na">year</span> <span class="p">=</span> <span class="s">{2023}</span><span class="p">,</span> 3237 <span class="na">organization</span> <span class="p">=</span> <span class="s">{IEEE}</span> 3238<span class="p">}</span></code></pre></figure> 3239 </div> 3240 3241 3242 3243 </div> 3244</div> 3245</li> 3246</ol> 3247<h2 class="bibliography">2022</h2> 3248<ol class="bibliography"> 3249<li> 3250<div class="row"> 3251 3252 <div class="col col-sm-2 abbr"> 3253 3254 3255 3256 3257 3258<figure> 3259 <picture> 3260 <!-- Auto scaling with imagemagick --> 3261 <!-- 3262 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 3263 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 3264 --> 3265 3266 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/catnipp-480.webp 480w,/assets/img/publication_preview/catnipp-800.webp 800w,/assets/img/publication_preview/catnipp-1400.webp 1400w," sizes="200px" type="image/webp"></source> 3267 3268 <img src="/assets/img/publication_preview/catnipp.gif" class="preview z-depth-1 rounded" width="100%" height="auto" alt="catnipp.gif" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 3269 </picture> 3270 3271 3272</figure> 3273 3274 3275 3276 </div> 3277 3278 3279 <!-- Entry bib key --> 3280 <div id="cao2023catnipp" class="col-sm-8"> 3281 <!-- Title --> 3282 <div class="title">CAtNIPP: Context-aware attention-based network for informative path planning</
3282div> 3283 <!-- Author --> 3284 <div class="author"> 3285 3286 3287 3288 <em>Yuhong 3289 Cao</em>, Yizhuo 3290 Wang, Apoorva 3291 Vashisth, Haolin 3292 Fan, and Guillaume Adrien 3293 Sartoretti 3294 </div> 3295 3296 <!-- Journal/Book title and date --> 3297 3298 3299 3300 3301 3302 3303 3304 3305 3306 3307 <div class="periodical"> 3308 <em>In Conference on Robot Learning</em> , 2022 3309 </div> 3310 <div class="periodical"> 3311 3312 </div> 3313 3314 <!-- Links/Buttons --> 3315 <div class="links"> 3316 3317 3318 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 3319 3320 3321 3322 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 3323 3324 3325 3326 <a href="https://proceedings.mlr.press/v205/cao23b.html" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 3327 3328 3329 3330 3331 <a href="https://proceedings.mlr.press/v205/cao23b/cao23b.pdf" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 3332 3333 3334 3335 3336 <a href="https://drive.google.com/file/d/1yo5TrJCzEUE65jNeEgIOP_ON7uTVZtPv/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 3337 3338 3339 3340 <a href="https://github.com/marmotlab/CAtNIPP" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 3341 3342 3343 3344 3345 </div> 3346 3347 3348 3349 3350 3351 3352 3353 3354 3355 3356 <!-- Hidden abstract block --> 3357 <div class="abstract hidden"> 3358 <p>Informative path planning (IPP) is an NP-hard problem, which aims at planning a path allowing an agent to build an accurate belief about a quantity of interest throughout a given search domain, within constraints on resource budget (e.g., path length for robots with limited battery life). IPP requires frequent online replanning as this belief is updated with every new measurement (i.e., adaptive IPP), while balancing short-term exploitation and longer-term exploration to avoid suboptimal, myopic behaviors. Encouraged by the recent developments in deep reinforcement learning, we introduce CAtNIPP, a fully reactive, neural approach to the adaptive IPP problem. CAtNIPP relies on self-attention for its powerful ability to capture 3359dependencies in data at multiple spatial scales. Specifically, our agent learns to form a context of its belief over the entire domain, which it uses to sequence local movement decisions 3360that optimize short- and longer-term search objectives. We experimentally demonstrate that CAtNIPP significantly outperforms state-of-the-art non-learning IPP solvers in terms of solution quality and computing time once trained, and present experimental results on hardware.</p> 3361 </div> 3362 3363 3364 3365 <!-- Hidden bibtex block --> 3366 <div class="bibtex hidden"> 3367 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">cao2023catnipp</span><span class="p">,</span> 3368 <span class="na">title</span> <span class="p">=</span> <span class="s">{CAtNIPP: Context-aware attention-based network for informative path planning}</span><span class="p">,</span> 3369 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Wang, Yizhuo and Vashisth, Apoorva and Fan, Haolin and Sartoretti, Guillaume Adrien}</span><span class="p">,</span> 3370 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{Conference on Robot Learning}</span><span class="p">,</span> 3371 <span class="na">pages</span> <span class="p">=</span> <span class="s">{1928--1937}</span><span class="p">,</span> 3372 <span class="na">year</span> <span class="p">=</span> <span class="s">{2022}</span><span class="p">,</span> 3373 <span class="na">organization</span> <span class="p">=</span> <span class="s">{PMLR}</span> 3374<span class="p">}</span></code></pre></figure> 3375 </div> 3376 3377 3378 3379 </div> 3380</div> 3381</li> 3382<li> 3383<div class="row"> 3384 3385 <div class="col col-sm-2 abbr"> 3386 3387 3388 3389 3390 3391<figure> 3392 <picture> 3393 <!-- Auto scaling with imagemagick --> 3394 <!-- 3395 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 3396 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 3397 --> 3398 3399 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/distributed_learning_review-480.webp 480w,/assets/img/publication_preview/distributed_learning_review-800.webp 800w,/assets/img/publication_preview/distributed_learning_review-1400.webp 1400w," sizes="200px" type="image/webp"></source> 3400 3401 <img src="/assets/img/publication_preview/distributed_learning_review.jpg" class="preview z-depth-1 rounded" width="100%" height="auto" alt="distributed_learning_review.jpg" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 3402 </picture> 3403 3404 3405</figure> 3406 3407 3408 3409 </div> 3410 3411 3412 <!-- Entry bib key --> 3413 <div id="wang2022distributed" class="col-sm-8"> 3414 <!-- Title --> 3415 <div class="title">Distributed reinforcement learning for robot teams: A review</div> 3416 <!-- Author --> 3417 <div class="author"> 3418 3419 3420 3421 Yutong 3422 Wang, Mehul 3423 Damani, Pamela 3424 Wang, <em>Yuhong 3425 Cao</em>, and Guillaume 3426 Sartoretti 3427 </div> 3428 3429 <!-- Journal/Book title and date --> 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 <div class="periodical"> 3441 <em>Current Robotics Reports</em>, 2022 3442 </div> 3443 <div class="periodical"> 3444 3445 </div> 3446 3447 <!-- Links/Buttons --> 3448 <div class="links"> 3449 3450 3451 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 3452 3453 3454 3455 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 3456 3457 3458 3459 <a href="https://link.springer.com/article/10.1007/s43154-022-00091-8" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 3460 3461 3462 3463 3464 <a href="https://arxiv.org/pdf/2204.03516" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 3465 3466 3467 3468 3469 3470 3471 3472 3473 3474 </div> 3475 3476 3477 3478 3479 3480 3481 3482 3483 3484 3485 <!-- Hidden abstract block --> 3486 <div class="abstract hidden"> 3487 <p>Recent advances in sensing, actuation, and computation have opened the 3488door to multi-robot systems consisting of hundreds/thousands of robots, with promising applications 3489to automated manufacturing, disaster relief, harvesting, last-mile delivery, port/airport operations, 3490or search and rescue. The community has leveraged model-free multi-agent reinfor
3490cement learning 3491(MARL) to devise efficient, scalable controllers for multi-robot systems (MRS). This review aims 3492to provide an analysis of the state-of-the-art in distributed MARL for multi-robot cooperation. Decentralized MRS face fundamental challenges, such as non-stationarity and 3493partial observability. Building upon the âcentralized training, decentralized executionâ paradigm, 3494recent MARL approaches include independent learning, centralized critic, value decomposition, and 3495communication learning approaches. Cooperative behaviors are demonstrated through AI benchmarks and fundamental real-world robotic capabilities such as multi-robot motion/path planning. This survey reports the challenges surrounding decentralized model-free MARL 3496for multi-robot cooperation and existing classes of approaches. We present benchmarks 3497and robotic applications along with a discussion on current open avenues for research.</p> 3498 </div> 3499 3500 3501 3502 <!-- Hidden bibtex block --> 3503 <div class="bibtex hidden"> 3504 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@article</span><span class="p">{</span><span class="nl">wang2022distributed</span><span class="p">,</span> 3505 <span class="na">title</span> <span class="p">=</span> <span class="s">{Distributed reinforcement learning for robot teams: A review}</span><span class="p">,</span> 3506 <span class="na">author</span> <span class="p">=</span> <span class="s">{Wang, Yutong and Damani, Mehul and Wang, Pamela and Cao, Yuhong and Sartoretti, Guillaume}</span><span class="p">,</span> 3507 <span class="na">journal</span> <span class="p">=</span> <span class="s">{Current Robotics Reports}</span><span class="p">,</span> 3508 <span class="na">volume</span> <span class="p">=</span> <span class="s">{3}</span><span class="p">,</span> 3509 <span class="na">number</span> <span class="p">=</span> <span class="s">{4}</span><span class="p">,</span> 3510 <span class="na">pages</span> <span class="p">=</span> <span class="s">{239--257}</span><span class="p">,</span> 3511 <span class="na">year</span> <span class="p">=</span> <span class="s">{2022}</span><span class="p">,</span> 3512 <span class="na">publisher</span> <span class="p">=</span> <span class="s">{Springer}</span> 3513<span class="p">}</span></code></pre></figure> 3514 </div> 3515 3516 3517 3518 </div> 3519</div> 3520</li> 3521<li> 3522<div class="row"> 3523 3524 <div class="col col-sm-2 abbr"> 3525 3526 3527 3528 3529 3530<figure> 3531 <picture> 3532 <!-- Auto scaling with imagemagick --> 3533 <!-- 3534 See https://www.debugbear.com/blog/responsive-images#w-descriptors-and-the-sizes-attribute and 3535 https://developer.mozilla.org/en-US/docs/Learn/HTML/Multimedia_and_embedding/Responsive_images for info on defining 'sizes' for responsive images 3536 --> 3537 3538 <source class="responsive-img-srcset" srcset="/assets/img/publication_preview/DAN-480.webp 480w,/assets/img/publication_preview/DAN-800.webp 800w,/assets/img/publication_preview/DAN-1400.webp 1400w," sizes="200px" type="image/webp"></source> 3539 3540 <img src="/assets/img/publication_preview/DAN.jpg" class="preview z-depth-1 rounded" width="100%" height="auto" alt="DAN.jpg" data-zoomable="" loading="eager" onerror="this.onerror=null; $('.responsive-img-srcset').remove();"> 3541 </picture> 3542 3543 3544</figure> 3545 3546 3547 3548 </div> 3549 3550 3551 <!-- Entry bib key --> 3552 <div id="cao2022dan" class="col-sm-8"> 3553 <!-- Title --> 3554 <div class="title">Dan: Decentralized attention-based neural network for the minmax multiple traveling salesman problem</div> 3555 <!-- Author --> 3556 <div class="author"> 3557 3558 3559 3560 <em>Yuhong 3561 Cao</em>, Zhanhong 3562 Sun, and Guillaume 3563 Sartoretti 3564 </div> 3565 3566 <!-- Journal/Book title and date --> 3567 3568 3569 3570 3571 3572 3573 3574 3575 3576 3577 <div class="periodical"> 3578 <em>In International Symposium on Distributed Autonomous Robotic Systems</em> , 2022 3579 </div> 3580 <div class="periodical"> 3581 3582 </div> 3583 3584 <!-- Links/Buttons --> 3585 <div class="links"> 3586 3587 <a class="award btn btn-sm z-depth-0" role="button">Awareded</a> 3588 3589 3590 <a class="abstract btn btn-sm z-depth-0" role="button">Abs</a> 3591 3592 3593 3594 <a class="bibtex btn btn-sm z-depth-0" role="button">Bib</a> 3595 3596 3597 3598 <a href="https://link.springer.com/chapter/10.1007/978-3-031-51497-5_15" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">HTML</a> 3599 3600 3601 3602 3603 <a href="https://arxiv.org/pdf/2109.04205" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">PDF</a> 3604 3605 3606 3607 3608 <a href="https://drive.google.com/file/d/14oUQ-Ap5SYWLJPfWagcbmvhejsfU3JwA/view?usp=sharing" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Video</a> 3609 3610 3611 3612 <a href="https://github.com/marmotlab/DAN" class="btn btn-sm z-depth-0" role="button" rel="external nofollow noopener" target="_blank">Code</a> 3613 3614 3615 3616 3617 </div> 3618 3619 3620 3621 3622 3623 3624 3625 3626 <!-- Hidden Award block --> 3627 <div class="award hidden d-print-inline"> 3628 <p></p> 3629<p>Best student paper</p> 3630 3631 </div> 3632 3633 3634 3635 <!-- Hidden abstract block --> 3636 <div class="abstract hidden"> 3637 <p>The multiple traveling salesman problem (mTSP) is a well-known NP-hard problem with numerous real-world applications. In particular, this work addresses MinMax mTSP, where the objective is to minimize the max tour length among all agents. Many robotic deployments require recomputing potentially large mTSP instances frequently, making the natural trade-off between computing time and solution quality of great importance. However, exact and heuristic algorithms become inefficient as the number of cities increases, due to their computational complexity. Encouraged by the recent developments in deep reinforcement learning (dRL), this work approaches the mTSP as a cooperative task and introduces DAN, a decentralized attention-based neural method that aims at tackling this key trade-off. In DAN, agents learn fully decentralized policies to collaboratively construct a tour, by predicting each otherâs future decisions. Our model relies on attention mechanisms and is trained using multi-agent RL with parameter sharing, providing natural scalability to the numbers of agents and cities. Our experimental results on
3637small- to large-scale mTSP instances (50 to 1000 cities and 5 to 20 agents) show that DAN is able to match or outperform state-of-the-art solvers while keeping planning times low. In particular, given the same computation time budget, DAN outperforms all conventional and dRL-based baselines on larger-scale instances (more than 100 cities, more than 5 agents), and exhibits enhanced agent collaboration.</p> 3638 </div> 3639 3640 3641 3642 <!-- Hidden bibtex block --> 3643 <div class="bibtex hidden"> 3644 <figure class="highlight"><pre><code class="language-bibtex" data-lang="bibtex"><span class="nc">@inproceedings</span><span class="p">{</span><span class="nl">cao2022dan</span><span class="p">,</span> 3645 <span class="na">title</span> <span class="p">=</span> <span class="s">{Dan: Decentralized attention-based neural network for the minmax multiple traveling salesman problem}</span><span class="p">,</span> 3646 <span class="na">author</span> <span class="p">=</span> <span class="s">{Cao, Yuhong and Sun, Zhanhong and Sartoretti, Guillaume}</span><span class="p">,</span> 3647 <span class="na">booktitle</span> <span class="p">=</span> <span class="s">{International Symposium on Distributed Autonomous Robotic Systems}</span><span class="p">,</span> 3648 <span class="na">pages</span> <span class="p">=</span> <span class="s">{202--215}</span><span class="p">,</span> 3649 <span class="na">year</span> <span class="p">=</span> <span class="s">{2022}</span><span class="p">,</span> 3650 <span class="na">organization</span> <span class="p">=</span> <span class="s">{Springer}</span> 3651<span class="p">}</span></code></pre></figure> 3652 </div> 3653 3654 3655 3656 </div> 3657</div> 3658</li> 3659</ol> 3660 3661</div> 3662 3663 </article> 3664 3665 3666 3667 3668</div> 3669 3670 3671 </div> 3672 3673 <!-- Footer --> 3674 3675 <footer class="fixed-bottom" role="contentinfo"> 3676 <div class="container mt-0"> 3677 © Copyright 2026 3678 Yuhong 3679 3680 Cao. Powered by <a href="https://jekyllrb.com/" target="_blank" rel="external nofollow noopener">Jekyll</a> with <a href="https://github.com/alshedivat/al-folio" rel="external nofollow noopener" target="_blank">al-folio</a> theme. 3681 3682 3683 3684 Last updated: June 20, 2026. 3685 3686 </div> 3687 </footer> 3688 3689 3690 3691 <!-- JavaScripts --> 3692 <!-- jQuery -->
3693<script src="https://cdn.jsdelivr.net/npm/[email protected]/dist/jquery.min.js" integrity="sha256-/xUj+3OJU5yExlq6GSYGSHk7tPXikynS7ogEvDej/m4=" crossorigin="anonymous"></script>
3693 3694 3695 <!-- Bootsrap & MDB scripts -->
3696<script src="/assets/js/bootstrap.bundle.min.js"></script>
3696 3697<!--
3697<script src="/assets/js/mdb.min.js"></script>
3697 -->
3698<script src="https://cdn.jsdelivr.net/npm/[email protected]/js/mdb.min.js" integrity="sha256-NdbiivsvWt7VYCt6hYNT3h/th9vSTL4EDWeGs5SN3DA=" crossorigin="anonymous"></script>
3698 3699 3700 3701 <!-- Masonry & imagesLoaded --> 3702
3702<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/masonry.pkgd.min.js" integrity="sha256-Nn1q/fx0H7SNLZMQ5Hw5JLaTRZp0yILA/FRexe19VdI=" crossorigin="anonymous"></script>
3702 3703
3703<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/imagesloaded.pkgd.min.js" integrity="sha256-htrLFfZJ6v5udOG+3kNLINIKh2gvoKqwEhHYfTTMICc=" crossorigin="anonymous"></script>
3703 3704
3704<script defer src="/assets/js/masonry.js" type="text/javascript"></script>
3704 3705 3706 3707 3708 3709 3710 3711 3712 3713 3714 3715 3716 3717 3718 3719 3720 3721 3722 3723 3724 3725 <!-- Medium Zoom JS --> 3726
3726<script defer src="https://cdn.jsdelivr.net/npm/[email protected]/dist/medium-zoom.min.js" integrity="sha256-ZgMyDAIYDYGxbcpJcfUnYwNevG/xi9OHKaR/8GK+jWc=" crossorigin="anonymous"></script>
3726 3727
3727<script defer src="/assets/js/zoom.js?85ddb88934d28b74e78031fd54cf8308"></script>
3727 3728 3729 3730 3731<!-- Bootstrap Table --> 3732 3733 3734<!-- Load Common JS -->
3735<script src="/assets/js/no_defer.js?2781658a0a2b13ed609542042a859126"></script>
vendor: 1 bytes, line 3735
3735
3736<script defer src="/assets/js/common.js?06cae41083477f121be8cd9797ad8e2f"></script>
vendor: 1 bytes, line 3736
3736
3737<script defer src="/assets/js/copy_code.js?12775fdf7f95e901d7119054556e495f" type="text/javascript"></script>
3737 3738 3739<!-- Jupyter Open External Links New Tab -->
3740<script defer src="/assets/js/jupyter_new_tab.js?d9f17b6adc2311cbabd747f4538bb15f"></script>
3740 3741 3742 3743 3744 3745
3745<script async src="https://d1bxh8uas1mnw7.cloudfront.net/assets/embed.js"></script>
3745 3746 3747 3748
3748<script async src="https://badge.dimensions.ai/badge.js"></script>
3748 3749 3750 3751 3752 3753 <!-- MathJax --> 3754
3754<script type="text/javascript"> 3755 window.MathJax = { 3756 tex: { 3757 tags: 'ams', 3758 }, 3759 }; 3760 </script>
3760 3761
3761<script defer type="text/javascript" id="MathJax-script" src="https://cdn.jsdelivr.net/npm/[email protected]/es5/tex-mml-chtml.min.js" integrity="sha256-rjmgmaB99riUNcdlrDtcAiwtLIojSxNyUFdl+Qh+rB4=" crossorigin="anonymous"></script>
3761 3762
3762<script defer src="https://cdnjs.cloudflare.com/polyfill/v3/polyfill.min.js?features=es6" crossorigin="anonymous"></script>
3762 3763 3764 3765 3766 3767 3768 3769 <!-- Global site tag (gtag.js) - Google Analytics -->
vendor: 66 bytes, lines 3769-3770
3769 3770 <script async src="https://www.googletagmanager.com/gtag/js?id=
3770G-2F7M79CZKB
vendor: 14 bytes, lines 3770-3771
3770"></script> 3771
3771<script> 3772 window.dataLayer = window.dataLayer || []; 3773 function gtag() { 3774 window.dataLayer.push(arguments); 3775 } 3776 gtag('js', new Date()); 3777 gtag('config', 'G-2F7M79CZKB'); 3778 </script>
3778 3779 3780 3781 3782 3783 3784 <!-- Scrolling Progress Bar --> 3785
3785<script type="text/javascript"> 3786 /* 3787 * This JavaScript code has been adapted from the article 3788 * https://css-tricks.com/reading-position-indicator/ authored by Pankaj Parashar, 3789 * published on the website https://css-tricks.com on the 7th of May, 2014. 3790 * Couple of changes were made to the original code to make it compatible 3791 * with the `al-foio` theme. 3792 */ 3793 const progressBar = $('#progress'); 3794 /* 3795 * We set up the bar after all elements are done loading. 3796 * In some cases, if the images in the page are larger than the intended 3797 * size they'll have on the page, they'll be resized via CSS to accomodate 3798 * the desired size. This mistake, however, breaks the computations as the 3799 * scroll size is computed as soon as the elements finish loading. 3800 * To account for this, a minimal delay was introduced before computing the 3801 * values. 3802 */ 3803 window.onload = function () { 3804 setTimeout(progressBarSetup, 50); 3805 }; 3806 /* 3807 * We set up the bar according to the browser. 3808 * If the browser supports the progress element we use that. 3809 * Otherwise, we resize the bar thru CSS styling 3810 */ 3811 function progressBarSetup() { 3812 if ('max' in document.createElement('progress')) { 3813 initializeProgressElement(); 3814 $(document).on('scroll', function () { 3815 progressBar.attr({ value: getCurrentScrollPosition() }); 3816 }); 3817 $(window).on('resize', initializeProgressElement); 3818 } else { 3819 resizeProgressBar(); 3820 $(document).on('scroll', resizeProgressBar); 3821 $(window).on('resize', resizeProgressBar); 3822 } 3823 } 3824 /* 3825 * The vertical scroll position is the same as the number of pixels that 3826 * are hidden from view above the scrollable area. Thus, a value > 0 is 3827 * how much the user has scrolled from the top 3828 */ 3829 function getCurrentScrollPosition() { 3830 return $(window).scrollTop(); 3831 } 3832 3833 function initializeProgressElement() { 3834 let navbarHeight = $('#navbar').outerHeight(true); 3835 $('body').css({ 'padding-top': navbarHeight }); 3836 $('progress-container').css({ 'padding-top': navbarHeight }); 3837 progressBar.css({ top: navbarHeight }); 3838 progressBar.attr({ 3839 max: getDistanceToScroll(), 3840 value: getCurrentScrollPosition(), 3841 }); 3842 } 3843 /* 3844 * The offset between the html document height and the browser viewport 3845 * height will be greater than zero if vertical scroll is possible. 3846 * This is the distance the user can scroll 3847 */ 3848 function getDistanceToScroll() { 3849 return $(document).height() - $(window).height(); 3850 } 3851 3852 function resizeProgressBar() { 3853 progressBar.css({ width: getWidthPercentage() + '%' }); 3854 } 3855 // The scroll ratio equals the percentage to resize the bar 3856 function getWidthPercentage() { 3857 return (getCurrentScrollPosition() / getDistanceToScroll()) * 100; 3858 } 3859 </script>
3859 3860 3861 3862 3863 3864 3865 3866 3867 3868 3869
3869<script src="/assets/js/vanilla-back-to-top.min.js?f40d453793ff4f64e238e420181a1d17"></script>
3869 3870
3870<script> 3871 addBackToTop(); 3872 </script>
3872 3873 3874 3875 3876 3877
3877<script src="/assets/js/shortcut-key.js"></script>
3877 3878 </body> 3879</html>
Line numbers count LF bytes from the start of the resource, as the search results do. Vendor segments are library code the classifier recognised; they are stored but not indexed. Bytes are shown as Latin1 characters, one per byte.