index.html

<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Yiyuan Pan</title>
<meta name="description" content="Personal academic homepage of Yiyuan Pan.">
<meta name="theme-color" content="#F0EFEB">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:opsz,wght@14..32,400..700&display=swap" rel="stylesheet">
<link rel="stylesheet" href="style.css?v=2026-10-04.5">  <!-- bump ?v= with every style.css change, so a new page never meets a cached old stylesheet -->
<link rel="icon" href="assets/favicon.svg" type="image/svg+xml">
<script>
  // Motion on scroll: flag JS support before first paint (so content stays visible without JS), then mark each
  // publication, the timeline, and the publications title (which sets the skater riding) with .in as it enters
  // the viewport, staggered 90ms apart. The timeline and the skater play once per visit: coming back to the page
  // (from the misc page, say) they are already in place (.seen-timeline, .seen-skater).
  if ('IntersectionObserver' in window) {
    const root = document.documentElement;
    const seen = (key) => { try { return sessionStorage.getItem(key) === 'played'; } catch { return false; } };
    const remember = (key) => { try { sessionStorage.setItem(key, 'played'); } catch {} };
    root.classList.add('js');
    if (seen('timeline')) root.classList.add('seen-timeline');
    if (seen('skater')) root.classList.add('seen-skater');
    addEventListener('DOMContentLoaded', () => {
      const io = new IntersectionObserver((entries) => {
        let n = 0;
        for (const e of entries) {
          if (!e.isIntersecting) continue;
          e.target.style.setProperty('--delay', Math.min(n++, 8) * 90 + 'ms');
          e.target.classList.add('in');
          if (e.target.matches('.timeline')) remember('timeline');
          if (e.target.matches('#publications h2')) remember('skater');
          io.unobserve(e.target);
        }
      }, { threshold: 0.2 });
      const targets = ['#publications .entry'];
      if (!seen('timeline')) targets.push('.timeline');
      if (!seen('skater')) targets.push('#publications h2');
      document.querySelectorAll(targets.join(', ')).forEach((el) => io.observe(el));
    });
  }
</script>
</head>
<body>

<!-- Background doodles: faint, tilted, bleeding off the page edges. Per doodle: --top (from the page top),
     --x (offset from the page center), --w (width), --r (rotation), --o (opacity); the -sm values apply on phones.
     A doodle with data-beside is pinned to an entry by the script at the end of the page: its middle at the entry's
     middle (data-at="middle") or in the gap above it (data-at="above"); its --top is then only the fallback. -->
<div class="doodle-field" aria-hidden="true">
  <img src="assets/doodles/profile.svg" alt="" width="109" height="106" data-beside="#cmu" data-at="middle" style="--top: 560px; --x: 475px; --w: 220px; --r: -12deg; --x-sm: 40px; --w-sm: 170px">
  <img src="assets/doodles/leap.svg" alt="" class="sm-hide" loading="lazy" style="--top: 32%; --x: -735px; --w: 290px; --r: 7deg">
  <img src="assets/doodles/celebrate.svg" alt="" width="572" height="1039" data-beside="#dream-to-recall" data-at="above" style="--top: 76%; --x: 455px; --w: 170px; --r: 10deg; --o: .065; --x-sm: 115px; --w-sm: 130px">
</div>

<div class="wrap">

  <header class="hero">
    <div class="hero-text">
      <h1>Yiyuan Pan</h1>
      <p class="tagline">M.S. in Robotics · Carnegie Mellon University</p>
      <p class="intro">
        I am a second-year MSR student at Carnegie Mellon University, advised by
        <a href="https://icontrol.ri.cmu.edu/people/changliu.html" target="_blank" rel="noopener noreferrer">Changliu Liu</a>. I also work with
        <a href="https://jiayuanm.com/" target="_blank" rel="noopener noreferrer">Jiayuan Mao</a>. My research focuses on what latent structure should remain explicit—or be explicitly induced—inside powerful learned models for planning and control.
      </p>

      <nav class="links" aria-label="Contact and profile links">
        <a href="mailto:yiyuanp@andrew.cmu.edu">Email</a>
        <a href="https://scholar.google.com/citations?user=xqixK_wAAAAJ&amp;hl=en&amp;oi=ao" target="_blank" rel="noopener noreferrer">Google Scholar</a>
        <a href="https://github.com/PyyWill" target="_blank" rel="noopener noreferrer">GitHub</a>
        <a href="assets/Yiyuan_Pan_CV.pdf" target="_blank" rel="noopener noreferrer">CV</a>
      </nav>
    </div>

    <div class="hero-photo">
      <img src="assets/photo.jpg" alt="Yiyuan Pan" width="220" height="220">
      <a class="misc-link" href="misc/">Misc <span class="misc-link-arrow" aria-hidden="true">→</span></a>
    </div>
  </header>

  <main>

    <section id="education">
      <h2>Education</h2>
      <div class="entry-list">

        <article class="entry" id="cmu">
          <div class="entry-year">2025—Present</div>
          <div class="entry-body">
            <h3>Carnegie Mellon University</h3>
            <div class="entry-meta">M.S. in Robotics</div>
            <div class="entry-meta">Advisor: <a href="https://icontrol.ri.cmu.edu/people/changliu.html" target="_blank" rel="noopener noreferrer">Prof. Changliu Liu</a></div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2021—2025</div>
          <div class="entry-body">
            <h3>Shanghai Jiao Tong University</h3>
            <div class="entry-meta">B.S. in Automation</div>
            <div class="entry-meta">Advisor: <a href="https://irmv.sjtu.edu.cn/liuzhe" target="_blank" rel="noopener noreferrer">Prof. Zhe Liu</a>, <a href="https://irmv.sjtu.edu.cn/wanghesheng" target="_blank" rel="noopener noreferrer">Prof. Hesheng Wang</a></div>
          </div>
        </article>

      </div>
    </section>

    <section id="experience">
      <h2>Experience</h2>
      <ol class="timeline">
        <li class="xp">
          <span class="xp-logo xp-logo--caltech"><img src="assets/logos/caltech.svg" alt=""></span>
          <span class="xp-rail" aria-hidden="true"></span>
          <span class="xp-year">2024</span>
          <a class="xp-title" href="https://netlab.caltech.edu/" target="_blank" rel="noopener noreferrer">Caltech · Netlab</a>
          <span class="xp-meta">Advisor: <a href="https://www.cms.caltech.edu/people/slow" target="_blank" rel="noopener noreferrer">Prof. Steven Low</a></span>
        </li>
        <li class="xp">
          <span class="xp-logo xp-logo--bytedance"><img src="assets/logos/bytedance.svg" alt=""></span>
          <span class="xp-rail" aria-hidden="true"></span>
          <span class="xp-year">2025</span>
          <a class="xp-title" href="https://seed.bytedance.com/en/" target="_blank" rel="noopener noreferrer">ByteDance Seed</a>
          <span class="xp-meta">Advisor: <a href="https://hangli-hl.github.io/" target="_blank" rel="noopener noreferrer">Dr. Hang Li</a></span>
        </li>
        <li class="xp xp--now">
          <span class="xp-logo xp-logo--penn"><img src="assets/logos/penn.svg" alt=""></span>
          <span class="xp-rail" aria-hidden="true"></span>
          <span class="xp-year">2026</span>
          <a class="xp-title" href="https://thelabone.org/" target="_blank" rel="noopener noreferrer">UPenn · LabOne</a>
          <span class="xp-meta">Advisor: <a href="https://jiayuanm.com/" target="_blank" rel="noopener noreferrer">Prof. Jiayuan Mao</a></span>
        </li>
      </ol>
    </section>

    <section id="publications">
      <h2>Selected Publications<img class="doodle doodle--skater" src="assets/doodles/skater.svg" alt="" width="64" height="54" loading="lazy" decoding="async"></h2>
      <div class="entry-list">
        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://arxiv.org/abs/2606.26588" target="_blank" rel="noopener noreferrer">Inference-Time Robot Behavior Steering through Physically-Aware Reconfiguration of Task-Structure</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Hanjiang Hu, Shangtao Li, Xusheng Luo, and Changliu Liu.</div>
            <div class="entry-venue">CoRL 2026</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://arxiv.org/abs/2603.25903" target="_blank" rel="noopener noreferrer">Emergent Neural Automaton Policies: Learning Symbolic Structure from Visuomotor Trajectories</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span>*</strong>, Xusheng Luo*, Hanjiang Hu, Peiqi Yu, and Changliu Liu.</div>
            <div class="entry-venue">RSS 2026</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3>What Makes a Good Symbol for Hybrid Discrete–Continuous Planning?</h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Changliu Liu, and Jiayuan Mao.</div>
            <div class="entry-venue">Under Review</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2025</div>
          <div class="entry-body">
            <h3><a href="https://proceedings.neurips.cc/paper_files/paper/2025/hash/23bff092672192694aa05ff412e997d9-Abstract-Conference.html" target="_blank" rel="noopener noreferrer">Seeing through Uncertainty: Robust Task-Oriented Optimization in Visual Navigation</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Yunzhe Xu, Zhe Liu, and Hesheng Wang.</div>
            <div class="entry-venue">NeurIPS 2025</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2025</div>
          <div class="entry-body">
            <h3><a href="https://proceedings.neurips.cc/paper_files/paper/2025/hash/a86f9983dee679f0d81ab712a8f7e17c-Abstract-Conference.html" target="_blank" rel="noopener noreferrer">Wonder Wins Ways: Curiosity-Driven Exploration through Multi-Agent Contextual Calibration</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Zhe Liu, and Hesheng Wang.</div>
            <div class="entry-venue">NeurIPS 2025</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2025</div>
          <div class="entry-body">
            <h3><a href="https://ojs.aaai.org/index.php/AAAI/article/view/32679" target="_blank" rel="noopener noreferrer">Planning from Imagination: Episodic Simulation and Episodic Memory for Vision-and-Language Navigation</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Yunzhe Xu, Zhe Liu, and Hesheng Wang.</div>
            <div class="entry-venue">AAAI 2025 · Oral</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2023</div>
          <div class="entry-body">
            <h3><a href="https://ieeexplore.ieee.org/abstract/document/10330971/" target="_blank" rel="noopener noreferrer">Robust Scheduling of Integrated Energy Systems with Decision-Dependent Uncertainties</a></h3>
            <div class="entry-meta"><strong><span class="author-me">Yiyuan Pan</span></strong>, Jihang Wang, Jiabei He, and Zhaojian Wang.</div>
            <div class="entry-venue">IEEE PowerCon 2023</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://arxiv.org/abs/2606.03954" target="_blank" rel="noopener noreferrer">VLESA: Vision-Language Embodied Safety Agent for Human Activity Monitoring</a></h3>
            <div class="entry-meta">Hanjiang Hu*, <strong><span class="author-me">Yiyuan Pan</span>*</strong>, Jiaxing Li, Xusheng Luo, Alexander Robey, Na Li, Yebin Wang, and Changliu Liu.</div>
            <div class="entry-venue">arXiv 2026</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://arxiv.org/abs/2603.22703" target="_blank" rel="noopener noreferrer">Learning Safe-Stoppability Monitors for Humanoid Robots</a></h3>
            <div class="entry-meta">Yifan Sun, <strong><span class="author-me">Yiyuan Pan</span></strong>, Shangtao Li, Caiwu Ding, Tao Cui, Lingyun Wang, and Changliu Liu.</div>
            <div class="entry-venue">IROS 2026</div>
          </div>
        </article>

        <article class="entry" id="dream-to-recall">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://ieeexplore.ieee.org/abstract/document/11459135/" target="_blank" rel="noopener noreferrer">Dream to Recall: Imagination-Guided Experience Retrieval for Memory-Persistent Vision-and-Language Navigation</a></h3>
            <div class="entry-meta">Yunzhe Xu, <strong><span class="author-me">Yiyuan Pan</span></strong>, and Zhe Liu.</div>
            <div class="entry-venue">IEEE TPAMI 2026</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2025</div>
          <div class="entry-body">
            <h3><a href="https://ojs.aaai.org/index.php/AAAI/article/view/32974" target="_blank" rel="noopener noreferrer">FLAME: Learning to Navigate with Multimodal LLM in Urban Environments</a></h3>
            <div class="entry-meta">Yunzhe Xu, <strong><span class="author-me">Yiyuan Pan</span></strong>, Zhe Liu, and Hesheng Wang.</div>
            <div class="entry-venue">AAAI 2025 · Oral</div>
          </div>
        </article>

        <article class="entry">
          <div class="entry-year">2026</div>
          <div class="entry-body">
            <h3><a href="https://proceedings.iclr.cc/paper_files/paper/2026/hash/ec370e882e8a02be955679ddbd2731b8-Abstract-Conference.html" target="_blank" rel="noopener noreferrer">Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory</a></h3>
            <div class="entry-meta">Lin Long*, Yichen He*, Wentao Ye, <strong><span class="author-me">Yiyuan Pan</span></strong>, Yuan Lin, Hang Li, Junbo Zhao, and Wei Li.</div>
            <div class="entry-venue">ICLR 2026</div>
          </div>
        </article>

      </div>
    </section>

  </main>

  <footer>
    <p>© 2026 Yiyuan Pan · <span class="credit-own">Doodles hand-drawn by me</span> · Layout adapted from <a href="https://pd-perry.github.io" target="_blank" rel="noopener noreferrer">Perry Dong</a><span class="visits" hidden> · <a class="visits-count" target="_blank" rel="noopener noreferrer"></a></span></p>
    <img class="doodle doodle--sticker" src="assets/doodles/sticker.svg" alt="" width="44" height="46" loading="lazy" decoding="async">
  </footer>

</div>

<script>
  // Pin the background doodles that have data-beside to their entry, measuring from the page top (offsets ignore
  // the publications' rise-in transform). Again once the fonts have loaded, and whenever the window resizes.
  (() => {
    const pageTop = (el) => { let y = 0; for (; el; el = el.offsetParent) y += el.offsetTop; return y; };
    const pin = () => document.querySelectorAll('.doodle-field [data-beside]').forEach((img) => {
      const entry = document.querySelector(img.dataset.beside);
      if (!entry) return;
      const y = img.dataset.at === 'above' ? pageTop(entry) - 24 : pageTop(entry) + entry.offsetHeight / 2;  // 24px up: the hairline between entries
      img.style.setProperty('--top', `${Math.round(y - img.offsetHeight / 2)}px`);
    });
    pin();
    document.fonts?.ready.then(pin);
    addEventListener('resize', pin);
  })();
</script>

<!-- Visitor statistics: GoatCounter (no cookies); the footer shows the total, locations are in its dashboard. -->
<script data-goatcounter="https://yiyuanp.goatcounter.com/count" async src="https://gc.zgo.at/count.js"></script>
<script src="assets/visits.js?v=2026-10-04.5" defer></script>

</body>
</html>