index.html

---
layout: default
title: joso
---

<div class="section">
  <div class="subsection">
    <button onclick="dark()" aria-label="Toggle dark mode">
      <div class="title-img">
          <img src="/media/img/headshot.jpg" alt="Headshot of John So">
          <script>
            function dark() {
              let element = document.body;
              element.classList.toggle("light");
            }
            </script>
      </div>
    </button>
    <div class="text-container title-text">
      <h1>john so</h1>
      <p><i>aspiring full-stack roboticist!</i></p>
      <p>[johnso backwards] &lt;at&gt; gmail &lt;dot&gt; com</p>
      <p> &#123;
        <!-- <a target="_blank" href="/media/pdf/johnso_2023.pdf">resume</a>, -->
        <a target="_blank" href="https://twitter.com/johnrso_">twitter</a>,
        <a target="_blank" href="https://www.linkedin.com/in/johnianrso/">linkedin</a>,
        <a target="_blank" href="https://github.com/johnrso">github</a>,
        <!-- <a target="_blank" href="/media/pdf/johnso_2024.pdf">cv</a>, -->
        <a target="_blank" href="https://scholar.google.com/citations?user=r0vGtTwAAAAJ">scholar</a>
      &#125; </p>
    </div>
  </div>
  <div class="text-container">
    <p>
      I currently work on robot learning at <a target="_blank" href="https://www.tesla.com/en_eu/AI">Tesla Optimus</a>, where I focus on policy learning and human data collection. I previously spent time as an AI resident at <a target="_blank" href="https://www.1x.tech/about">1X</a>, where I focused on data and training foundation policy models.
    </p>
    <p class="last-p">
      I received my MS CS at <a target="_blank" href="https://www.stanford.edu/">Stanford University</a>, where I was fortunate to be advised by
      <a target="_blank" href="https://shurans.github.io/">Shuran Song</a> as a member of Stanford <a target="_blank" href="https://real.stanford.edu/">REAL</a> (Robotics and Embodied Artificial Intelligence Lab), and my BS EECS from <a target="_blank" href="https://engineering.berkeley.edu/">UC Berkeley</a>, where I was advised by
      <a target="_blank" href="https://people.eecs.berkeley.edu/~pabbeel/">Pieter Abbeel</a>,
      <a target="_blank" href="https://stepjam.github.io/">Stephen James</a>, and
      <a target="_blank" href="https://xingyu-lin.github.io/">Xingyu Lin</a> as a part of Berkeley
      <a target="_blank" href="https://rll.berkeley.edu/">RLL</a> (Robot Learning Lab).
    </p>
  </div>
</div>
<div class="section">
  <div class="text-container">
    <h2>research<hr></h2>
  </div>
  <div class="subsection">
    <div class="text-container">
      <p>
        My dream is for robots to become an everyday household occurrence, and I believe that <strong>large scale data</strong> is fundamental to bridge this gap between dexterity and robust generalization to unseen scenarios; how can we learn priors for dexterity from non-robot data, such as human videos and simulation? How do we scale, curate, and augment robot data?</li>
      </ol>
      <p class="last-p">
        Long term, I hope to leverage perspectives from cognitive science and developmental psychology to inform how robots can learn from and like humans.
      </p>
    </div>
  </div>
  <div class="subsection project">
    <div class="project-content text-container">
      <h3>Any-point Trajectory Modeling for Policy Learning</h3>
      <div class="project-text">
      <p>
        <a target="_blank" href="https://alvinwen428.github.io/">Chuan Wen</a>*,
        <a target="_blank" href="https://xingyu-lin.github.io/">Xingyu Lin</a>*,
        <b><u>John So</u></b>*,
        <a target="_blank" href="https://www.cse.cuhk.edu.hk/~qdou/">Qi Dou</a>,
        <a target="_blank" href="https://ck-kai.github.io/">Kai Chen</a>,
        <a target="_blank" href="https://yang-gao.weebly.com/">Yang Gao</a>,
        <a target="_blank" href="https://people.eecs.berkeley.edu/~pabbeel/">Pieter Abbeel</a>
      </p>
      <p>
        <strong>TL;DR:</strong> We condition a policy on arbitrary point trajectories learned from actionless videos, enabling sample-efficient policy learning and positive cross-embodiment transfer.
      </p>
      <p>
        <span class="badge conference"><b>RSS 2024</b></span>
        &#123; <a target="_blank" href="/media/pdf/atm.pdf">paper</a>,
        <a target="_blank" href="https://arxiv.org/abs/2401.00025" >arXiv</a>,
        <a target="_blank" href="https://xingyu-lin.github.io/atm/">website</a>,
        <a target="_blank" href="https://github.com/Large-Trajectory-Model/ATM">code</a> &#125;
      </p>
      </div>
    </div>
    <a target="_blank" href="https://xingyu-lin.github.io/atm/" class="no-hover">
      <video class="project-media" src="/media/vid/atm.mp4" autoplay muted inline loop>
        Your browser does not support the video tag.
      </video>
    </a>
  </div>
  <hr>
  <div class="subsection project">
    <div class="project-content text-container">
      <h3>SpawnNet: Learning Generalizable Visuomotor Skills from Pre-trained Networks</h3>
      <div class="project-text">
      <p>
        <a target="_blank" href="https://xingyu-lin.github.io/">Xingyu Lin</a>*,
        <b><u>John So</u></b>*,
        <a target="_blank" href="https://sashwat-mahalingam.github.io">Sashwat Mahalingam</a>,
        <a target="_blank" href="https://fangchenliu.github.io">Fangchen Liu</a>,
        <a target="_blank" href="https://people.eecs.berkeley.edu/~pabbeel/">Pieter Abbeel</a>
      </p>
      <p>
        <strong>TL;DR:</strong> We propose a method to adapt dense features from pre-trained vision backbones for visuomotor policies, enabling sample-efficient generalization to unseen objects.
      </p>
      <p>
        <span class="badge conference"><b>ICRA 2024</b></span>
        &#123; <a target="_blank" href="/media/pdf/spawnnet.pdf">paper</a>,
        <a target="_blank" href="https://arxiv.org/abs/2307.03567" >arXiv</a>,
        <a target="_blank" href="https://xingyu-lin.github.io/spawnnet/">website</a>,
        <a target="_blank" href="https://github.com/johnrso/spawnnet">code</a> &#125;
      </p>
      </div>
    </div>
    <a target="_blank" href="https://xingyu-lin.github.io/spawnnet/" class="no-hover">
      <video class="project-media" src="/media/vid/spawnnet.mp4" autoplay muted inline loop>
        Your browser does not support the video tag.
      </video>
    </a>
  </div>
  <hr>
  <div class="subsection project">
    <div class="project-content text-container">
      <h3>Sim-to-Real via Sim-to-Seg: End-to-end Off-road Autonomous Driving Without Real Data</h3>
      <div class="project-text">
      <p>
        <b><u>John So</u></b>*,
        <a target="_blank" href="https://amberxie88.github.io/" >Amber Xie</a>*,
        <a target="_blank" href="https://www-robotics.jpl.nasa.gov/who-we-are/people/sunggoo-jung/" >Sunggoo Jung</a>,
        <a target="_blank" href="https://www-robotics.jpl.nasa.gov/who-we-are/people/jeffrey_edlund/" >Jeffrey Edlund</a>,
        <a target="_blank" href="https://www-robotics.jpl.nasa.gov/who-we-are/people/rohan_thakker/" >Rohan Thakker</a>,
        <a target="_blank" href="https://aliagha.site/" >Ali Agha-mohammadi</a>,
        <a target="_blank" href="https://people.eecs.berkeley.edu/~pabbeel/" >Pieter Abbeel</a>,
        <a target="_blank" href="https://stepjam.github.io/" >Stephen James</a>
      </p>
      <p>
        <strong>TL;DR:</strong> We train a navigation policy in simulation with RL in learned segmentation space, and deploy zero-shot to a real vehicle.
      </p>
      <p>
        <span class="badge conference"><b>CoRL 2022</b></span>
        &#123; <a target="_blank" href="/media/pdf/sim-to-seg.pdf" >paper</a>,
        <a target="_blank" href="https://arxiv.org/abs/2210.14721">arXiv</a>,
        <a target="_blank" href="https://sites.google.com/view/sim2segcorl2022/home">website</a>,
        <a target="_blank" href="https://github.com/rll-research/sim2seg">code</a> &#125;
      </p>
      </div>
    </div>
    <a target="_blank" href="https://sites.google.com/view/sim2segcorl2022/home" class="no-hover">
      <video class="project-media" src="/media/vid/sim2seg.mp4" autoplay muted inline loop>
        Your browser does not support the video tag.
      </video>
    </a>
  </div>
</div>
<div class="section">
  <div class="text-container">
    <h2>teaching<hr></h2>
  </div>
  <div class="subsection">
    <div class="text-container">
      <p class="last-p">
        <a target="_blank" href="https://cs229.stanford.edu/">CS 229: Machine Learning</a> — Winter 2025, Winter 2024 <br>
        <a target="_blank" href="https://cs221.stanford.edu/">CS 221: Artificial Intelligence: Principles and Techniques</a> — Fall 2024 <br>
        <a target="_blank" href="https://people.eecs.berkeley.edu/~jrs/189/">CS 189: Introduction to Machine Learning</a> — Spring 2023 <br>
        <a target="_blank" href="https://cs61a.org/">CS 61A: Structure and Interpretation of Computer Programs</a> — Fall 2022, Spring 2022, Fall 2021
      </p>
    </div>
  </div>
</div>
<div class="section">
  <div class="text-container">
    <h2>miscellaneous<hr></h2>
  </div>
  <div class="subsection">
    <div class="text-container">
      <p>
        I'm forever grateful to the communities which raised me, notably <a target="_blank" href="https://ml.berkeley.edu/">Machine Learning at Berkeley (ML@B)</a>, <a target="_blank" href="https://cs61a.org/">61A Staff</a>, <a target="_blank" href="https://eecs.berkeley.edu/resources/undergrads/accel/">Accel Scholars</a>, and <a target="_blank" href="https://www.felicis.com/fellows">Felicis Fellows</a>.
      </p>
      <p>
        In particular, I spent the majority of my undergrad learning from and organizing
        <a target="_blank" href="https://ml.berkeley.edu/" >ML@B</a>,
        serving as the organization's president in Fall 2022. We presented a white paper about our structure
        and initiatives at the NeurIPS 2022 <a target="_blank" href="https://sites.google.com/view/broadening-collaboration-in-ml/home">Broadening Research Collaborations in ML Workshop</a>;
        you may find a preprint <a target="_blank" href="/media/pdf/built_to_last.pdf">here</a>.
      </p>
      <p class="last-p">
        Generally, I like to think about how to best teach, learn, and optimize for fulfillment. Rarely, I <a target="_blank" href="blog.html">write my thoughts down</a>. Shoot me an email or DM if you'd like to chat. &#129312;
      </p>
    </div>
  </div>
</div>

<div class="section footer">
  <p>last updated: {{ site.time | date_to_string }}<a href="#">&#128640;</a></p>
</div>