<html>
  <head>

    <meta http-equiv="content-type" content="text/html; charset=UTF-8">
  </head>
  <body>
    <p>Hi Everyone,</p>
    <p>Please come hear Adam tell us about self-driving cars and RL in
      his thesis proposal!</p>
    <p>Jeff.</p>
    <p><br>
    </p>
    <div class="moz-forward-container"><br>
      <br>
      -------- Forwarded Message --------
      <table class="moz-email-headers-table" cellspacing="0"
        cellpadding="0" border="0">
        <tbody>
          <tr>
            <th valign="BASELINE" nowrap="nowrap" align="RIGHT">Subject:
            </th>
            <td>RI Ph.D. Thesis Proposal: Adam Villaflor</td>
          </tr>
          <tr>
            <th valign="BASELINE" nowrap="nowrap" align="RIGHT">Date: </th>
            <td>Wed, 7 Sep 2022 13:58:33 -0400</td>
          </tr>
          <tr>
            <th valign="BASELINE" nowrap="nowrap" align="RIGHT">From: </th>
            <td>Suzanne Muth <a class="moz-txt-link-rfc2396E" href="mailto:lyonsmuth@cmu.edu"><lyonsmuth@cmu.edu></a></td>
          </tr>
          <tr>
            <th valign="BASELINE" nowrap="nowrap" align="RIGHT">To: </th>
            <td><a class="moz-txt-link-abbreviated" href="mailto:ri-people@lists.andrew.cmu.edu">ri-people@lists.andrew.cmu.edu</a></td>
          </tr>
        </tbody>
      </table>
      <br>
      <br>
      <div dir="ltr">
        <div class="gmail_default" style="">
          <div style=""><font style="" face="arial, sans-serif">Date: 16
              September 2022<br>
              Time: 1:00 p.m. (ET)<br>
              Location: GHC 4405<br>
              Zoom Link: <a
href="https://cmu.zoom.us/j/95581369179?pwd=THdjUnVkQnBaUUFlNDdOcDBMcVhNQT09"
                target="_blank" style="" moz-do-not-send="true"
                class="moz-txt-link-freetext">https://cmu.zoom.us/j/95581369179?pwd=THdjUnVkQnBaUUFlNDdOcDBMcVhNQT09</a><br>
              Type: Ph.D. Thesis Proposal<br>
              Who: Adam Villaflor<br>
              Title: Combining Offline Reinforcement Learning with
              Stochastic Multi-Agent Planning for Autonomous Driving</font></div>
          <div style=""><font style="" face="arial, sans-serif"><br>
            </font></div>
          <div style=""><font face="arial, sans-serif">Abstract:</font></div>
          <div style=""><font face="arial, sans-serif">Fully autonomous
              vehicles have the potential to greatly reduce vehicular
              accidents and revolutionize how people travel and how we
              transport goods. Many of the major challenges for
              autonomous driving systems emerge from the numerous
              traffic situations that require complex interactions with
              other agents. For the foreseeable future, autonomous
              vehicles will have to share the road with human-drivers
              and pedestrians, and thus cannot rely on centralized
              communication to address these interactive scenarios.
              Therefore, autonomous driving systems need to be able to
              negotiate and respond to unknown agents that exhibit
              uncertain behavior. To tackle these problems, most
              commercial autonomous driving stacks use a modular
              approach that splits perception, agent forecasting, and
              planning into separately engineered modules. By
              decomposing autonomous driving into smaller modules, it
              allows for simplifying abstractions and greater
              parallelization of the engineering effort.<br>
              <br>
              However, fully separating prediction and planning makes it
              difficult to reason about how other vehicles will respond
              to the planned trajectory for the controlled ego-vehicle.
              Thus to maintain safety, many modular approaches have to
              be overly conservative when interacting with other agents.
              Ideally, we want autonomous vehicles to drive in a natural
              and confident manner, while still maintaining safety. We
              believe that to achieve this behavior we need 3 major
              components. First, we need an approach that unifies
              prediction and planning in a single probabilistic
              closed-loop planning framework. Second, we need to use a
              multi-agent formulation in combination with deep learning
              models that can scale to the complexities of real-world
              driving and effectively model the interactive multi-modal
              distributions of real-world traffic. Finally, we need
              approaches that can effectively search the space of
              potential multi-agent interactions across time efficiently
              in order to produce a suitable planned behavior. In this
              proposal, we will show our current progress in applying
              deep offline reinforcement learning to autonomous driving,
              and present future work to continue scaling deep learning
              approaches to more complicated and interactive autonomous
              driving problems.</font></div>
          <div style=""><font face="arial, sans-serif"><br>
            </font></div>
          <div style=""><font face="arial, sans-serif">Thesis Committee
              Members:</font></div>
          <div style=""><font face="arial, sans-serif">Jeff Schneider,
              Chair<br>
              John Dolan, Co-Chair<br>
              David Held<br>
              Philipp Krähenbühl (UT Austin)</font></div>
          <div style=""><font face="arial, sans-serif"><br>
              A draft of the thesis proposal document is available at<br>
            </font></div>
          <div style=""><a
href="https://drive.google.com/drive/folders/14DoKABsNlCx9AKK9O1ZY77Roau4ugEmG?usp=sharing"
              target="_blank" style="" moz-do-not-send="true"><font
                style="" face="arial, sans-serif">https://drive.google.com/drive/folders/14DoKABsNlCx9AKK9O1ZY77Roau4ugEmG?usp=sharing</font></a></div>
        </div>
      </div>
    </div>
  </body>
</html>