CapabilityDifferences.astro

<section id="mobile" class="section">
  <div class="section-inner">
    <div class="section-heading">
      <div>
        <h2>
          Capability Differences: Ahead on short-horizon tasks, behind on tabletop D&amp;P and
          long tasks
        </h2>
      </div>
      <p>
        Taking those annotations in turn gives a first account of where the deficit sits. Averaged
        within the benchmark's own task groups, GPT-6-Astra has the highest success rate of the
        eight models on the 11 short-horizon tasks (73.2%, against OpenWAM-&alpha;&rsquo;s 65.9%),
        but ranks sixth on the 15 long-horizon tasks (27.3%, against OpenWAM-&alpha;&rsquo;s 47.6%)
        and sixth on the seven tabletop D&amp;P tasks (20.0%, against Qwen-RobotManip&rsquo;s
        50.0%). The lead therefore comes from short procedures, eight of the eleven of which
        demand only low precision; once a task runs long or requires dexterous, precise contact,
        the post-trained policies stay ahead.
      </p>
    </div>
    <div id="task-gap-analysis" class="insight-block">
      <h3>Task-level strengths and weaknesses</h3>
    </div>
    <div id="mobile-content"></div>
  </div>
</section>