<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>🖼️ 多模态提示 on Elon&#39;s AD Insight</title>
    <link>https://auto-driving-blog.pages.dev/tags/%EF%B8%8F-%E5%A4%9A%E6%A8%A1%E6%80%81%E6%8F%90%E7%A4%BA/</link>
    <description>Recent content in 🖼️ 多模态提示 on Elon&#39;s AD Insight</description>
    <image>
      <title>Elon&#39;s AD Insight</title>
      <url>https://auto-driving-blog.pages.dev/images/share.png</url>
      <link>https://auto-driving-blog.pages.dev/images/share.png</link>
    </image>
    <generator>Hugo</generator>
    <language>zh-cn</language>
    <lastBuildDate>Thu, 22 Oct 2026 00:00:00 +0000</lastBuildDate>
    <atom:link href="https://auto-driving-blog.pages.dev/tags/%EF%B8%8F-%E5%A4%9A%E6%A8%A1%E6%80%81%E6%8F%90%E7%A4%BA/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>精读 WorldScape Policy 2.0：给 WAM 装上「推理增强的长短期记忆」——RoboTwin 平均 94.3%，真机 PiPER 也稳</title>
      <link>https://auto-driving-blog.pages.dev/posts/paper-reading/worldscape%E7%B2%BE%E8%AF%BB/</link>
      <pubDate>Thu, 22 Oct 2026 00:00:00 +0000</pubDate>
      <guid>https://auto-driving-blog.pages.dev/posts/paper-reading/worldscape%E7%B2%BE%E8%AF%BB/</guid>
      <description>WorldScape Policy 2.0（Manifold AI + 清华 + 上交，arXiv 2607.18840）精读：可操控 WAM，核心是「推理增强的长短期记忆」——因果短期视觉记忆把最近观测作为 DiT prefill 保住局部动力学；长短期事件记忆把历史 VLM 输出组织成 global-history / local-active / event-boundary，按进度检索并门控进当前推理；semantic forcing 把事件级指令语义灌进隐式规划路径。预训练数据 ManipEvent-5M 近 500 万事件段，对齐动作轨迹、整段指令、子任务字幕、目标图像与视频演示。RoboTwin 2.0 干净+随机平均 94.3%（比 π0.5 高 14.5 pp），C2R 60.1%；真机双臂 PiPER：折叠 75%、桌面清理 80%、shell game 75%、示范条件积木 60–70%。RoboDojo 榜单上 WorldScape-R2SI 39.35% SR / 46.6 progress，Memory/Open 分类是学习型最强（63.7% / 47.5%）。逐张讲解全部 11 张插图。</description>
    </item>
  </channel>
</rss>
