{"id":100999,"date":"2026-09-05T22:25:30","date_gmt":"2026-09-05T14:25:30","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/100999.html"},"modified":"2026-09-05T22:25:30","modified_gmt":"2026-09-05T14:25:30","slug":"dawn%ef%bc%9a%e4%b8%96%e7%95%8c-%e5%8a%a8%e4%bd%9c%e4%ba%a4%e4%ba%92%e6%a8%a1%e5%9e%8b%e7%9a%84%e9%bb%8e%e6%98%8e","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/100999.html","title":{"rendered":"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e"},"content":{"rendered":"

1. \u5f15\u8a00：\u81ea\u52a8\u9a7e\u9a76\u4e2d\u7684\u672a\u6765\u9884\u6d4b\u56f0\u5883<\/h3>\n

\u5728\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u7684\u51b3\u7b56\u8fc7\u7a0b\u4e2d，\u5b58\u5728\u4e00\u4e2a\u672c\u8d28\u6027\u7684\u77db\u76fe：\u672a\u6765\u573a\u666f\u7684\u6f14\u5316\u4f9d\u8d56\u4e8e\u5f53\u524d\u8003\u8651\u7684\u9a7e\u9a76\u52a8\u4f5c，\u800c\u5408\u7406\u7684\u9a7e\u9a76\u52a8\u4f5c\u53c8\u4f9d\u8d56\u4e8e\u5bf9\u672a\u6765\u573a\u666f\u7684\u51c6\u786e\u9884\u6d4b\u3002\u8fd9\u79cd\u76f8\u4e92\u4f9d\u8d56\u7684\u5173\u7cfb\u5728\u4f20\u7edf\u7684World Action Model（\u4e16\u754c\u52a8\u4f5c\u6a21\u578b，\u7b80\u79f0WAM）\u4e2d\u5e76\u672a\u5f97\u5230\u5145\u5206\u4f53\u73b0\u3002\u73b0\u6709\u7684WAM\u8981\u4e48\u91c7\u7528\u5e76\u884c\u5206\u652f\u72ec\u7acb\u9884\u6d4b\u4e16\u754c\u72b6\u6001\u548c\u9a7e\u9a76\u52a8\u4f5c，\u8981\u4e48\u91c7\u7528\u4e25\u683c\u7684\u9884\u6d4b-\u89c4\u5212\u6d41\u6c34\u7ebf，\u5148\u56fa\u5b9a\u672a\u6765\u573a\u666f\u9884\u6d4b，\u518d\u57fa\u4e8e\u6b64\u8fdb\u884c\u52a8\u4f5c\u89c4\u5212\u3002<\/p>\n

\u8fd9\u4e9b\u8bbe\u8ba1\u5b58\u5728\u6839\u672c\u6027\u5c40\u9650：\u5e76\u884c\u8bbe\u8ba1\u867d\u7136\u5141\u8bb8\u4e16\u754c\u548c\u52a8\u4f5c\u4ea7\u751f\u5173\u8054，\u4f46\u65e0\u6cd5\u8ba9\u4e8c\u8005\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u76f8\u4e92\u91cd\u5851；\u987a\u5e8f\u8bbe\u8ba1\u867d\u7136\u8ba9\u52a8\u4f5c\u4f9d\u8d56\u4e8e\u9884\u6d4b\u7684\u672a\u6765，\u4f46\u8fd9\u4e2a\u672a\u6765\u662f\u56fa\u5b9a\u7684\u5047\u8bbe，\u5e76\u975e\u968f\u7740\u52a8\u4f5c\u5047\u8bbe\u6f14\u5316\u7684\u52a8\u6001\u672a\u6765\u3002\u5728\u590d\u6742\u7684\u4ea4\u4e92\u5f0f\u9a7e\u9a76\u573a\u666f\u4e2d，\u8fd9\u79cd\u7ed3\u6784\u6027\u89e3\u8026\u5bfc\u81f4\u6a21\u578b\u65e0\u6cd5\u5145\u5206\u7406\u89e3\u52a8\u4f5c\u4f9d\u8d56\u6027\u7684\u672a\u6765\u6f14\u5316\u7279\u6027\u3002<\/p>\n

DAWN（Denoising Actions and World iNteractive model）\u6b63\u662f\u5728\u8fd9\u6837\u7684\u80cc\u666f\u4e0b\u8bde\u751f\u7684\u3002\u5b83\u63d0\u51fa\u4e86World-Action Interactive Model（\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b，\u7b80\u79f0WAIM）\u7684\u65b0\u8303\u5f0f，\u5c06\u672a\u6765\u4e16\u754c\u72b6\u6001\u548c\u9a7e\u9a76\u52a8\u4f5c\u89c6\u4e3a\u9700\u8981\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5171\u540c\u63a8\u65ad\u7684\u8026\u5408\u53d8\u91cf，\u800c\u975e\u72ec\u7acb\u751f\u6210\u6216\u5355\u5411\u4f9d\u8d56\u7684\u8f93\u51fa\u3002\u8fd9\u4e00\u7406\u5ff5\u7684\u6838\u5fc3\u5728\u4e8e：\u5728\u51b3\u7b56\u76f8\u5173\u7684\u672a\u6765\u672c\u8d28\u4e0a\u4f9d\u8d56\u4e8e\u6240\u8003\u8651\u7684\u52a8\u4f5c\u65f6，\u6a21\u578b\u4e0d\u5e94\u5148\u9884\u6d4b\u4e00\u4e2a\u88ab\u52a8\u7684\u4e16\u754c\u672a\u6765\u518d\u5728\u5176\u4e2d\u884c\u52a8，\u800c\u5e94\u8be5\u8054\u5408\u63a8\u65ad\u4e00\u4e2a\u4e16\u754c\u6f14\u5316\u4e0e\u51b3\u7b56\u5236\u5b9a\u4fdd\u6301\u76f8\u4e92\u5bf9\u9f50\u7684\u672a\u6765\u3002<\/p>\n

\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

2. \u4eceWAM\u5230WAIM：\u8303\u5f0f\u8f6c\u53d8\u7684\u672c\u8d28<\/h3>\n

2.1 \u4f20\u7edfWAM\u7684\u7ed3\u6784\u6027\u5c40\u9650<\/h4>\n

\u4f20\u7edf\u7684\u81ea\u52a8\u9a7e\u9a76\u7b56\u7565\u76f4\u63a5\u5efa\u6a21\u6761\u4ef6\u6982\u7387\u5206\u5e03 <\/p>\n

p<\/p>\n

(<\/p>\n

a<\/p>\n

1<\/p>\n

:<\/p>\n

H<\/p>\n

\u2223<\/p>\n

o<\/p>\n

,<\/p>\n

l<\/p>\n

)<\/p>\n

p(a_{1:H}\\\\mid o,l)<\/p>\n

<\/span><\/span>p<\/span>(<\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>o<\/span>,<\/span><\/span>l<\/span>)<\/span><\/span><\/span><\/span><\/span>，\u5176\u4e2d <\/p>\n

o<\/p>\n

o<\/p>\n

<\/span><\/span>o<\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u5f53\u524d\u89c2\u6d4b，<\/p>\n

l<\/p>\n

l<\/p>\n

<\/span><\/span>l<\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u4efb\u52a1\u6307\u4ee4，<\/p>\n

a<\/p>\n

1<\/p>\n

:<\/p>\n

H<\/p>\n

a_{1:H}<\/p>\n

<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u672a\u6765 <\/p>\n

H<\/p>\n

H<\/p>\n

<\/span><\/span>H<\/span><\/span><\/span><\/span><\/span> \u6b65\u7684\u52a8\u4f5c\u5e8f\u5217\u3002\u8fd9\u79cd\u76f4\u63a5\u6620\u5c04\u5ffd\u7565\u4e86\u672a\u6765\u4e16\u754c\u6f14\u5316\u5bf9\u51b3\u7b56\u7684\u5f71\u54cd\u3002<\/p>\n

World Action Model\u901a\u8fc7\u5f15\u5165\u672a\u6765\u4e16\u754c\u8868\u793a <\/p>\n

v<\/p>\n

1<\/p>\n

:<\/p>\n

T<\/p>\n

v_{1:T}<\/p>\n

<\/span><\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>，\u5c06\u95ee\u9898\u6269\u5c55\u4e3a\u8054\u5408\u5efa\u6a21 <\/p>\n

p<\/p>\n

(<\/p>\n

v<\/p>\n

1<\/p>\n

:<\/p>\n

T<\/p>\n

,<\/p>\n

a<\/p>\n

1<\/p>\n

:<\/p>\n

H<\/p>\n

\u2223<\/p>\n

o<\/p>\n

,<\/p>\n

l<\/p>\n

)<\/p>\n

p(v_{1:T}, a_{1:H}\\\\mid o,l)<\/p>\n

<\/span><\/span>p<\/span>(<\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>o<\/span>,<\/span><\/span>l<\/span>)<\/span><\/span><\/span><\/span><\/span>\u3002\u7406\u8bba\u4e0a，\u52a8\u4f5c\u5206\u5e03\u53ef\u4ee5\u901a\u8fc7\u5bf9\u6240\u6709\u53ef\u80fd\u672a\u6765\u7684\u8fb9\u7f18\u5316\u5f97\u5230：<\/p>\n

<\/p>\n

p<\/p>\n

(<\/p>\n

a<\/p>\n

1<\/p>\n

:<\/p>\n

H<\/p>\n

\u2223<\/p>\n

o<\/p>\n

,<\/p>\n

l<\/p>\n

)<\/p>\n

=<\/p>\n

\u222b<\/p>\n

p<\/p>\n

(<\/p>\n

v<\/p>\n

1<\/p>\n

:<\/p>\n

T<\/p>\n

,<\/p>\n

a<\/p>\n

1<\/p>\n

:<\/p>\n

H<\/p>\n

\u2223<\/p>\n

o<\/p>\n

,<\/p>\n

l<\/p>\n

)
\n \u2009<\/p>\n

d<\/p>\n

v<\/p>\n

1<\/p>\n

:<\/p>\n

T<\/p>\n

p(a_{1:H}\\\\mid o,l)=\\\\int p(v_{1:T},a_{1:H}\\\\mid o,l)\\\\,dv_{1:T} <\/p>\n

<\/span><\/span>p<\/span>(<\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>o<\/span>,<\/span><\/span>l<\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>\u222b<\/span><\/span>p<\/span>(<\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2223<\/span><\/span><\/span><\/span>o<\/span>,<\/span><\/span>l<\/span>)<\/span><\/span>d<\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

\u7136\u800c，\u73b0\u6709WAM\u7684\u5b9e\u73b0\u5728\u63a8\u7406\u65f6\u5f80\u5f80\u91c7\u53d6\u4e24\u79cd\u6781\u7aef\u7b56\u7565：<\/p>\n

  • \n

    \u5e76\u884c\u751f\u6210：\u4f7f\u7528\u72ec\u7acb\u7684\u5206\u652f\u4ece\u5171\u4eab\u7684\u89c6\u89c9\u4e0a\u4e0b\u6587\u540c\u65f6\u9884\u6d4b\u672a\u6765\u573a\u666f\u548c\u52a8\u4f5c，\u4e8c\u8005\u5728\u8bad\u7ec3\u65f6\u76f8\u5173\u4f46\u5728\u751f\u6210\u65f6\u5404\u81ea\u72ec\u7acb\u3002<\/p>\n<\/li>\n

  • \n

    \u987a\u5e8f\u6d41\u6c34\u7ebf：\u5148\u9884\u6d4b\u672a\u6765\u89c2\u6d4b\u3001\u5360\u7528\u6805\u683c\u6216\u6f5c\u5728\u573a\u666f\u72b6\u6001，\u7136\u540e\u57fa\u4e8e\u8fd9\u4e9b\u9884\u6d4b\u7684\u672a\u6765\u8fdb\u884c\u52a8\u4f5c\u89c4\u5212\u3002<\/p>\n<\/li>\n

    \u8fd9\u4e24\u79cd\u7b56\u7565\u7684\u5171\u540c\u95ee\u9898\u5728\u4e8e：\u5728\u751f\u6210\u65f6\u523b，\u4e00\u65b9\u76f8\u5bf9\u4e8e\u53e6\u4e00\u65b9\u662f\u56fa\u5b9a\u7684\u3002\u5e76\u884c\u8bbe\u8ba1\u5141\u8bb8\u76f8\u5173\u6027\u4f46\u4e0d\u5141\u8bb8\u8fed\u4ee3\u91cd\u5851；\u987a\u5e8f\u8bbe\u8ba1\u8ba9\u52a8\u4f5c\u4f9d\u8d56\u4e8e\u4e00\u4e2a\u51bb\u7ed3\u7684\u672a\u6765\u5047\u8bbe，\u800c\u975e\u968f\u52a8\u4f5c\u5047\u8bbe\u5171\u540c\u6f14\u5316\u7684\u672a\u6765\u3002<\/p>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    2.2 WAIM\u7684\u6838\u5fc3\u601d\u60f3<\/h4>\n

    WAIM\u5c06World-Action Interactive Model\u5b9a\u4e49\u4e3a\u4e00\u7c7b\u7279\u6b8a\u7684WAM，\u5176\u4e2d\u672a\u6765\u4e16\u754c\u548c\u672a\u6765\u52a8\u4f5c\u88ab\u63a8\u65ad\u4e3a\u8026\u5408\u53d8\u91cf，\u800c\u975e\u72ec\u7acb\u751f\u6210\u6216\u6309\u56fa\u5b9a\u5355\u5411\u987a\u5e8f\u751f\u6210\u3002\u5f62\u5f0f\u5316\u5730，WAIM\u5bfb\u6c42\u4e00\u4e2a\u81ea\u6d3d\u7684\u5bf9 <\/p>\n

    (<\/p>\n

    v<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    T<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    )<\/p>\n

    (\\\\hat{v}_{1:T}, \\\\hat{a}_{1:H})<\/p>\n

    <\/span><\/span>(<\/span><\/span>v<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span><\/span>a<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span>，\u4f7f\u5f97：<\/p>\n

    <\/p>\n

    v<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    T<\/p>\n

    =<\/p>\n

    F<\/p>\n

    \u03b8<\/p>\n

    (<\/p>\n

    o<\/p>\n

    ,<\/p>\n

    l<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    =<\/p>\n

    G<\/p>\n

    \u03d5<\/p>\n

    (<\/p>\n

    o<\/p>\n

    ,<\/p>\n

    l<\/p>\n

    ,<\/p>\n

    v<\/p>\n

    ^<\/p>\n

    1<\/p>\n

    :<\/p>\n

    T<\/p>\n

    )<\/p>\n

    \\\\hat{v}_{1:T}=F_{\\\\theta}(o,l,\\\\hat{a}_{1:H}),\\\\qquad \\\\hat{a}_{1:H}=G_{\\\\phi}(o,l,\\\\hat{v}_{1:T}) <\/p>\n

    <\/span><\/span><\/span>v<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>F<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>,<\/span><\/span>l<\/span>,<\/span><\/span><\/span>a<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span>,<\/span><\/span><\/span><\/span>a<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>G<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>,<\/span><\/span>l<\/span>,<\/span><\/span><\/span>v<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u5728\u5b9e\u8df5\u4e2d，\u8fd9\u53ef\u4ee5\u901a\u8fc7\u8fed\u4ee3\u4ea4\u4e92\u5b9e\u73b0：<\/p>\n

    <\/p>\n

    (<\/p>\n

    v<\/p>\n

    1<\/p>\n

    :<\/p>\n

    T<\/p>\n

    (<\/p>\n

    k<\/p>\n

    +<\/p>\n

    1<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    k<\/p>\n

    +<\/p>\n

    1<\/p>\n

    )<\/p>\n

    )<\/p>\n

    =<\/p>\n

    I<\/p>\n

    \u0398<\/p>\n

    (<\/p>\n

    v<\/p>\n

    1<\/p>\n

    :<\/p>\n

    T<\/p>\n

    (<\/p>\n

    k<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    k<\/p>\n

    )<\/p>\n

    ;<\/p>\n

    o<\/p>\n

    ,<\/p>\n

    l<\/p>\n

    )<\/p>\n

    (v_{1:T}^{(k+1)},a_{1:H}^{(k+1)}) =\\\\mathcal{I}_{\\\\Theta}(v_{1:T}^{(k)},a_{1:H}^{(k)};o,l) <\/p>\n

    <\/span><\/span>(<\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>(<\/span>k<\/span>+<\/span>1<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>k<\/span>+<\/span>1<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span>=<\/span><\/span><\/span><\/span>I<\/span><\/span>\u0398<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>v<\/span><\/span>1<\/span>:<\/span>T<\/span><\/span><\/span><\/span><\/span>(<\/span>k<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>k<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>;<\/span><\/span>o<\/span>,<\/span><\/span>l<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u5176\u4e2d <\/p>\n

    I<\/p>\n

    \u0398<\/p>\n

    \\\\mathcal{I}_{\\\\Theta}<\/p>\n

    <\/span><\/span>I<\/span><\/span>\u0398<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u8868\u793a\u4e00\u4e2a\u4ea4\u4e92\u7b97\u5b50，\u5b83\u540c\u65f6\u66f4\u65b0\u4e16\u754c\u5047\u8bbe\u548c\u52a8\u4f5c\u5047\u8bbe\u3002<\/p>\n

    \u5173\u952e\u533a\u522b\u5728\u4e8e：WAM\u8054\u5408\u5efa\u6a21\u672a\u6765\u4e16\u754c\u548c\u52a8\u4f5c，\u800cWAIM\u901a\u8fc7\u4ea4\u4e92\u8054\u5408\u63a8\u65ad\u5b83\u4eec\u3002\u8fd9\u79cd\u533a\u522b\u5728\u51b3\u7b56\u76f8\u5173\u7684\u672a\u6765\u4f9d\u8d56\u4e8e\u6240\u8003\u8651\u7684\u52a8\u4f5c，\u800c\u975e\u4ec5\u4f9d\u8d56\u4e8e\u573a\u666f\u52a8\u529b\u5b66\u65f6\u5c24\u4e3a\u91cd\u8981\u3002<\/p>\n

    \u5728\u672c\u5730\u4ee3\u7801\u91cc，WAIM\u7684\u4ea4\u4e92\u601d\u60f3\u4e3b\u8981\u843d\u5728\u9a8c\u8bc1\u5165\u53e3 DAWN\/scripts\/infer_nuscenes_val.py：\u5148\u6784\u5efa encoder\/predictor\/planner\/token_ae，\u518d\u628a\u8fd9\u4e9b\u6a21\u5757\u4ea4\u7ed9 run_validation \u6267\u884c\u9884\u6d4b\u548c\u89c4\u5212\u95ed\u73af\u3002<\/p>\n

    metrics =<\/span> run_validation(<\/span>
    \n encoder=<\/span>encoder,<\/span>
    \n predictor=<\/span>predictor,<\/span>
    \n planner=<\/span>planner,<\/span>
    \n val_loader=<\/span>val_loader,<\/span>
    \n val_sampler=<\/span>val_sampler,<\/span>
    \n config=<\/span>config,<\/span>
    \n epoch=<\/span>args.<\/span>epoch,<\/span>
    \n rank=<\/span>rank,<\/span>
    \n world_size=<\/span>world_size,<\/span>
    \n use_tubelet_repeat=<\/span>config.<\/span>data.<\/span>use_tubelet_repeat,<\/span>
    \n vis_output_dir=<\/span>vis_output_dir,<\/span>
    \n token_ae=<\/span>token_ae,<\/span>
    \n )<\/span><\/p>\n

    3. DAWN\u67b6\u6784\u8bbe\u8ba1：\u7d27\u51d1\u6f5c\u5728\u7a7a\u95f4\u4e2d\u7684\u4ea4\u4e92\u5f0f\u63a8\u7406<\/h3>\n

    3.1 \u6574\u4f53\u67b6\u6784\u6982\u89c8<\/h4>\n

    DAWN\u5728\u7d27\u51d1\u7684\u8bed\u4e49\u6f5c\u5728\u7a7a\u95f4\u4e2d\u5b9e\u4f8b\u5316WAIM，\u907f\u514d\u4e86\u6602\u8d35\u7684\u50cf\u7d20\u7ea7\u672a\u6765\u6e32\u67d3\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u4f7f\u7528\u77ed\u671f\u663e\u5f0f\u6f5c\u5728\u63a8\u6f14\u6765\u652f\u6301\u590d\u6742\u4ea4\u4e92\u573a\u666f\u4e2d\u7684\u957f\u671f\u52a8\u4f5c\u751f\u6210\u3002\u67b6\u6784\u5305\u542b\u4ee5\u4e0b\u5173\u952e\u7ec4\u4ef6：<\/p>\n

  • Student Vision-Encoder：\u4ece\u5f53\u524d\u89c2\u6d4b\u4e2d\u63d0\u53d6\u5bc6\u96c6\u89c6\u89c9\u7279\u5f81<\/li>\n
  • Teacher Vision-Encoder：\u4ec5\u5728\u8bad\u7ec3\u65f6\u4f7f\u7528，\u4e3a\u672a\u6765\u89c2\u6d4b\u63d0\u4f9b\u76d1\u7763\u4fe1\u53f7<\/li>\n
  • Auto-Encoder Resampler：\u5c06\u5bc6\u96c6\u7f16\u7801\u5668\u7279\u5f81\u538b\u7f29\u4e3a\u7d27\u51d1\u7684\u6f5c\u5728\u4e16\u754c\u8868\u793a<\/li>\n
  • World Predictor：\u5728\u6f5c\u5728\u7a7a\u95f4\u4e2d\u9884\u6d4b\u672a\u6765\u4e16\u754c\u72b6\u6001<\/li>\n
  • World-Conditioned Action Denoiser：\u57fa\u4e8e\u9884\u6d4b\u7684\u672a\u6765\u4e16\u754c\u53bb\u566a\u52a8\u4f5c\u5047\u8bbe<\/li>\n
  • Action Head：\u5c06\u53bb\u566a\u540e\u7684\u52a8\u4f5c\u72b6\u6001\u89e3\u7801\u4e3a\u6700\u7ec8\u8f68\u8ff9<\/li>\n

    \u8fd9\u79cd\u8bbe\u8ba1\u7684\u72ec\u7279\u4e4b\u5904\u5728\u4e8e，\u5b83\u4e0d\u9700\u8981\u5728\u6574\u4e2a\u52a8\u4f5c\u89c4\u5212\u65f6\u57df\u5185\u63a8\u6f14\u4e16\u754c\u72b6\u6001，\u4e5f\u4e0d\u9700\u8981\u5728\u50cf\u7d20\u7a7a\u95f4\u4e2d\u8fdb\u884c\u672a\u6765\u6e32\u67d3\u3002\u76f8\u53cd，\u5b83\u5728\u4e00\u4e2a\u7d27\u51d1\u7684\u6f5c\u5728\u7a7a\u95f4\u4e2d\u6267\u884c\u77ed\u671f\u4e16\u754c\u63a8\u6f14，\u8fd9\u8db3\u4ee5\u652f\u6301\u957f\u671f\u8f68\u8ff9\u751f\u6210\u3002<\/p>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \u3001<\/p>\n\n\u8bba\u6587\u6a21\u5757\u672c\u5730\u4ee3\u7801\u4f4d\u7f6e\u4f5c\u7528<\/tr>\n\n\n\n\n\n\n\n
    Runtime\u5165\u53e3<\/td>\nDAWN\/scripts\/infer_nuscenes_val.py<\/td>\n\u6784\u5efaencoder\u3001predictor\u3001planner\u3001dataloader，\u5e76\u52a0\u8f7dcheckpoint<\/td>\n<\/tr>\n
    \u6a21\u578b\u521d\u59cb\u5316<\/td>\nDAWN\/app\/vjepa_cowa_world_model\/training\/models.py<\/td>\n\u521d\u59cb\u5316V-JEPA encoder\u3001TokenAE\u3001World Predictor\u3001Diffusion Planner<\/td>\n<\/tr>\n
    Auto-Encoder Resampler<\/td>\nDAWN\/app\/vjepa_cowa_world_model\/models\/token_ae.py<\/td>\n\u5c06\u5bc6\u96c6ViT tokens\u538b\u7f29\u4e3a\u5c11\u91cflatent world tokens<\/td>\n<\/tr>\n
    World Predictor<\/td>\nDAWN\/src\/models\/ac_predictor.py<\/td>\nVisionTransformerPredictorAC，\u52a8\u4f5c\u6761\u4ef6\u7684\u672a\u6765latent\u9884\u6d4b\u5668<\/td>\n<\/tr>\n
    Action Denoiser<\/td>\nDAWN\/app\/vjepa_cowa_world_model\/models\/diffusion_planner.py<\/td>\nDiffusionPlanner \/ TrajectoryDiT，\u57fa\u4e8elatent world tokens\u751f\u6210\u8f68\u8ff9<\/td>\n<\/tr>\n
    \u9a8c\u8bc1\u95ed\u73af<\/td>\nDAWN\/app\/vjepa_cowa_world_model\/val_command.py<\/td>\n\u6267\u884cencoder -> predictor rollout -> planner -> metrics<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    3.2 \u89c6\u89c9\u7f16\u7801\u4e0e\u6f5c\u5728\u538b\u7f29<\/h4>\n

    \u7ed9\u5b9a\u5f53\u524d\u89c2\u6d4b <\/p>\n

    o<\/p>\n

    o<\/p>\n

    <\/span><\/span>o<\/span><\/span><\/span><\/span><\/span>，Student Vision-Encoder\u63d0\u53d6\u5bc6\u96c6\u89c6\u89c9tokens：<\/p>\n

    <\/p>\n

    u<\/p>\n

    =<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    o<\/p>\n

    )<\/p>\n

    u=E_{\\\\mathrm{stu}}(o) <\/p>\n

    <\/span><\/span>u<\/span><\/span>=<\/span><\/span><\/span><\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u5728\u5b9e\u73b0\u4e2d，Student\u548cTeacher\u5206\u652f\u90fd\u4f7f\u7528V-JEPA 2 Large\u4f5c\u4e3a\u89c6\u89c9\u9aa8\u5e72\u7f51\u7edc\u3002\u7531\u4e8e\u5bc6\u96c6\u7f16\u7801\u5668tokens\u7684\u76f4\u63a5\u63a8\u6f14\u6210\u672c\u8fc7\u9ad8，DAWN\u5f15\u5165\u4e86Auto-Encoder Resampler，\u8fd9\u662f\u4e00\u4e2a\u5728token\u7a7a\u95f4\u4e2d\u64cd\u4f5c\u7684\u5b66\u4e60\u74f6\u9888\u81ea\u7f16\u7801\u5668：<\/p>\n

    <\/p>\n

    z<\/p>\n

    =<\/p>\n

    R<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    u<\/p>\n

    )<\/p>\n

    z=R_{\\\\mathrm{stu}}(u) <\/p>\n

    <\/span><\/span>z<\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>u<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u4ea7\u751f\u4e86\u4e00\u4e2a\u7d27\u51d1\u7684\u6f5c\u5728\u4e16\u754c\u8868\u793a，\u7528\u4e8e\u4e0b\u6e38\u4ea4\u4e92\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d，\u672a\u6765\u89c2\u6d4b <\/p>\n

    o<\/p>\n

    +<\/p>\n

    o^{+}<\/p>\n

    <\/span><\/span>o<\/span><\/span>+<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u901a\u8fc7Teacher Vision-Encoder\u53ca\u5176\u5bf9\u5e94\u7684resampler\u5904\u7406，\u751f\u6210\u76ee\u6807\u672a\u6765\u6f5c\u5728\u8868\u793a：<\/p>\n

    <\/p>\n

    z<\/p>\n

    t<\/p>\n

    a<\/p>\n

    r<\/p>\n

    g<\/p>\n

    e<\/p>\n

    t<\/p>\n

    =<\/p>\n

    R<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    E<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    o<\/p>\n

    +<\/p>\n

    )<\/p>\n

    )<\/p>\n

    z_{\\\\mathrm{target}}=R_{\\\\mathrm{tea}}(E_{\\\\mathrm{tea}}(o^{+})) <\/p>\n

    <\/span><\/span>z<\/span><\/span>target<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span><\/span>+<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>))<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u79cd\u8bbe\u8ba1\u7684\u5173\u952e\u521b\u65b0\u5728\u4e8e：\u5b83\u4e0d\u662f\u5728\u9ad8\u7ef4\u50cf\u7d20\u7a7a\u95f4\u6216\u5bc6\u96c6\u7279\u5f81\u7a7a\u95f4\u4e2d\u8fdb\u884c\u4e16\u754c\u5efa\u6a21，\u800c\u662f\u5b66\u4e60\u4e00\u4e2a\u6781\u5176\u7d27\u51d1\u7684\u6f5c\u5728\u8868\u793a（\u4f8b\u5982，\u4ec516\u4e2atokens），\u8fd9\u4e2a\u8868\u793a\u4fdd\u7559\u4e86\u52a8\u4f5c\u89c4\u5212\u6240\u9700\u7684\u5173\u952e\u573a\u666f\u4fe1\u606f，\u540c\u65f6\u5927\u5e45\u964d\u4f4e\u4e86\u63a8\u6f14\u7684\u8ba1\u7b97\u6210\u672c\u3002<\/p>\n

    \u5b9e\u73b0\u7ec6\u8282\u663e\u793a，Resampler\u4f7f\u752816\u4e2a\u6ce8\u610f\u529b\u5934，4\u5c42\u7f16\u7801\u5668\u548c2\u5c42\u89e3\u7801\u5668\u67b6\u6784，MLP\u6bd4\u7387\u4e3a4.0，\u5e76\u91c7\u7528\u53ef\u5b66\u4e60\u7684\u4f4d\u7f6e\u5d4c\u5165\u3002\u8fd9\u79cd\u914d\u7f6e\u5728\u8868\u8fbe\u80fd\u529b\u548c\u8ba1\u7b97\u6548\u7387\u4e4b\u95f4\u53d6\u5f97\u4e86\u826f\u597d\u5e73\u8861\u3002<\/p>\n

    \u672c\u5730\u5b9e\u73b0\u91cc，TokenAE\u7684\u63a8\u7406\u538b\u7f29\u63a5\u53e3\u5728 DAWN\/app\/vjepa_cowa_world_model\/models\/token_ae.py，\u9a8c\u8bc1\u6d41\u7a0b\u91cc\u7684\u5b9e\u9645\u8c03\u7528\u5728 DAWN\/app\/vjepa_cowa_world_model\/val_command.py\u3002<\/p>\n

    \u6765\u6e90：DAWN\/app\/vjepa_cowa_world_model\/models\/token_ae.py<\/p>\n

    def<\/span> compress<\/span>(<\/span>self,<\/span> x:<\/span> torch.<\/span>Tensor,<\/span> num_frames:<\/span> int<\/span>)<\/span> –<\/span>><\/span> torch.<\/span>Tensor:<\/span>
    \n """Compress tokens for inference (no reconstruction, no loss).<\/p>\n

    Parameters
    \n ———-
    \n x : [B, T * tokens_per_frame, D]
    \n num_frames : int<\/p>\n

    Returns
    \n ——-
    \n z : [B, T * num_latent_tokens, D]
    \n """<\/span>
    \n return<\/span> self.<\/span>encode(<\/span>x,<\/span> num_frames)<\/span><\/p>\n

    \u6765\u6e90：DAWN\/app\/vjepa_cowa_world_model\/val_command.py<\/p>\n

    def<\/span> _compress_tokens_with_token_ae<\/span>(<\/span>token_ae,<\/span> tokens:<\/span> torch.<\/span>Tensor,<\/span> num_frames:<\/span> int<\/span>)<\/span> –<\/span>><\/span> torch.<\/span>Tensor:<\/span>
    \n """Apply frozen Token AE compression on concatenated frame tokens."""<\/span>
    \n ae_tokens_per_frame =<\/span> int<\/span>(<\/span>getattr<\/span>(<\/span>token_ae,<\/span> "tokens_per_frame"<\/span>)<\/span>)<\/span>
    \n expected_tokens =<\/span> int<\/span>(<\/span>num_frames)<\/span> *<\/span> ae_tokens_per_frame
    \n if<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> !=<\/span> expected_tokens:<\/span>
    \n if<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> %<\/span> ae_tokens_per_frame !=<\/span> 0<\/span>:<\/span>
    \n raise<\/span> ValueError(<\/span>
    \n "Cannot infer TokenAE frame count: "<\/span>
    \n f"tokens=<\/span>{<\/span>tokens.<\/span>size(<\/span>1<\/span>)<\/span>}<\/span><\/span>, num_frames=<\/span>{<\/span>num_frames}<\/span><\/span>, ae_tokens_per_frame=<\/span>{<\/span>ae_tokens_per_frame}<\/span><\/span>"<\/span><\/span>
    \n )<\/span>
    \n num_frames =<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> \/\/<\/span> ae_tokens_per_frame
    \n return<\/span> token_ae.<\/span>encode(<\/span>tokens,<\/span> num_frames=<\/span>num_frames)<\/span><\/p>\n

    3.3 \u9012\u5f52\u4ea4\u4e92\u673a\u5236：\u4e16\u754c\u9884\u6d4b\u4e0e\u52a8\u4f5c\u53bb\u566a\u7684\u8026\u5408<\/h4>\n

    DAWN\u7684\u6838\u5fc3\u521b\u65b0\u5728\u4e8eWorld Predictor\u548cWorld-Conditioned Action Denoiser\u4e4b\u95f4\u7684\u9012\u5f52\u4ea4\u4e92\u3002World Predictor\u5b9e\u73b0\u4e3a\u4e00\u4e2a\u56e0\u679cTransformer，\u4ece\u5f53\u524d\u6f5c\u5728\u4e0a\u4e0b\u6587\u548c\u5f53\u524d\u52a8\u4f5c\u5047\u8bbe\u9884\u6d4b\u672a\u6765\u6f5c\u5728\u4e16\u754ctokens\u3002World-Conditioned Action Denoiser\u5b9e\u73b0\u4e3aDiT（Diffusion Transformer），\u5b83\u57fa\u4e8e\u6f5c\u5728\u4e0a\u4e0b\u6587\u548c\u9884\u6d4b\u7684\u672a\u6765\u4e16\u754c\u5bf9\u52a8\u4f5ctokens\u8fdb\u884c\u53bb\u566a\u3002<\/p>\n

    \u8bbe\u6761\u4ef6tokens c\u5305\u542b\u81ea\u8f66\u72b6\u6001\u548c\u9ad8\u7ea7\u52a8\u4f5c\u6216\u8def\u7ebftokens\u3002Action Denoiser\u989d\u5916\u63a5\u6536\u89d2\u8272\u7279\u5b9a\u67e5\u8be2，\u6307\u793a\u5b83\u662f\u751f\u6210\u521d\u59cb\u63d0\u8bae\u8fd8\u662f\u4f7f\u7528\u9884\u6d4b\u5668\u63a8\u6f14\u6765\u7ec6\u5316\u52a8\u4f5c\u3002DAWN\u6267\u884c\u4ee5\u4e0b\u9012\u5f52\u8fc7\u7a0b：<\/p>\n

    <\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    0<\/p>\n

    )<\/p>\n

    =<\/p>\n

    G<\/p>\n

    \u03d5<\/p>\n

    (<\/p>\n

    q<\/p>\n

    p<\/p>\n

    r<\/p>\n

    o<\/p>\n

    p<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    )<\/p>\n

    a_{1:H}^{(0)}=G_{\\\\phi}(q_{\\\\mathrm{prop}},c,z) <\/p>\n

    <\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>0<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>G<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>q<\/span><\/span>prop<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>z<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    z<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    =<\/p>\n

    P<\/p>\n

    \u03b8<\/p>\n

    (<\/p>\n

    z<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    )<\/p>\n

    z_{\\\\mathrm{future}}^{(r)}=P_{\\\\theta}(z,c,a_{1:H}^{(r)}) <\/p>\n

    <\/span><\/span>z<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>P<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>z<\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    +<\/p>\n

    1<\/p>\n

    )<\/p>\n

    =<\/p>\n

    G<\/p>\n

    \u03d5<\/p>\n

    (<\/p>\n

    q<\/p>\n

    r<\/p>\n

    e<\/p>\n

    f<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    )<\/p>\n

    a_{1:H}^{(r+1)} =G_{\\\\phi}(q_{\\\\mathrm{ref}}^{(r)},c,z_{\\\\mathrm{future}}^{(r)},a_{1:H}^{(r)}) <\/p>\n

    <\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>+<\/span>1<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>G<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>q<\/span><\/span>ref<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>z<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u91cc\u7684\u5173\u952e\u8bbe\u8ba1\u662f：<\/p>\n

  • \n

    q<\/p>\n

    p<\/p>\n

    r<\/p>\n

    o<\/p>\n

    p<\/p>\n

    q_{\\\\mathrm{prop}}<\/p>\n

    <\/span><\/span>q<\/span><\/span>prop<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u548c <\/p>\n

    q<\/p>\n

    r<\/p>\n

    e<\/p>\n

    f<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    q_{\\\\mathrm{ref}}^{(r)}<\/p>\n

    <\/span><\/span>q<\/span><\/span>ref<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u662f\u89d2\u8272\u7279\u5b9a\u7684\u67e5\u8be2\u5d4c\u5165，\u7528\u4e8e\u63d0\u8bae\u751f\u6210\u548c\u7ec6\u5316<\/li>\n

  • \u53bb\u566a\u5668\u6743\u91cd\u5728\u4e24\u4e2a\u89d2\u8272\u4e4b\u95f4\u5171\u4eab，\u4ec5\u8f93\u5165\u6e90\u548c\u67e5\u8be2\u5d4c\u5165\u4e0d\u540c<\/li>\n
  • \u9884\u6d4b\u7684\u672a\u6765\u4e16\u754c \n

    z<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    z_{\\\\mathrm{future}}^{(r)}<\/p>\n

    <\/span><\/span>z<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u6761\u4ef6\u52a8\u4f5c\u53bb\u566a<\/li>\n

  • \u53bb\u566a\u7684\u52a8\u4f5c\u5047\u8bbe \n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    a_{1:H}^{(r)}<\/p>\n

    <\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span> \u88ab\u53cd\u9988\u4ee5\u66f4\u65b0\u4e16\u754c\u9884\u6d4b<\/li>\n

    \u8fd9\u79cd\u8bbe\u8ba1\u7684\u6df1\u523b\u4e4b\u5904\u5728\u4e8e：\u5b83\u5728\u8bad\u7ec3\u548c\u63a8\u7406\u65f6\u90fd\u4fdd\u6301\u76f8\u540c\u7684\u4ea4\u4e92\u6a21\u5f0f\u3002\u5728\u8bad\u7ec3\u9636\u6bb5，\u6a21\u578b\u5b66\u4e60\u4e24\u4e2a\u89d2\u8272：\u9996\u5148\u4eceresampler\u6f5c\u5728\u4e0a\u4e0b\u6587\u751f\u6210\u521d\u59cb\u63d0\u8bae，\u7136\u540e\u57fa\u4e8e\u9884\u6d4b\u5668\u63a8\u6f14\u7ec6\u5316\u52a8\u4f5c\u3002\u4e0d\u540c\u7684\u67e5\u8be2\u548c\u6e90\u5d4c\u5165\u6307\u5b9a\u53bb\u566a\u5668\u5f53\u524d\u64cd\u4f5c\u7684\u662f\u63d0\u8bae\u8fd8\u662f\u4ea4\u4e92\u5f0f\u7ec6\u5316\u89d2\u8272\u3002<\/p>\n

    \u4ee3\u7801\u5b9e\u73b0\u4e2d，World Predictor\u4f7f\u752812\u5c42Transformer，\u5d4c\u5165\u7ef4\u5ea6384，12\u4e2a\u6ce8\u610f\u529b\u5934，RoPE\u4f4d\u7f6e\u7f16\u7801，\u4ee5\u53ca\u6fc0\u6d3b\u68c0\u67e5\u70b9\u4ee5\u8282\u7701\u5185\u5b58\u3002Action Denoiser\u91c7\u7528DiT-style\u9aa8\u5e72，\u9690\u85cf\u7ef4\u5ea6384，12\u5c42，12\u4e2a\u6ce8\u610f\u529b\u5934，MLP\u6bd4\u73874.0\u3002\u4e0e\u521d\u59cb\u6269\u6563\u89c4\u5212\u5668\u5757\u76f8\u6bd4，DiT\u5757\u66f4\u7d27\u5bc6\u5730\u9075\u5faa\u539f\u59cbadaLN-Zero\u8bbe\u8ba1：\u65f6\u95f4\u6b65\/\u72b6\u6001\u6761\u4ef6\u5411\u91cf\u4e0d\u4ec5\u8c03\u5236\u81ea\u6ce8\u610f\u529b\u548cMLP\u5206\u652f，\u8fd8\u8c03\u5236\u4ea4\u53c9\u6ce8\u610f\u529b\u5206\u652f\u5230\u6f5c\u5728\u4e16\u754ctokens\u3002<\/p>\n

    VisionTransformerPredictorAC\u5728forward\u91cc\u663e\u5f0f\u63a5\u6536\u89c6\u89c9tokens\u3001\u52a8\u4f5c\u3001\u72b6\u6001\u548c\u53ef\u9009\u5916\u53c2；\u8fd9\u5bf9\u5e94\u8bba\u6587\u4e2d\u7684 <\/p>\n

    P<\/p>\n

    \u03b8<\/p>\n

    (<\/p>\n

    z<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    )<\/p>\n

    P_{\\\\theta}(z,c,a)<\/p>\n

    <\/span><\/span>P<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>z<\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>a<\/span>)<\/span><\/span><\/span><\/span><\/span>\u3002DiffusionPlanner.forward\u63a5\u6536\u9884\u6d4b\u5668\u8f93\u51fa\u7684 z_ar \u548c\u81ea\u8f66\u72b6\u6001 status_feature，\u8fd9\u5bf9\u5e94\u4e16\u754c\u6761\u4ef6\u52a8\u4f5c\u53bb\u566a\u5668\u3002<\/p>\n

    \u6765\u6e90：DAWN\/src\/models\/ac_predictor.py<\/p>\n

    # Fwd prop<\/span>
    \n for<\/span> i,<\/span> blk in<\/span> enumerate<\/span>(<\/span>self.<\/span>predictor_blocks)<\/span>:<\/span>
    \n if<\/span> self.<\/span>use_activation_checkpointing:<\/span>
    \n x =<\/span> torch.<\/span>utils.<\/span>checkpoint.<\/span>checkpoint(<\/span>
    \n blk,<\/span>
    \n x,<\/span>
    \n mask=<\/span>None<\/span>,<\/span>
    \n attn_mask=<\/span>attn_mask,<\/span>
    \n T=<\/span>T,<\/span>
    \n H=<\/span>self.<\/span>grid_height,<\/span>
    \n W=<\/span>self.<\/span>grid_width,<\/span>
    \n action_tokens=<\/span>cond_tokens,<\/span>
    \n use_reentrant=<\/span>False<\/span>,<\/span>
    \n )<\/span>
    \n else<\/span>:<\/span>
    \n x =<\/span> blk(<\/span>
    \n x,<\/span>
    \n mask=<\/span>None<\/span>,<\/span>
    \n attn_mask=<\/span>attn_mask,<\/span>
    \n T=<\/span>T,<\/span>
    \n H=<\/span>self.<\/span>grid_height,<\/span>
    \n W=<\/span>self.<\/span>grid_width,<\/span>
    \n action_tokens=<\/span>cond_tokens,<\/span>
    \n )<\/span><\/p>\n

    # Split out action and frame tokens<\/span>
    \n x =<\/span> x.<\/span>view(<\/span>B,<\/span> T,<\/span> cond_tokens +<\/span> self.<\/span>grid_height *<\/span> self.<\/span>grid_width,<\/span> D)<\/span> # [B, T, K+H*W, D]<\/span>
    \n x =<\/span> x[<\/span>:<\/span>,<\/span> :<\/span>,<\/span> cond_tokens:<\/span>,<\/span> :<\/span>]<\/span>.<\/span>flatten(<\/span>1<\/span>,<\/span> 2<\/span>)<\/span><\/p>\n

    x =<\/span> self.<\/span>predictor_norm(<\/span>x)<\/span>
    \n x =<\/span> self.<\/span>predictor_proj(<\/span>x)<\/span><\/p>\n

    return<\/span> x<\/p>\n

    \u6765\u6e90：DAWN\/app\/vjepa_cowa_world_model\/models\/diffusion_planner.py<\/p>\n

    def<\/span> forward<\/span>(<\/span>
    \n self,<\/span>
    \n z_ar:<\/span> torch.<\/span>Tensor,<\/span>
    \n status_feature:<\/span> torch.<\/span>Tensor,<\/span>
    \n z_context:<\/span> Optional[<\/span>torch.<\/span>Tensor]<\/span> =<\/span> None<\/span>,<\/span>
    \n z_observed:<\/span> Optional[<\/span>torch.<\/span>Tensor]<\/span> =<\/span> None<\/span>,<\/span>
    \n action_history:<\/span> Optional[<\/span>torch.<\/span>Tensor]<\/span> =<\/span> None<\/span>,<\/span>
    \n gt_trajectory:<\/span> Optional[<\/span>torch.<\/span>Tensor]<\/span> =<\/span> None<\/span>,<\/span>
    \n anchor_state:<\/span> Optional[<\/span>torch.<\/span>Tensor]<\/span> =<\/span> None<\/span>,<\/span>
    \n )<\/span> –<\/span>><\/span> Dict[<\/span>str<\/span>,<\/span> torch.<\/span>Tensor]<\/span>:<\/span>
    \n """
    \n Forward pass.<\/p>\n

    Args:
    \n z_ar: [B, T*P, encoder_dim] \u2014 predictor output tokens
    \n status_feature: [B, status_dim] \u2014 ego status
    \n z_context: [B, P, encoder_dim] \u2014 optional first-frame encoder tokens
    \n z_observed: [B, T_obs*P, encoder_dim] \u2014 optional observed-frame encoder tokens
    \n action_history: [B, T_obs, action_history_dim] \u2014 optional observed history tokens
    \n gt_trajectory: [B, num_poses, 6] \u2014 ground truth in 6-dim format
    \n (x, y, vx, vy, cos_yaw, sin_yaw). Only needed for training.
    \n anchor_state: [B, 6] or [B, 1, 6] \u2014 current ego state for anchor frame.
    \n Only used when ``use_anchor_frame=True``. If *None*, a
    \n default zero-origin anchor is used.<\/p>\n

    Returns:
    \n Training: {"loss", "reg_loss", "conf_loss", "cover_loss"}
    \n Inference: {"trajectories": [B, K, num_poses, 3], "confidences": [B, K]}
    \n """<\/span>
    \n # Prepare conditioning<\/span>
    \n context_tokens =<\/span> self.<\/span>_prepare_context(<\/span>z_ar,<\/span> z_context,<\/span> z_observed,<\/span> action_history)<\/span> # [B, N, hidden_dim]<\/span>
    \n status_emb =<\/span> self.<\/span>_prepare_status(<\/span>status_feature)<\/span> # [B, hidden_dim]<\/span><\/p>\n

    if<\/span> gt_trajectory is<\/span> not<\/span> None<\/span> and<\/span> self.<\/span>training:<\/span>
    \n return<\/span> self.<\/span>_training_forward(<\/span>context_tokens,<\/span> status_emb,<\/span> gt_trajectory,<\/span> anchor_state)<\/span>
    \n else<\/span>:<\/span>
    \n return<\/span> self.<\/span>_inference_forward(<\/span>context_tokens,<\/span> status_emb,<\/span> anchor_state)<\/span><\/p>\n

    \"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"<\/p>\n

    4. \u8bad\u7ec3\u6d41\u7a0b：\u56db\u9636\u6bb5\u6e10\u8fdb\u5f0f\u5b66\u4e60\u7b56\u7565<\/h3>\n

    4.1 \u9636\u6bb5\u4e00：\u5927\u89c4\u6a21\u89c6\u89c9\u9884\u8bad\u7ec3<\/h4>\n

    DAWN\u91c7\u7528\u56db\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565，\u6bcf\u4e2a\u9636\u6bb5\u90fd\u6709\u660e\u786e\u7684\u5b66\u4e60\u76ee\u6807\u3002\u7b2c\u4e00\u9636\u6bb5\u5728\u5927\u89c4\u6a21\u9a7e\u9a76\u89c6\u9891\u6570\u636e\u4e0a\u9884\u8bad\u7ec3Student Vision-Encoder，\u5305\u62ecOpenScene\u3001DrivingDojo\u548cCoVLA\u7b49\u6570\u636e\u96c6\u3002\u6240\u6709\u6570\u636e\u96c6\u88ab\u8f6c\u6362\u4e3a\u7edf\u4e00\u7684\u89c6\u9891\u683c\u5f0f，\u4f7f\u7528\u6ed1\u52a8\u7a97\u53e3\u91c7\u6837\u3002<\/p>\n

    \u9884\u8bad\u7ec3\u5728256\u00d7512\u5206\u8fa8\u7387\u548c2Hz\u5e27\u7387\u4e0b\u8fdb\u884c\u3002\u8fd9\u4e00\u9636\u6bb5\u7684\u5173\u952e\u662f\u5efa\u7acb\u5f3a\u5927\u7684\u89c6\u89c9\u5148\u9a8c，\u4f7f\u7f16\u7801\u5668\u80fd\u591f\u4ece\u539f\u59cb\u50cf\u7d20\u4e2d\u63d0\u53d6\u9a7e\u9a76\u573a\u666f\u7684\u8bed\u4e49\u4fe1\u606f\u3002\u9884\u8bad\u7ec3\u76ee\u6807\u901a\u5e38\u662f\u89c6\u9891\u7ea7\u522b\u7684\u81ea\u76d1\u7763\u4efb\u52a1，\u5982V-JEPA\u4e2d\u7684masked prediction，\u4f7f\u6a21\u578b\u5b66\u4e60\u573a\u666f\u7684\u65f6\u7a7a\u7ed3\u6784\u548c\u52a8\u6001\u7279\u6027\u3002<\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u53ef\u4ee5\u6982\u62ec\u4e3a：<\/p>\n

    <\/p>\n

    u<\/p>\n

    =<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    o<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    L<\/p>\n

    s<\/p>\n

    s<\/p>\n

    l<\/p>\n

    =<\/p>\n

    \u2113<\/p>\n

    s<\/p>\n

    s<\/p>\n

    l<\/p>\n

    (<\/p>\n

    u<\/p>\n

    ,<\/p>\n

    m<\/p>\n

    a<\/p>\n

    s<\/p>\n

    k<\/p>\n

    e<\/p>\n

    d<\/p>\n

    \u00a0<\/p>\n

    t<\/p>\n

    a<\/p>\n

    r<\/p>\n

    g<\/p>\n

    e<\/p>\n

    t<\/p>\n

    s<\/p>\n

    )<\/p>\n

    u=E_{\\\\mathrm{stu}}(o),\\\\qquad \\\\mathcal{L}_{\\\\mathrm{ssl}}=\\\\ell_{\\\\mathrm{ssl}}(u,\\\\mathrm{masked\\\\ targets}) <\/p>\n

    <\/span><\/span>u<\/span><\/span>=<\/span><\/span><\/span><\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>)<\/span>,<\/span><\/span><\/span>L<\/span><\/span>ssl<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u2113<\/span><\/span>ssl<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>u<\/span>,<\/span><\/span>masked<\/span>\u00a0<\/span>targets<\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    Student Encoder\u901a\u8fc7\u81ea\u76d1\u7763\u635f\u5931\u66f4\u65b0，Teacher Encoder\u901a\u8fc7EMA\u540c\u6b65：<\/p>\n

    <\/p>\n

    E<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    \u2190<\/p>\n

    E<\/p>\n

    M<\/p>\n

    A<\/p>\n

    (<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    ,<\/p>\n

    E<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    )<\/p>\n

    E_{\\\\mathrm{tea}}\\\\leftarrow \\\\mathrm{EMA}(E_{\\\\mathrm{stu}},E_{\\\\mathrm{tea}}) <\/p>\n

    <\/span><\/span>E<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>\u2190<\/span><\/span><\/span><\/span>EMA<\/span><\/span>(<\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>E<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u7684\u8bad\u7ec3\u6570\u636e\u89c4\u6a21\u901a\u5e38\u8fbe\u5230\u6570\u767e\u4e07\u751a\u81f3\u4e0a\u5343\u4e07\u89c6\u9891\u7247\u6bb5，\u786e\u4fdd\u7f16\u7801\u5668\u80fd\u591f\u6cdb\u5316\u5230\u5404\u79cd\u9a7e\u9a76\u573a\u666f\u3001\u5929\u6c14\u6761\u4ef6\u548c\u5149\u7167\u53d8\u5316\u3002<\/p>\n

    DAWN\/app\/vjepa_cowa_world_model\/training\/models.py<\/p>\n

    def<\/span> init_encoder<\/span>(<\/span>
    \n config:<\/span> TrainingConfig,<\/span>
    \n device:<\/span> torch.<\/span>device,<\/span>
    \n)<\/span> –<\/span>><\/span> Tuple[<\/span>nn.<\/span>Module,<\/span> nn.<\/span>Module]<\/span>:<\/span>
    \n """
    \n \u521d\u59cb\u5316 encoder \u548c target_encoder<\/p>\n

    \u6839\u636e config.model.backbone \u9009\u62e9 V-JEPA 2 \u6216 V-JEPA 2.1 \u7f16\u7801\u5668\u3002<\/p>\n

    Args:
    \n config: \u8bad\u7ec3\u914d\u7f6e
    \n device: \u8bbe\u5907<\/p>\n

    Returns:
    \n Tuple[nn.Module, nn.Module]: (encoder, target_encoder)
    \n """<\/span>
    \n encoder =<\/span> init_context_encoder(<\/span>config,<\/span> device)<\/span><\/p>\n

    target_encoder =<\/span> copy.<\/span>deepcopy(<\/span>encoder)<\/span>
    \n logger.<\/span>info(<\/span>"end init encoder"<\/span>)<\/span><\/p>\n

    # \u6253\u5370\u53c2\u6570\u91cf（\u4ec5\u4e3b\u8fdb\u7a0b）<\/span>
    \n encoder_params =<\/span> sum<\/span>(<\/span>p.<\/span>numel(<\/span>)<\/span> for<\/span> p in<\/span> encoder.<\/span>parameters(<\/span>)<\/span>)<\/span>
    \n target_encoder_params =<\/span> sum<\/span>(<\/span>p.<\/span>numel(<\/span>)<\/span> for<\/span> p in<\/span> target_encoder.<\/span>parameters(<\/span>)<\/span>)<\/span>
    \n if<\/span> _is_main_process(<\/span>)<\/span>:<\/span>
    \n logger.<\/span>info(<\/span>f"init encoder_params: <\/span>{<\/span>encoder_params \/<\/span> 1e6<\/span>:<\/span>>8.2f<\/span>}<\/span><\/span>M"<\/span><\/span>)<\/span>
    \n logger.<\/span>info(<\/span>f"init target_encoder_params: <\/span>{<\/span>target_encoder_params \/<\/span> 1e6<\/span>:<\/span>>8.2f<\/span>}<\/span><\/span>M"<\/span><\/span>)<\/span><\/p>\n

    return<\/span> encoder,<\/span> target_encoder<\/p>\n

    DAWN\/app\/vjepa_cowa_world_model\/training\/ema.py<\/p>\n

    @torch.<\/span>no_grad<\/span>(<\/span>)<\/span>
    \n def<\/span> update_ema<\/span>(<\/span>m:<\/span> float<\/span>)<\/span> –<\/span>><\/span> None<\/span>:<\/span>
    \n """
    \n \u4f7f\u7528\u52a8\u6001 momentum \u66f4\u65b0 target encoder \u7684\u53c2\u6570
    \n \u91c7\u7528\u9ad8\u6548\u7684\u539f\u5730\u64cd\u4f5c\u65b9\u5f0f<\/p>\n

    Args:
    \n m: \u5f53\u524d\u7684 momentum \u503c (\u4ece momentum_scheduler \u83b7\u53d6)
    \n """<\/span>
    \n # \u6536\u96c6\u6240\u6709\u53c2\u6570<\/span>
    \n params_k =<\/span> [<\/span>]<\/span>
    \n params_q =<\/span> [<\/span>]<\/span>
    \n for<\/span> param_q,<\/span> param_k in<\/span> zip<\/span>(<\/span>encoder.<\/span>parameters(<\/span>)<\/span>,<\/span> target_encoder.<\/span>parameters(<\/span>)<\/span>)<\/span>:<\/span>
    \n params_k.<\/span>append(<\/span>param_k)<\/span>
    \n params_q.<\/span>append(<\/span>param_q)<\/span><\/p>\n

    # \u9ad8\u6548\u7684\u539f\u5730\u64cd\u4f5c<\/span>
    \n torch.<\/span>_foreach_mul_(<\/span>params_k,<\/span> m)<\/span>
    \n torch.<\/span>_foreach_add_(<\/span>params_k,<\/span> params_q,<\/span> alpha=<\/span>1<\/span> –<\/span> m)<\/span><\/p>\n

    4.2 \u9636\u6bb5\u4e8c：Token\u7a7a\u95f4\u74f6\u9888\u5b66\u4e60<\/h4>\n

    \u7b2c\u4e8c\u9636\u6bb5\u5728\u76f8\u540c\u7684\u9884\u8bad\u7ec3\u8bed\u6599\u4e0a\u8bad\u7ec3Auto-Encoder Resampler\u3002\u8fd9\u4e2a\u9636\u6bb5\u4ece\u9884\u8bad\u7ec3\u7684\u7f16\u7801\u5668\u5f00\u59cb，\u5b66\u4e60\u4e00\u4e2a\u7d27\u51d1\u7684token\u7a7a\u95f4\u74f6\u9888，\u5c06\u5bc6\u96c6\u89c6\u89c9tokens\u538b\u7f29\u4e3a\u6f5c\u5728\u4e16\u754ctokens，\u540c\u65f6\u4fdd\u7559\u672a\u6765\u9884\u6d4b\u548c\u52a8\u4f5c\u751f\u6210\u6240\u9700\u7684\u4fe1\u606f\u3002<\/p>\n

    \u5173\u952e\u521b\u65b0\u662f\u540c\u65f6\u8bad\u7ec3\u4e00\u4e2a\u8f85\u52a9\u6269\u6563\u89c4\u5212\u5668\u5934，\u4e0eAction Denoiser\u4f7f\u7528\u76f8\u540c\u7684DiT-style\u914d\u7f6e\u3002\u8fd9\u4e2a\u8f85\u52a9\u5934\u9f13\u52b1\u538b\u7f29tokens\u4fdd\u7559\u52a8\u4f5c\u76f8\u5173\u4fe1\u606f，\u786e\u4fdd\u74f6\u9888\u8868\u793a\u4e0d\u4ec5\u9002\u5408\u4e16\u754c\u5efa\u6a21，\u66f4\u9002\u5408\u4e0b\u6e38\u89c4\u5212\u4efb\u52a1\u3002<\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u53ef\u4ee5\u6982\u62ec\u4e3a：<\/p>\n

    <\/p>\n

    u<\/p>\n

    =<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    o<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    =<\/p>\n

    R<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    u<\/p>\n

    )<\/p>\n

    u=E_{\\\\mathrm{stu}}(o),\\\\qquad z=R_{\\\\mathrm{stu}}(u) <\/p>\n

    <\/span><\/span>u<\/span><\/span>=<\/span><\/span><\/span><\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>)<\/span>,<\/span><\/span><\/span>z<\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>u<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    u<\/p>\n

    ^<\/p>\n

    =<\/p>\n

    R<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    d<\/p>\n

    e<\/p>\n

    c<\/p>\n

    (<\/p>\n

    z<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    L<\/p>\n

    r<\/p>\n

    e<\/p>\n

    c<\/p>\n

    =<\/p>\n

    \u2113<\/p>\n

    r<\/p>\n

    e<\/p>\n

    c<\/p>\n

    (<\/p>\n

    u<\/p>\n

    ^<\/p>\n

    ,<\/p>\n

    u<\/p>\n

    )<\/p>\n

    \\\\hat{u}=R_{\\\\mathrm{stu}}^{\\\\mathrm{dec}}(z),\\\\qquad \\\\mathcal{L}_{\\\\mathrm{rec}}=\\\\ell_{\\\\mathrm{rec}}(\\\\hat{u},u) <\/p>\n

    <\/span><\/span><\/span>u<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>dec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>z<\/span>)<\/span>,<\/span><\/span><\/span>L<\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>\u2113<\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span><\/span>u<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>u<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u4e3a\u4e86\u8ba9\u538b\u7f29\u540e\u7684tokens\u4fdd\u7559\u52a8\u4f5c\u76f8\u5173\u4fe1\u606f，\u8fd8\u5f15\u5165\u8f85\u52a9\u89c4\u5212\u635f\u5931：<\/p>\n

    <\/p>\n

    L<\/p>\n

    A<\/p>\n

    E<\/p>\n

    =<\/p>\n

    L<\/p>\n

    r<\/p>\n

    e<\/p>\n

    c<\/p>\n

    +<\/p>\n

    L<\/p>\n

    a<\/p>\n

    u<\/p>\n

    x<\/p>\n

    _<\/p>\n

    a<\/p>\n

    c<\/p>\n

    t<\/p>\n

    \\\\mathcal{L}_{\\\\mathrm{AE}}=\\\\mathcal{L}_{\\\\mathrm{rec}}+\\\\mathcal{L}_{\\\\mathrm{aux\\\\_act}} <\/p>\n

    <\/span><\/span>L<\/span><\/span>AE<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>L<\/span><\/span>rec<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>L<\/span><\/span>aux_act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u4e00\u8bbe\u8ba1\u7684\u6df1\u523b\u4e4b\u5904\u5728\u4e8e：\u5b83\u4e0d\u662f\u7b80\u5355\u7684\u964d\u7ef4，\u800c\u662f\u5b66\u4e60\u4e00\u4e2a\u4efb\u52a1\u7279\u5b9a\u7684\u538b\u7f29，\u4f7f\u5f97\u7d27\u51d1\u8868\u793a\u65e2\u80fd\u652f\u6301\u672a\u6765\u4e16\u754c\u63a8\u6f14，\u53c8\u80fd\u652f\u6301\u52a8\u4f5c\u89c4\u5212\u3002<\/p>\n

    DAWN\/app\/vjepa_cowa_world_model\/training\/models.py<\/p>\n

    def<\/span> prepare_runtime_tokens<\/span>(<\/span>
    \n tokens:<\/span> torch.<\/span>Tensor,<\/span>
    \n num_frames:<\/span> int<\/span>,<\/span>
    \n normalize_reps:<\/span> bool<\/span>,<\/span>
    \n token_ae:<\/span> Optional[<\/span>nn.<\/span>Module]<\/span> =<\/span> None<\/span>,<\/span>
    \n)<\/span> –<\/span>><\/span> torch.<\/span>Tensor:<\/span>
    \n """Compress per-frame tokens with TokenAE if present, then apply runtime normalization."""<\/span>
    \n input_dtype =<\/span> tokens.<\/span>dtype
    \n if<\/span> token_ae is<\/span> not<\/span> None<\/span>:<\/span>
    \n ae_tokens_per_frame =<\/span> int<\/span>(<\/span>getattr<\/span>(<\/span>token_ae,<\/span> "tokens_per_frame"<\/span>)<\/span>)<\/span>
    \n expected_tokens =<\/span> int<\/span>(<\/span>num_frames)<\/span> *<\/span> ae_tokens_per_frame
    \n if<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> !=<\/span> expected_tokens:<\/span>
    \n if<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> %<\/span> ae_tokens_per_frame !=<\/span> 0<\/span>:<\/span>
    \n raise<\/span> ValueError(<\/span>
    \n "Cannot infer TokenAE frame count: "<\/span>
    \n f"tokens=<\/span>{<\/span>tokens.<\/span>size(<\/span>1<\/span>)<\/span>}<\/span><\/span>, num_frames=<\/span>{<\/span>num_frames}<\/span><\/span>, "<\/span><\/span>
    \n f"ae_tokens_per_frame=<\/span>{<\/span>ae_tokens_per_frame}<\/span><\/span>"<\/span><\/span>
    \n )<\/span>
    \n num_frames =<\/span> tokens.<\/span>size(<\/span>1<\/span>)<\/span> \/\/<\/span> ae_tokens_per_frame
    \n token_ae_parameters =<\/span> getattr<\/span>(<\/span>token_ae,<\/span> "parameters"<\/span>,<\/span> None<\/span>)<\/span>
    \n ae_param =<\/span> next<\/span>(<\/span>token_ae_parameters(<\/span>)<\/span>,<\/span> None<\/span>)<\/span> if<\/span> callable<\/span>(<\/span>token_ae_parameters)<\/span> else<\/span> None<\/span>
    \n if<\/span> ae_param is<\/span> not<\/span> None<\/span> and<\/span> tokens.<\/span>is_floating_point(<\/span>)<\/span> and<\/span> tokens.<\/span>dtype !=<\/span> ae_param.<\/span>dtype:<\/span>
    \n tokens =<\/span> tokens.<\/span>to(<\/span>dtype=<\/span>ae_param.<\/span>dtype)<\/span>
    \n tokens =<\/span> token_ae.<\/span>encode(<\/span>tokens,<\/span> num_frames=<\/span>num_frames)<\/span>
    \n if<\/span> tokens.<\/span>is_floating_point(<\/span>)<\/span> and<\/span> tokens.<\/span>dtype !=<\/span> input_dtype:<\/span>
    \n tokens =<\/span> tokens.<\/span>to(<\/span>dtype=<\/span>input_dtype)<\/span>
    \n if<\/span> normalize_reps:<\/span><\/p>\n

    4.3 \u9636\u6bb5\u4e09：World Predictor\u8bad\u7ec3<\/h4>\n

    \u7b2c\u4e09\u9636\u6bb5\u5728\u4e0b\u6e38\u4efb\u52a1\u6570\u636e\u96c6\u5982nuScenes\u548cNAVSIM\u4e0a\u8bad\u7ec3World Predictor\u3002\u6b64\u9636\u6bb5，\u9884\u6d4b\u5668\u5b66\u4e60\u4ece\u9884\u8bad\u7ec3\u7f16\u7801\u5668\u548cresampler\u4ea7\u751f\u7684\u7d27\u51d1\u6f5c\u5728\u4e0a\u4e0b\u6587\u4e2d\u63a8\u6f14\u4efb\u52a1\u76f8\u5173\u7684\u672a\u6765\u6f5c\u5728\u4e16\u754c\u72b6\u6001\u3002<\/p>\n

    World Predictor\u88ab\u8bbe\u8ba1\u4e3a\u56e0\u679cTransformer，\u80fd\u591f\u81ea\u56de\u5f52\u5730\u751f\u6210\u672a\u6765\u7684\u6f5c\u5728tokens\u5e8f\u5217\u3002\u8bad\u7ec3\u76ee\u6807\u662f\u6700\u5c0f\u5316\u9884\u6d4b\u7684\u672a\u6765\u6f5c\u5728\u8868\u793a\u4e0eTeacher\u5206\u652f\u751f\u6210\u7684\u76ee\u6807\u4e4b\u95f4\u7684\u8ddd\u79bb：<\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u53ef\u4ee5\u6982\u62ec\u4e3a：<\/p>\n

    <\/p>\n

    z<\/p>\n

    =<\/p>\n

    R<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    o<\/p>\n

    )<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    t<\/p>\n

    a<\/p>\n

    r<\/p>\n

    g<\/p>\n

    e<\/p>\n

    t<\/p>\n

    =<\/p>\n

    R<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    E<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    o<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    )<\/p>\n

    )<\/p>\n

    z=R_{\\\\mathrm{stu}}(E_{\\\\mathrm{stu}}(o)),\\\\qquad z_{\\\\mathrm{target}}=R_{\\\\mathrm{tea}}(E_{\\\\mathrm{tea}}(o_{\\\\mathrm{future}})) <\/p>\n

    <\/span><\/span>z<\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>))<\/span>,<\/span><\/span><\/span>z<\/span><\/span>target<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>))<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    z<\/p>\n

    ^<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    =<\/p>\n

    P<\/p>\n

    \u03b8<\/p>\n

    (<\/p>\n

    z<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    L<\/p>\n

    W<\/p>\n

    M<\/p>\n

    =<\/p>\n

    d<\/p>\n

    (<\/p>\n

    z<\/p>\n

    ^<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    t<\/p>\n

    a<\/p>\n

    r<\/p>\n

    g<\/p>\n

    e<\/p>\n

    t<\/p>\n

    )<\/p>\n

    \\\\hat{z}_{\\\\mathrm{future}}=P_{\\\\theta}(z,c),\\\\qquad \\\\mathcal{L}_{\\\\mathrm{WM}}=d(\\\\hat{z}_{\\\\mathrm{future}},z_{\\\\mathrm{target}}) <\/p>\n

    <\/span><\/span><\/span>z<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>P<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>z<\/span>,<\/span><\/span>c<\/span>)<\/span>,<\/span><\/span><\/span>L<\/span><\/span>WM<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>d<\/span>(<\/span><\/span>z<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>z<\/span><\/span>target<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u7684\u5173\u952e\u662f\u4f7f\u9884\u6d4b\u5668\u5b66\u4e60\u9a7e\u9a76\u573a\u666f\u7684\u52a8\u6001\u6f14\u5316\u89c4\u5f8b，\u5305\u62ec\u5176\u4ed6\u8f66\u8f86\u7684\u8fd0\u52a8\u6a21\u5f0f\u3001\u4ea4\u901a\u89c4\u5219\u7684\u9690\u5f0f\u7ea6\u675f，\u4ee5\u53ca\u573a\u666f\u51e0\u4f55\u7684\u65f6\u7a7a\u53d8\u5316\u3002<\/p>\n

    DAWN\/app\/vjepa_cowa_world_model\/training\/predictor_loss.py<\/p>\n

    observed_steps =<\/span> int<\/span>(<\/span>num_observed_steps)<\/span> if<\/span> num_observed_steps is<\/span> not<\/span> None<\/span> else<\/span> config.<\/span>train.<\/span>num_observed_frames<\/p>\n

    if<\/span> bool<\/span>(<\/span>getattr<\/span>(<\/span>config.<\/span>train,<\/span> "use_parallel_predictor"<\/span>,<\/span> False<\/span>)<\/span>)<\/span>:<\/span>
    \n future_start_step =<\/span> observed_steps if<\/span> config.<\/span>train.<\/span>predictor_inference_consistent else<\/span> 1<\/span>
    \n future_offset =<\/span> future_start_step *<\/span> tokens_per_frame
    \n jloss =<\/span> loss_fn(<\/span>
    \n z_tf[<\/span>:<\/span>,<\/span> future_offset:<\/span>]<\/span>,<\/span>
    \n h_target,<\/span>
    \n offset=<\/span>future_offset,<\/span>
    \n )<\/span>
    \n if<\/span> config.<\/span>train.<\/span>predictor_use_z_ar_supervision:<\/span>
    \n if<\/span> z_ar is<\/span> None<\/span>:<\/span>
    \n raise<\/span> ValueError(<\/span>"z_ar must be provided when predictor_use_z_ar_supervision=True"<\/span>)<\/span>
    \n sloss =<\/span> loss_fn(<\/span>z_ar,<\/span> h_target,<\/span> offset=<\/span>future_offset)<\/span>
    \n else<\/span>:<\/span>
    \n sloss =<\/span> jloss *<\/span> 0.0<\/span>
    \n jepa_loss =<\/span> jloss +<\/span> sloss
    \n return<\/span> jepa_loss,<\/span> jloss,<\/span> sloss<\/p>\n

    4.4 \u9636\u6bb5\u56db：\u8054\u5408\u4e16\u754c-\u52a8\u4f5c\u8bad\u7ec3<\/h4>\n

    \u6700\u540e\u9636\u6bb5\u4ece\u9636\u6bb5\u4e09\u521d\u59cb\u5316World Predictor，\u9644\u52a0World-Conditioned Action Denoiser\u548cAction Head，\u5728\u76ee\u6807\u6570\u636e\u96c6\u4e0a\u8054\u5408\u8bad\u7ec3\u4e16\u754c\u548c\u52a8\u4f5c\u5206\u652f\u3002\u8fd9\u662fWAIM\u8303\u5f0f\u5b9e\u73b0\u7684\u5173\u952e\u9636\u6bb5\u3002<\/p>\n

    \u5728\u6b64\u9636\u6bb5，\u9884\u6d4b\u5668\u548c\u52a8\u4f5c\u53bb\u566a\u5668\u540c\u65f6\u4f18\u5316\u3002Action Denoiser\u5728\u4e24\u4e2a\u89d2\u8272\u4e0b\u8bad\u7ec3，\u6743\u91cd\u5171\u4eab：\u9996\u5148\u4eceresampler\u6f5c\u5728\u4e0a\u4e0b\u6587\u751f\u6210\u521d\u59cb\u63d0\u8bae，\u7136\u540e\u57fa\u4e8e\u9884\u6d4b\u5668\u63a8\u6f14\u7ec6\u5316\u52a8\u4f5c\u3002\u4e0d\u540c\u7684\u67e5\u8be2\u548c\u6e90\u5d4c\u5165\u6307\u5b9a\u53bb\u566a\u5668\u5f53\u524d\u64cd\u4f5c\u7684\u89d2\u8272\u3002<\/p>\n

    \u8fd9\u4e00\u9636\u6bb5\u53ef\u4ee5\u6982\u62ec\u4e3a：<\/p>\n

    <\/p>\n

    z<\/p>\n

    =<\/p>\n

    R<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    E<\/p>\n

    s<\/p>\n

    t<\/p>\n

    u<\/p>\n

    (<\/p>\n

    o<\/p>\n

    )<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    t<\/p>\n

    a<\/p>\n

    r<\/p>\n

    g<\/p>\n

    e<\/p>\n

    t<\/p>\n

    =<\/p>\n

    R<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    E<\/p>\n

    t<\/p>\n

    e<\/p>\n

    a<\/p>\n

    (<\/p>\n

    o<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    )<\/p>\n

    )<\/p>\n

    z=R_{\\\\mathrm{stu}}(E_{\\\\mathrm{stu}}(o)),\\\\qquad z_{\\\\mathrm{target}}=R_{\\\\mathrm{tea}}(E_{\\\\mathrm{tea}}(o_{\\\\mathrm{future}})) <\/p>\n

    <\/span><\/span>z<\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>stu<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span>))<\/span>,<\/span><\/span><\/span>z<\/span><\/span>target<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>R<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>E<\/span><\/span>tea<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>o<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>))<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    0<\/p>\n

    )<\/p>\n

    =<\/p>\n

    G<\/p>\n

    \u03d5<\/p>\n

    (<\/p>\n

    q<\/p>\n

    p<\/p>\n

    r<\/p>\n

    o<\/p>\n

    p<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    )<\/p>\n

    a_{1:H}^{(0)}=G_{\\\\phi}(q_{\\\\mathrm{prop}},c,z) <\/p>\n

    <\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>0<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>G<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>q<\/span><\/span>prop<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>z<\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u7b2c <\/p>\n

    r<\/p>\n

    r<\/p>\n

    <\/span><\/span>r<\/span><\/span><\/span><\/span><\/span> \u8f6e\u9012\u5f52\u4ea4\u4e92\u4e3a：<\/p>\n

    <\/p>\n

    z<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    =<\/p>\n

    P<\/p>\n

    \u03b8<\/p>\n

    (<\/p>\n

    z<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    )<\/p>\n

    z_{\\\\mathrm{future}}^{(r)}=P_{\\\\theta}(z,c,a_{1:H}^{(r)}) <\/p>\n

    <\/span><\/span>z<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>P<\/span><\/span>\u03b8<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>z<\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    +<\/p>\n

    1<\/p>\n

    )<\/p>\n

    =<\/p>\n

    G<\/p>\n

    \u03d5<\/p>\n

    (<\/p>\n

    q<\/p>\n

    r<\/p>\n

    e<\/p>\n

    f<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    c<\/p>\n

    ,<\/p>\n

    z<\/p>\n

    f<\/p>\n

    u<\/p>\n

    t<\/p>\n

    u<\/p>\n

    r<\/p>\n

    e<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    ,<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    r<\/p>\n

    )<\/p>\n

    )<\/p>\n

    a_{1:H}^{(r+1)} =G_{\\\\phi}(q_{\\\\mathrm{ref}}^{(r)},c,z_{\\\\mathrm{future}}^{(r)},a_{1:H}^{(r)}) <\/p>\n

    <\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>+<\/span>1<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>G<\/span><\/span>\u03d5<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>q<\/span><\/span>ref<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>c<\/span>,<\/span><\/span>z<\/span><\/span>future<\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>,<\/span><\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>r<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u6700\u7ec8\u8f68\u8ff9\u548c\u8054\u5408\u635f\u5931\u4e3a：<\/p>\n

    <\/p>\n

    \u03c4<\/p>\n

    ^<\/p>\n

    =<\/p>\n

    H<\/p>\n

    a<\/p>\n

    c<\/p>\n

    t<\/p>\n

    (<\/p>\n

    a<\/p>\n

    1<\/p>\n

    :<\/p>\n

    H<\/p>\n

    (<\/p>\n

    R<\/p>\n

    )<\/p>\n

    )<\/p>\n

    \\\\hat{\\\\tau}=H_{\\\\mathrm{act}}(a_{1:H}^{(R)}) <\/p>\n

    <\/span><\/span><\/span>\u03c4<\/span><\/span><\/span>^<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>H<\/span><\/span>act<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>(<\/span>a<\/span><\/span>1<\/span>:<\/span>H<\/span><\/span><\/span><\/span><\/span>(<\/span>R<\/span>)<\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>)<\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    <\/p>\n

    L<\/p>\n

    t<\/p>\n

    o<\/p>\n

    t<\/p>\n

    a<\/p>\n

    l<\/p>\n

    =<\/p>\n

    L<\/p>\n

    W<\/p>\n

    M<\/p>\n

    +<\/p>\n

    L<\/p>\n

    p<\/p>\n

    l<\/p>\n

    a<\/p>\n

    n<\/p>\n

    \\\\mathcal{L}_{\\\\mathrm{total}} =\\\\mathcal{L}_{\\\\mathrm{WM}}+\\\\mathcal{L}_{\\\\mathrm{plan}} <\/p>\n

    <\/span><\/span>L<\/span><\/span>total<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>=<\/span><\/span><\/span><\/span>L<\/span><\/span>WM<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span>+<\/span><\/span><\/span><\/span>L<\/span><\/span>plan<\/span><\/span><\/span><\/span><\/span><\/span>\u200b<\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/span><\/p>\n

    \u8bad\u7ec3\u65b9\u6848\u7684\u6838\u5fc3\u5728\u4e8e：\u6a21\u578b\u4e0d\u4ec5\u5b66\u4e60\u9884\u6d4b\u672a\u6765\u4e16\u754c\u548c\u751f\u6210\u52a8\u4f5c，\u66f4\u91cd\u8981\u7684\u662f\u5b66\u4e60\u5982\u4f55\u8ba9\u8fd9\u4e24\u4e2a\u8fc7\u7a0b\u901a\u8fc7\u9012\u5f52\u4ea4\u4e92\u76f8\u4e92\u589e\u5f3a\u3002<\/p>\n

    DAWN\/app\/vjepa_cowa_world_model\/training\/models.py<\/p>\n

    token_ae_enabled =<\/span> bool<\/span>(<\/span>getattr<\/span>(<\/span>config,<\/span> "token_ae"<\/span>,<\/span> None<\/span>)<\/span> and<\/span> config.<\/span>token_ae.<\/span>enabled)<\/span>
    \n if<\/span> token_ae_enabled:<\/span>
    \n effective_tokens_per_frame =<\/span> resolve_effective_tokens_per_frame(<\/span>config)<\/span>
    \n else<\/span>:<\/span>
    \n effective_tokens_per_frame =<\/span> (<\/span>
    \n int<\/span>(<\/span>raw_tokens_per_frame_override)<\/span>
    \n if<\/span> raw_tokens_per_frame_override is<\/span> not<\/span> None<\/span>
    \n else<\/span> resolve_effective_tokens_per_frame(<\/span>config)<\/span>
    \n )<\/span>
    \n runtime_normalize_reps =<\/span> resolve_predictor_runtime_normalize_reps(<\/span>config)<\/span>
    \n token_ae =<\/span> None<\/span><\/p>\n

    if<\/span> token_ae_enabled:<\/span>
    \n token_ae,<\/span> runtime_normalize_reps =<\/span> load_frozen_token_ae(<\/span>
    \n config,<\/span>
    \n device=<\/span>device,<\/span>
    \n encoder_embed_dim=<\/span>encoder_embed_dim,<\/span>
    \n tokens_per_frame=<\/span>raw_tokens_per_frame,<\/span>
    \n normalize_reps=<\/span>runtime_normalize_reps,<\/span>
    \n dtype=<\/span>config.<\/span>dtype,<\/span>
    \n )<\/span>
    \n predictor =<\/span> init_predictor_for_ae(<\/span>
    \n config,<\/span>
    \n device=<\/span>device,<\/span>
    \n encoder_embed_dim=<\/span>encoder_embed_dim,<\/span>
    \n num_latent_tokens=<\/span>effective_tokens_per_frame,<\/span>
    \n latent_grid_size=<\/span>getattr<\/span>(<\/span>token_ae,<\/span> "latent_grid_size"<\/span>,<\/span> config.<\/span>token_ae.<\/span>latent_grid_size)<\/span>,<\/span>
    \n )<\/span>
    \n else<\/span>:<\/span>
    \n predictor =<\/span> init_predictor(<\/span>
    \n config,<\/span>
    \n device,<\/span>
    \n encoder_embed_dim,<\/span>
    \n predictor_img_size_override=<\/span>predictor_img_size_override,<\/span>
    \n )<\/span><\/p>\n

    return<\/span> predictor,<\/span> token_ae,<\/span> effective_tokens_per_frame,<\/span> runtime_normalize_reps<\/p>\n

    \u2026\u8be6\u60c5\u8bf7\u53c2\u7167\u53e4\u6708\u5c45<\/h3>\n","protected":false},"excerpt":{"rendered":"

    1. \u5f15\u8a00：\u81ea\u52a8\u9a7e\u9a76\u4e2d\u7684\u672a\u6765\u9884\u6d4b\u56f0\u5883
    \n\u5728\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u7684\u51b3\u7b56\u8fc7\u7a0b\u4e2d，\u5b58\u5728\u4e00\u4e2a\u672c\u8d28\u6027\u7684\u77db\u76fe：\u672a\u6765\u573a\u666f\u7684\u6f14\u5316\u4f9d\u8d56\u4e8e\u5f53\u524d\u8003\u8651\u7684\u9a7e\u9a76\u52a8\u4f5c，\u800c\u5408\u7406\u7684\u9a7e\u9a76\u52a8\u4f5c\u53c8\u4f9d\u8d56\u4e8e\u5bf9\u672a\u6765\u573a\u666f\u7684\u51c6\u786e\u9884\u6d4b\u3002\u8fd9\u79cd\u76f8\u4e92\u4f9d\u8d56\u7684\u5173\u7cfb\u5728\u4f20\u7edf\u7684World Action Model（\u4e16\u754c\u52a8\u4f5c\u6a21\u578b，\u7b80\u79f0WAM）\u4e2d\u5e76\u672a\u5f97\u5230\u5145\u5206\u4f53\u73b0\u3002\u73b0\u6709\u7684WAM\u8981\u4e48\u91c7\u7528\u5e76\u884c\u5206\u652f\u72ec\u7acb\u9884\u6d4b\u4e16\u754c\u72b6\u6001\u548c\u9a7e\u9a76\u52a8\u4f5c，\u8981\u4e48\u91c7\u7528\u4e25\u683c\u7684\u9884\u6d4b-\u89c4\u5212\u6d41<\/p>\n","protected":false},"author":2,"featured_media":100994,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1666],"topic":[],"class_list":["post-100999","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-1666"],"yoast_head":"\nDAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/100999.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"1. \u5f15\u8a00：\u81ea\u52a8\u9a7e\u9a76\u4e2d\u7684\u672a\u6765\u9884\u6d4b\u56f0\u5883 \u5728\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u7684\u51b3\u7b56\u8fc7\u7a0b\u4e2d，\u5b58\u5728\u4e00\u4e2a\u672c\u8d28\u6027\u7684\u77db\u76fe：\u672a\u6765\u573a\u666f\u7684\u6f14\u5316\u4f9d\u8d56\u4e8e\u5f53\u524d\u8003\u8651\u7684\u9a7e\u9a76\u52a8\u4f5c，\u800c\u5408\u7406\u7684\u9a7e\u9a76\u52a8\u4f5c\u53c8\u4f9d\u8d56\u4e8e\u5bf9\u672a\u6765\u573a\u666f\u7684\u51c6\u786e\u9884\u6d4b\u3002\u8fd9\u79cd\u76f8\u4e92\u4f9d\u8d56\u7684\u5173\u7cfb\u5728\u4f20\u7edf\u7684World Action Model（\u4e16\u754c\u52a8\u4f5c\u6a21\u578b，\u7b80\u79f0WAM）\u4e2d\u5e76\u672a\u5f97\u5230\u5145\u5206\u4f53\u73b0\u3002\u73b0\u6709\u7684WAM\u8981\u4e48\u91c7\u7528\u5e76\u884c\u5206\u652f\u72ec\u7acb\u9884\u6d4b\u4e16\u754c\u72b6\u6001\u548c\u9a7e\u9a76\u52a8\u4f5c，\u8981\u4e48\u91c7\u7528\u4e25\u683c\u7684\u9884\u6d4b-\u89c4\u5212\u6d41\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/100999.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-09-05T14:25:30+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905142528-6a9c2658d0d32.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"14 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/100999.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/100999.html\",\"name\":\"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-09-05T14:25:30+00:00\",\"dateModified\":\"2026-09-05T14:25:30+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/100999.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/100999.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/100999.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/100999.html","og_locale":"zh_CN","og_type":"article","og_title":"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"1. \u5f15\u8a00：\u81ea\u52a8\u9a7e\u9a76\u4e2d\u7684\u672a\u6765\u9884\u6d4b\u56f0\u5883 \u5728\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u7684\u51b3\u7b56\u8fc7\u7a0b\u4e2d，\u5b58\u5728\u4e00\u4e2a\u672c\u8d28\u6027\u7684\u77db\u76fe：\u672a\u6765\u573a\u666f\u7684\u6f14\u5316\u4f9d\u8d56\u4e8e\u5f53\u524d\u8003\u8651\u7684\u9a7e\u9a76\u52a8\u4f5c，\u800c\u5408\u7406\u7684\u9a7e\u9a76\u52a8\u4f5c\u53c8\u4f9d\u8d56\u4e8e\u5bf9\u672a\u6765\u573a\u666f\u7684\u51c6\u786e\u9884\u6d4b\u3002\u8fd9\u79cd\u76f8\u4e92\u4f9d\u8d56\u7684\u5173\u7cfb\u5728\u4f20\u7edf\u7684World Action Model（\u4e16\u754c\u52a8\u4f5c\u6a21\u578b，\u7b80\u79f0WAM）\u4e2d\u5e76\u672a\u5f97\u5230\u5145\u5206\u4f53\u73b0\u3002\u73b0\u6709\u7684WAM\u8981\u4e48\u91c7\u7528\u5e76\u884c\u5206\u652f\u72ec\u7acb\u9884\u6d4b\u4e16\u754c\u72b6\u6001\u548c\u9a7e\u9a76\u52a8\u4f5c，\u8981\u4e48\u91c7\u7528\u4e25\u683c\u7684\u9884\u6d4b-\u89c4\u5212\u6d41","og_url":"https:\/\/www.wsisp.com\/helps\/100999.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-09-05T14:25:30+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2026\/09\/20260905142528-6a9c2658d0d32.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"14 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/100999.html","url":"https:\/\/www.wsisp.com\/helps\/100999.html","name":"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-09-05T14:25:30+00:00","dateModified":"2026-09-05T14:25:30+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/100999.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/100999.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/100999.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"DAWN\uff1a\u4e16\u754c-\u52a8\u4f5c\u4ea4\u4e92\u6a21\u578b\u7684\u9ece\u660e"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/100999","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=100999"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/100999\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/100994"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=100999"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=100999"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=100999"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=100999"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}