{"id":5721,"date":"2025-02-25T13:39:07","date_gmt":"2025-02-25T05:39:07","guid":{"rendered":"https:\/\/nullthought.net\/?p=5721"},"modified":"2025-03-20T11:07:35","modified_gmt":"2025-03-20T03:07:35","slug":"%e5%9c%a8jepa-%ef%bc%88joint-embedding-predictive-architecture%ef%bc%8c%e8%81%94%e5%90%88%e5%b5%8c%e5%85%a5%e9%a2%84%e6%b5%8b%e6%9e%b6%e6%9e%84%ef%bc%89%e6%a8%a1%e5%9e%8b%e7%9a%84%e6%bd%9c%e5%9c%a8","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5721","title":{"rendered":"\u5728JEPA \uff08Joint Embedding Predictive Architecture\uff0c\u8054\u5408\u5d4c\u5165\u9884\u6d4b\u67b6\u6784\uff09\u6a21\u578b\u7684\u6f5c\u5728\u7a7a\u95f4\u4e2d\u8fdb\u884c\u89c4\u5212\uff0c\u53ef\u5728\u65b0\u73af\u5883\u5e03\u5c40\u6cdb\u5316\u3001\u8f68\u8ff9\u62fc\u63a5\u548c\u6570\u636e\u6548\u7387\u65b9\u9762\u83b7\u5f97\u826f\u597d\u6548\u679c"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2502.14819\" target=\"_blank\" rel=\"noreferrer noopener\">Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models<\/a><\/strong>\u4e2d\u7684\u7814\u7a76\u4e2d\uff0c\u4f7f\u7528\u4e86\u7ecf\u8fc7\u65b9\u5dee-\u534f\u65b9\u5dee\u6b63\u5219\u5316\uff08Variance-Covariance regularization\uff09\u8bad\u7ec3\u7684JEPA \uff08Joint Embedding Predictive Architecture\uff0c\u8054\u5408\u5d4c\u5165\u9884\u6d4b\u67b6\u6784\uff09\u6a21\u578b\uff0c\u5e76\u5728\u5176\u6f5c\u5728\u7a7a\u95f4\u4e2d\u8fdb\u884c\u89c4\u5212\u3002\u7ed3\u679c\u8868\u660e\uff0c\u5f53\u6709\u5927\u91cf\u9ad8\u8d28\u91cf\u6570\u636e\u65f6\uff0c\u57fa\u4e8e\u6a21\u578b\u81ea\u7531\u7684\u5f3a\u5316\u5b66\u4e60\u8868\u73b0\u4f18\u8d8a\uff0c\u800c\u57fa\u4e8e\u6a21\u578b\u7684\u89c4\u5212\u65b9\u6cd5\u5728\u65b0\u73af\u5883\u5e03\u5c40\u7684\u6cdb\u5316\u3001\u8f68\u8ff9\u62fc\u63a5\u548c\u6570\u636e\u6548\u7387\u65b9\u9762\u8868\u73b0\u7a81\u51fa\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aVlad Sobal, Wancong Zhang, Kynghyun Cho, Randall Balestriero, Tim G. J. Rudner, Yann LeCun\uff0c\u6765\u81ea\u6765\u81eaNew York University, Genentech, Brown University\u548cMeta\u3002<a href=\"https:\/\/nullthought.net\/?s=%E6%9D%A8%E7%AB%8B%E6%98%86\" target=\"_blank\" rel=\"noreferrer noopener\">\u6768\u7acb\u6606\uff08Yann LeGun\uff09\u6559\u6388<\/a>\u4e5f\u662f\u8bba\u6587\u4f5c\u8005\u4e4b\u4e00\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"924\" height=\"390\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-17.png\" alt=\"\" class=\"wp-image-5722\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-17.png 924w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-17-300x127.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-17-768x324.png 768w\" sizes=\"auto, (max-width: 924px) 100vw, 924px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2502.14819\" target=\"_blank\" rel=\"noreferrer noopener\">Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e00\u3001\u5f15\u8a00<\/h5>\n\n\n\n<p>\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u4e00\u4e2a\u957f\u671f\u76ee\u6807\u662f\u6784\u5efa\u80fd\u591f\u5728\u4e0d\u540c\u73af\u5883\u4e2d\u6267\u884c\u5404\u79cd\u4efb\u52a1\u7684\u667a\u80fd\u4f53\uff0c\u5c24\u5176\u662f\u80fd\u5728\u672a\u66fe\u89c1\u8fc7\u7684\u73af\u5883\u4e2d\u6267\u884c\u4efb\u52a1\u3002\u4f20\u7edf\u4e0a\uff0c\u8fd9\u4e00\u6311\u6218\u901a\u8fc7\u4e24\u79cd\u4e3b\u8981\u65b9\u6cd5\u8fdb\u884c\u89e3\u51b3\uff1a(i) \u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\uff0c\u901a\u8fc7\u4e0e\u73af\u5883\u7684\u4ea4\u4e92\uff0c\u4e0d\u65ad\u8bd5\u9519\u4ee5\u5b66\u4e60\u6700\u4f18\u7b56\u7565\uff1b(ii) \u6700\u4f18\u63a7\u5236\uff0c\u901a\u8fc7\u5b66\u4e60\u6216\u5df2\u77e5\u7684\u52a8\u6001\u6a21\u578b\u6765\u8ba1\u5212\u6700\u4f73\u884c\u52a8\u3002\u7136\u800c\uff0c\u8fd9\u4e24\u79cd\u65b9\u6cd5\u7684\u4f18\u7f3a\u70b9\uff0c\u5728\u667a\u80fd\u4f53\u9700\u4ece\u79bb\u7ebf\u8f68\u8ff9\uff08\u5373\u6ca1\u6709\u5956\u52b1\u6807\u6ce8\u7684\u6570\u636e\uff09\u4e2d\u5b66\u4e60\u65f6\uff0c\u5c1a\u672a\u5f97\u5230\u5145\u5206\u7684\u63a2\u7d22\u3002\u672c\u7814\u7a76\u901a\u8fc7\u7cfb\u7edf\u5730\u5206\u6790\u4e0d\u540c\u7684\u5f3a\u5316\u5b66\u4e60\u548c\u63a7\u5236\u65b9\u6cd5\u5728\u4e0d\u540c\u6570\u636e\u96c6\u8d28\u91cf\u4e0b\u7684\u8868\u73b0\uff0c\u5c1d\u8bd5\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\u3002\u6211\u4eec\u4ece\u5956\u52b1\u6761\u4ef6\u5f3a\u5316\u5b66\u4e60\uff08Goal-conditioned RL\uff09\u3001\u96f6\u6837\u672c\uff08zero-shot\uff09\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u3001\u4ee5\u53ca\u57fa\u4e8e\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\u7684\u6700\u4f18\u63a7\u5236\u65b9\u6cd5\uff08Planning with Latent Dynamics Models, PLDM\uff09\u7b49\u591a\u4e2a\u89d2\u5ea6\u8fdb\u884c\u4e86\u6df1\u5165\u6bd4\u8f83\uff0c\u63a2\u8ba8\u5b83\u4eec\u5728\u4ece\u6ca1\u6709\u5956\u52b1\u6807\u6ce8\u7684\u79bb\u7ebf\u6570\u636e\u4e2d\u5b66\u4e60\u65f6\u7684\u4f18\u52bf\u4e0e\u4e0d\u8db3\u3002<\/p>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728\u6570\u636e\u5145\u8db3\u4e14\u8d28\u91cf\u9ad8\u7684\u60c5\u51b5\u4e0b\uff0c\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u8868\u73b0\u4f18\u8d8a\uff0c\u800c\u57fa\u4e8e\u6a21\u578b\u7684\u89c4\u5212\u65b9\u6cd5\u5219\u5728\u65b0\u73af\u5883\u5e03\u5c40\u7684\u6cdb\u5316\u3001\u8f68\u8ff9\u62fc\u63a5\u4ee5\u53ca\u6570\u636e\u6548\u7387\u65b9\u9762\u5c55\u73b0\u51fa\u660e\u663e\u7684\u4f18\u52bf\u3002\u5c24\u5176\u662fPLDM\uff0c\u5728\u96f6\u6837\u672c\uff08zero-shot\uff09\u4efb\u52a1\u7684\u6cdb\u5316\u80fd\u529b\u4e0a\u8868\u73b0\u7a81\u51fa\uff0c\u80fd\u591f\u6709\u6548\u5e94\u5bf9\u6b21\u4f18\u6570\u636e\u7684\u6311\u6218\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e8c\u3001\u76f8\u5173\u5de5\u4f5c<\/h5>\n\n\n\n<p>\u5728\u6ca1\u6709\u5956\u52b1\u4fe1\u53f7\u7684\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u9886\u57df\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u4e00\u4e9b\u65b9\u6cd5\u6765\u4ece\u79bb\u7ebf\u6570\u636e\u4e2d\u63d0\u53d6\u901a\u7528\u884c\u4e3a\uff0c\u5e76\u5c06\u5176\u5e94\u7528\u4e8e\u4e0b\u6e38\u4efb\u52a1\u3002\u76ee\u6807\u6761\u4ef6\u5f3a\u5316\u5b66\u4e60\uff08Goal-Conditioned RL\uff09\u5c31\u662f\u4e00\u79cd\u5178\u578b\u7684\u7b56\u7565\uff0c\u901a\u8fc7\u5bf9\u76ee\u6807\u7684\u6761\u4ef6\u5316\u6765\u5f15\u5bfc\u5b66\u4e60\u8fc7\u7a0b\uff0c\u589e\u5f3a\u667a\u80fd\u4f53\u5728\u590d\u6742\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\u6cdb\u5316\u80fd\u529b\uff0c\u96f6\u6837\u672c\uff08zero-shot\uff09\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u91c7\u7528\u6f5c\u5728\u4efb\u52a1\u53d8\u91cf\u7684\u65b9\u5f0f\uff0c\u76ee\u6807\u4e0d\u4ec5\u662f\u5b66\u4f1a\u4e00\u4e2a\u56fa\u5b9a\u4efb\u52a1\uff0c\u800c\u662f\u5b66\u4f1a\u9002\u5e94\u65b0\u7684\u4efb\u52a1\u548c\u76ee\u6807\u3002<\/p>\n\n\n\n<p>\u4e0e\u5f3a\u5316\u5b66\u4e60\u7684\u7b56\u7565\u4e0d\u540c\uff0c\u6700\u4f18\u63a7\u5236\u65b9\u6cd5\u91c7\u7528\u5df2\u77e5\u7684\u52a8\u529b\u5b66\u6a21\u578b\u6765\u9009\u62e9\u884c\u52a8\uff0c\u5e76\u6839\u636e\u8be5\u6a21\u578b\u8fdb\u884c\u89c4\u5212\u3002\u5728\u52a8\u529b\u5b66\u6a21\u578b\u65e0\u6cd5\u7cbe\u786e\u63cf\u8ff0\u7684\u60c5\u51b5\u4e0b\uff0c\u8fd1\u5e74\u6765\u8bb8\u591a\u65b9\u6cd5\u5c1d\u8bd5\u5b66\u4e60\u73af\u5883\u7684\u52a8\u6001\u6a21\u578b\uff0c\u5e76\u901a\u8fc7\u8be5\u6a21\u578b\u8fdb\u884c\u89c4\u5212\u3002\u8fd9\u4e9b\u65b9\u6cd5\u4e0d\u4f9d\u8d56\u4e8e\u5956\u52b1\u4fe1\u53f7\uff0c\u800c\u662f\u901a\u8fc7\u5bf9\u73af\u5883\u52a8\u529b\u5b66\u7684\u5efa\u6a21\u6765\u4f18\u5316\u667a\u80fd\u4f53\u7684\u884c\u4e3a\u3002\u8fd1\u5e74\u6765\uff0c\u4e00\u4e9b\u7814\u7a76\u5df2\u7ecf\u8bc1\u660e\uff0c\u57fa\u4e8e\u6a21\u578b\u7684\u63a7\u5236\u65b9\u6cd5\u5728\u67d0\u4e9b\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\uff0c\u5c24\u5176\u662f\u9762\u5bf9\u79bb\u7ebf\u6570\u636e\u65f6\u7684\u8868\u73b0\uff0c\u5177\u6709\u6781\u5927\u7684\u6f5c\u529b\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e09\u3001\u65b9\u6cd5\u6982\u8ff0<\/h5>\n\n\n\n<p>\u672c\u7814\u7a76\u63d0\u51fa\u4e86\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\uff08PLDM\uff09\u8fdb\u884c\u89c4\u5212\u7684\u65b9\u6cd5\uff0c\u5e76\u4e0e\u591a\u79cd\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u8fdb\u884c\u4e86\u6bd4\u8f83\u3002PLDM\u65b9\u6cd5\u7684\u6838\u5fc3\u601d\u60f3\u662f\u901a\u8fc7\u81ea\u76d1\u7763\u5b66\u4e60\uff08SSL\uff09\u76ee\u6807\u8bad\u7ec3\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\u3002\u6211\u4eec\u4e0d\u4f9d\u8d56\u4f20\u7edf\u7684\u56fe\u50cf\u91cd\u5efa\u4efb\u52a1\uff0c\u800c\u662f\u4e13\u6ce8\u4e8e\u901a\u8fc7\u6f5c\u5728\u7a7a\u95f4\u7684\u8868\u793a\u6765\u5b66\u4e60\u52a8\u529b\u5b66\uff0c\u4ece\u800c\u907f\u514d\u4e86\u91cd\u5efa\u8fc7\u7a0b\u4e2d\u7684\u6f5c\u5728\u4fe1\u606f\u4e22\u5931\u95ee\u9898\u3002\u5728\u6d4b\u8bd5\u9636\u6bb5\uff0c\u6211\u4eec\u4f7f\u7528\u89c4\u5212\u4f18\u5316\u76ee\u6807\u4efb\u52a1\uff0c\u901a\u8fc7\u4f18\u5316\u89c4\u5212\u8fc7\u7a0b\u4e2d\u7684\u4ee3\u4ef7\u51fd\u6570\u6765\u8fbe\u5230\u76ee\u6807\u72b6\u6001\u3002<\/p>\n\n\n\n<p>\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0c\u9664\u4e86PLDM\u65b9\u6cd5\u5916\uff0c\u8fd8\u9009\u7528\u4e86\u591a\u4e2a\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\uff0c\u5305\u62ec\u76ee\u6807\u6761\u4ef6\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08GCIQL\uff09\u3001\u5c42\u6b21\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08HIQL\uff09\u3001\u4ee5\u53ca\u96f6\u6837\u672c\uff08zero-shot\uff09\u5b66\u4e60\u65b9\u6cd5\uff08HILP\uff09\u7b49\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5206\u522b\u4ee3\u8868\u4e86\u4e0d\u540c\u7684\u5f3a\u5316\u5b66\u4e60\u601d\u8def\uff0c\u4ece\u76ee\u6807\u5bfc\u5411\u3001\u7b56\u7565\u5b66\u4e60\u5230\u6f5c\u5728\u7a7a\u95f4\u7684\u63a2\u7d22\uff0c\u90fd\u5177\u6709\u4e00\u5b9a\u7684\u5e94\u7528\u4ef7\u503c\u548c\u7814\u7a76\u610f\u4e49\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">\u56db\u3001\u5b9e\u9a8c\u4e0e\u7ed3\u679c<\/h5>\n\n\n\n<p>\u5728\u672c\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\u6765\u8bc4\u4f30\u4e0d\u540c\u65b9\u6cd5\u5728\u4e0d\u540c\u6761\u4ef6\u4e0b\u7684\u8868\u73b0\uff0c\u91cd\u70b9\u8003\u5bdf\u4e86\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6570\u636e\u8d28\u91cf\u8f83\u597d\u7684\u60c5\u51b5\u4e0b\uff0c\u54ea\u4e9b\u65b9\u6cd5\u80fd\u591f\u53d6\u5f97\u6700\u4f73\u8868\u73b0<\/strong><br>\u6211\u4eec\u9996\u5148\u901a\u8fc7\u4e00\u4e2a\u5927\u6570\u636e\u96c6\u8fdb\u884c\u6d4b\u8bd5\uff0c\u5176\u4e2d\u6570\u636e\u7684\u8d28\u91cf\u8f83\u9ad8\uff0c\u72b6\u6001\u8986\u76d6\u9762\u5e7f\uff0c\u8f68\u8ff9\u957f\u5ea6\u5145\u8db3\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6240\u6709\u65b9\u6cd5\u5747\u8fbe\u5230\u4e86\u6700\u4f73\u8868\u73b0\uff0c\u5176\u4e2dHIQL\u548cHILP\u8fbe\u5230\u4e86\u63a5\u8fd1100%\u7684\u6210\u529f\u7387\u3002\u7136\u800c\uff0c\u7531\u4e8ePLDM\u7684\u8bad\u7ec3\u76ee\u6807\u662f\u5b66\u4e60\u52a8\u529b\u5b66\u800c\u975e\u7b56\u7565\uff0c\u56e0\u6b64\u5728\u8fd9\u4e00\u9636\u6bb5PLDM\u7684\u8868\u73b0\u7565\u900a\u4e00\u7b79\u3002<\/li>\n\n\n\n<li><strong>\u5728\u6570\u636e\u7a00\u7f3a\u65f6\uff0c\u54ea\u4e9b\u65b9\u6cd5\u80fd\u66f4\u6709\u6548\u5730\u5b66\u4e60<\/strong><br>\u901a\u8fc7\u63a7\u5236\u6570\u636e\u96c6\u7684\u5927\u5c0f\uff0c\u6211\u4eec\u6d4b\u8bd5\u4e86\u5404\u65b9\u6cd5\u5728\u7a00\u7f3a\u6570\u636e\u73af\u5883\u4e0b\u7684\u8868\u73b0\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cPLDM\u5728\u6570\u636e\u7a00\u7f3a\u65f6\u8868\u73b0\u51fa\u8f83\u5f3a\u7684\u6837\u672c\u6548\u7387\uff0c\u80fd\u591f\u4ece\u5c11\u91cf\u6570\u636e\u4e2d\u6709\u6548\u5b66\u4e60\uff0c\u5e76\u5728\u6d4b\u8bd5\u4e2d\u53d6\u5f97\u826f\u597d\u7ed3\u679c\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cHILP\u65b9\u6cd5\u5728\u6570\u636e\u7a00\u7f3a\u65f6\u7684\u8868\u73b0\u8f83\u5dee\uff0c\u9700\u8981\u66f4\u591a\u7684\u8bad\u7ec3\u6570\u636e\u6765\u8fbe\u5230\u8f83\u597d\u7684\u6548\u679c\u3002<\/li>\n\n\n\n<li><strong>\u80fd\u5426\u4ece\u8f83\u77ed\u7684\u8f68\u8ff9\u4e2d\u5b66\u4e60\u5e76\u62fc\u63a5\u51fa\u5b8c\u6574\u7684\u8f68\u8ff9<\/strong><br>\u5728\u6b64\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u4e86\u957f\u5ea6\u4e0d\u540c\u7684\u8f68\u8ff9\u6570\u636e\u96c6\uff0c\u91cd\u70b9\u8003\u5bdf\u667a\u80fd\u4f53\u5728\u9762\u5bf9\u77ed\u8f68\u8ff9\u6570\u636e\u65f6\u7684\u8868\u73b0\u3002\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c\u57fa\u4e8e\u76ee\u6807\u6761\u4ef6\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u5728\u9047\u5230\u77ed\u8f68\u8ff9\u65f6\u5bb9\u6613\u51fa\u73b0\u76ee\u6807\u5206\u5e03\u5916\u7684\u60c5\u51b5\uff0c\u96be\u4ee5\u4ece\u77ed\u671f\u8f68\u8ff9\u4e2d\u5b66\u4e60\u5230\u5b8c\u6574\u7684\u884c\u4e3a\u5e8f\u5217\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cPLDM\u901a\u8fc7\u5b66\u4e60\u52a8\u529b\u5b66\u6a21\u578b\u80fd\u591f\u5728\u6d4b\u8bd5\u9636\u6bb5\u6709\u6548\u5730\u62fc\u63a5\u8f68\u8ff9\uff0c\u5b8c\u6210\u4efb\u52a1\u3002<\/li>\n\n\n\n<li><strong>\u80fd\u5426\u4ece\u968f\u673a\u7b56\u7565\u7684\u8f68\u8ff9\u4e2d\u5b66\u4e60<\/strong><br>\u5728\u8fd9\u9879\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u751f\u6210\u4e86\u968f\u673a\u7b56\u7565\u7684\u8f68\u8ff9\u6570\u636e\uff0c\u91cd\u70b9\u6d4b\u8bd5\u667a\u80fd\u4f53\u5728\u9762\u5bf9\u566a\u58f0\u8f83\u5927\u7684\u6570\u636e\u65f6\u7684\u5b66\u4e60\u80fd\u529b\u3002\u7ed3\u679c\u8868\u660e\uff0cPLDM\u548cHILP\u65b9\u6cd5\u80fd\u591f\u6709\u6548\u5730\u4ece\u968f\u673a\u7b56\u7565\u7684\u8f68\u8ff9\u4e2d\u5b66\u4e60\uff0c\u800c\u5176\u4ed6\u57fa\u4e8e\u76ee\u6807\u6761\u4ef6\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u5219\u8868\u73b0\u8f83\u5dee\u3002PLDM\u901a\u8fc7\u5b66\u4e60\u73af\u5883\u7684\u52a8\u529b\u5b66\u80fd\u591f\u8f83\u597d\u5730\u5e94\u5bf9\u8fd9\u79cd\u4f4e\u8d28\u91cf\u6570\u636e\uff0cHILP\u5219\u901a\u8fc7\u5b66\u4e60\u6f5c\u5728\u7a7a\u95f4\u7684\u7ed3\u6784\uff0c\u6709\u6548\u5e94\u5bf9\u4e86\u968f\u673a\u8f68\u8ff9\u7684\u6311\u6218\u3002<\/li>\n\n\n\n<li><strong>\u5404\u65b9\u6cd5\u5728\u96f6\u6837\u672c\uff08zero-shot\uff09\u4efb\u52a1\u4e2d\u7684\u6cdb\u5316\u80fd\u529b<\/strong><br>\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u65b0\u7684\u4efb\u52a1\u2014\u2014\u907f\u514d\u8ffd\u9010\u4efb\u52a1\uff0c\u6d4b\u8bd5\u4e0d\u540c\u65b9\u6cd5\u5728\u65b0\u4efb\u52a1\u4e2d\u7684\u6cdb\u5316\u80fd\u529b\u3002\u5b9e\u9a8c\u8868\u660e\uff0cPLDM\u80fd\u591f\u901a\u8fc7\u7b80\u5355\u7684\u8c03\u6574\u4ee3\u4ef7\u51fd\u6570\uff0c\u5b9e\u73b0\u5bf9\u65b0\u4efb\u52a1\u7684\u96f6\u6837\u672c\uff08zero-shot\uff09\u5b66\u4e60\uff0c\u800cHILP\u5219\u96be\u4ee5\u5728\u6ca1\u6709\u989d\u5916\u8bad\u7ec3\u7684\u60c5\u51b5\u4e0b\u5b8c\u6210\u8be5\u4efb\u52a1\u3002<\/li>\n\n\n\n<li><strong>\u5404\u65b9\u6cd5\u5728\u65b0\u73af\u5883\u53d8\u5316\u4e2d\u7684\u6cdb\u5316\u80fd\u529b<\/strong><br>\u5728\u8fd9\u4e2a\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u6d4b\u8bd5\u4e86\u4e0d\u540c\u65b9\u6cd5\u5728\u9762\u5bf9\u65b0\u73af\u5883\u5e03\u5c40\u65f6\u7684\u6cdb\u5316\u80fd\u529b\u3002PLDM\u5728\u8fd9\u79cd\u8bbe\u7f6e\u4e0b\u8868\u73b0\u51fa\u4e86\u6700\u597d\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u5373\u4f7f\u5728\u4ec5\u7528\u4e94\u79cd\u73af\u5883\u5e03\u5c40\u8fdb\u884c\u8bad\u7ec3\u7684\u60c5\u51b5\u4e0b\uff0c\u4ecd\u7136\u80fd\u591f\u5728\u65b0\u73af\u5883\u4e2d\u6210\u529f\u5b8c\u6210\u4efb\u52a1\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5176\u4ed6\u65b9\u6cd5\u968f\u7740\u6d4b\u8bd5\u73af\u5883\u7684\u53d8\u5316\u8868\u73b0\u5927\u5e45\u4e0b\u964d\u3002<\/li>\n<\/ol>\n\n\n\n<h5 class=\"wp-block-heading\">\u4e94\u3001\u7ed3\u8bba<\/h5>\n\n\n\n<p>\u672c\u7814\u7a76\u901a\u8fc7\u5bf9\u73b0\u6709\u5f3a\u5316\u5b66\u4e60\u548c\u6700\u4f18\u63a7\u5236\u65b9\u6cd5\u7684\u5168\u9762\u6bd4\u8f83\uff0c\u5f97\u51fa\u4e86\u4e00\u4e9b\u5173\u952e\u7ed3\u8bba\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\u7684\u5b66\u4e60\u548c\u89c4\u5212\u65b9\u6cd5\uff08PLDM\uff09\u8868\u73b0\u4f18\u8d8a<\/strong>\uff0c\u80fd\u591f\u5728\u8f83\u5dee\u7684\u6570\u636e\u8d28\u91cf\u4e0b\u5c55\u73b0\u51fa\u6781\u5f3a\u7684\u9c81\u68d2\u6027\u548c\u6570\u636e\u6548\u7387\uff0c\u540c\u65f6\u5728\u65b0\u73af\u5883\u548c\u65b0\u4efb\u52a1\u4e2d\u7684\u6cdb\u5316\u80fd\u529b\u4e5f\u6700\u4e3a\u7a81\u51fa\u3002<\/li>\n\n\n\n<li><strong>\u76ee\u6807\u6761\u4ef6\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08\u5982HIQL\uff09\u548c\u5c42\u6b21\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\uff08\u5982HILP\uff09<\/strong>\u5728\u6570\u636e\u5145\u8db3\u65f6\u8868\u73b0\u6700\u4f73\uff0c\u4f46\u5728\u6570\u636e\u7a00\u7f3a\u548c\u73af\u5883\u53d8\u5316\u8f83\u5927\u65f6\uff0c\u5176\u8868\u73b0\u8f83\u5dee\u3002<\/li>\n\n\n\n<li><strong>\u57fa\u4e8e\u6a21\u578b\u7684\u63a7\u5236\u65b9\u6cd5\uff08\u5982PLDM\uff09\u5728\u6570\u636e\u7a00\u7f3a\u65f6\u5177\u6709\u66f4\u9ad8\u7684\u6837\u672c\u6548\u7387<\/strong>\uff0c\u80fd\u591f\u4ece\u5c11\u91cf\u7684\u6570\u636e\u4e2d\u63d0\u53d6\u66f4\u591a\u6709\u7528\u7684\u4fe1\u606f\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u7814\u7a76\u8fd8\u8868\u660e\uff0c\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\u548c\u57fa\u4e8e\u89c4\u5212\u7684\u5b66\u4e60\u65b9\u6cd5\uff0c\u5728\u5e94\u5bf9\u4e0d\u786e\u5b9a\u6027\u548c\u53d8\u5316\u8f83\u5927\u7684\u4efb\u52a1\u65f6\uff0c\u5177\u6709\u8f83\u5f3a\u7684\u9002\u5e94\u6027\u548c\u6f5c\u529b\u3002\u672a\u6765\u7684\u7814\u7a76\u53ef\u4ee5\u8fdb\u4e00\u6b65\u4f18\u5316PLDM\u7684\u8ba1\u7b97\u6548\u7387\uff0c\u5e76\u63a2\u7d22\u5982\u4f55\u5728\u66f4\u590d\u6742\u7684\u7269\u7406\u73af\u5883\u548c\u673a\u5668\u4eba\u63a7\u5236\u4efb\u52a1\u4e2d\u5e94\u7528\u8fd9\u4e00\u65b9\u6cd5\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>PLDM \u901a\u8fc7\u81ea\u76d1\u7763\u5b66\u4e60\uff08Self-Supervised Learning, SSL\uff09\u5b66\u4e60\u73af\u5883\u7684 <strong>\u6f5c\u5728\uff08\u7a7a\u95f4\uff09\u52a8\u529b\u5b66\u6a21\u578b\uff08latent dynamics model\uff09<\/strong>\uff0c\u5373\u5728\u4e00\u4e2a\u4f4e\u7ef4\u7684\u6f5c\u5728\u7a7a\u95f4\uff08latent space\uff09\u4e2d\u8fdb\u884c\u52a8\u6001\u5efa\u6a21\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u5728\u539f\u59cb\u89c2\u6d4b\u7a7a\u95f4\uff08\u5982\u50cf\u7d20\u7a7a\u95f4\u6216\u9ad8\u7ef4\u72b6\u6001\u7a7a\u95f4\uff09\u4e2d\u8fdb\u884c\u5efa\u6a21\u3002<\/p>\n\n\n\n<p>\u5177\u4f53\u800c\u8a00\uff0cPLDM \u7684\u6838\u5fc3\u601d\u60f3\u662f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4f7f\u7528\u7f16\u7801\u5668\uff08encoder\uff09\u5c06\u539f\u59cb\u72b6\u6001\u6620\u5c04\u5230\u6f5c\u5728\u7a7a\u95f4<\/strong>\uff1a\u72b6\u6001 s \u7ecf\u8fc7\u7f16\u7801\u5668 h<sub>\u03b8<\/sub>\u53d8\u6362\uff0c\u5f97\u5230\u5bf9\u5e94\u7684\u6f5c\u5728\u8868\u793a z\u3002<\/li>\n\n\n\n<li><strong>\u5728\u6f5c\u5728\u7a7a\u95f4\u4e2d\u8fdb\u884c\u52a8\u529b\u5b66\u5efa\u6a21<\/strong>\uff1a\u4f7f\u7528\u4e00\u4e2a\u9884\u6d4b\u6a21\u578b\uff08dynamics model\uff09 f<sub>\u03b8<\/sub>\u5728\u6f5c\u5728\u7a7a\u95f4\u5185\u5bf9\u72b6\u6001\u8f6c\u79fb\u8fdb\u884c\u5efa\u6a21\uff0c\u5373\u5b66\u4e60 z<sub>t+1<\/sub>=f<sub>\u03b8<\/sub>(z<sub>t<\/sub>,a<sub>t<\/sub>)\u3002<\/li>\n\n\n\n<li><strong>\u57fa\u4e8e\u6f5c\u5728\u7a7a\u95f4\u8fdb\u884c\u89c4\u5212<\/strong>\uff1a\u5728\u6d4b\u8bd5\u65f6\uff0c\u7ed9\u5b9a\u521d\u59cb\u72b6\u6001\u548c\u76ee\u6807\u72b6\u6001\uff0cPLDM \u5728\u6f5c\u5728\u7a7a\u95f4\u5185\u641c\u7d22\u6700\u4f18\u7684\u52a8\u4f5c\u5e8f\u5217\uff0c\u4f7f\u5f97\u667a\u80fd\u4f53\u80fd\u591f\u8fbe\u5230\u76ee\u6807\u72b6\u6001\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u76f8\u6bd4\u4e8e\u5728\u9ad8\u7ef4\u72b6\u6001\u7a7a\u95f4\u4e2d\u76f4\u63a5\u5efa\u6a21\uff0cPLDM \u901a\u8fc7 <strong>\u6f5c\u5728\u7a7a\u95f4\u5b66\u4e60\uff08latent space learning\uff09<\/strong> \u53ef\u4ee5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u907f\u514d\u5197\u4f59\u4fe1\u606f<\/strong>\uff1a\u4f8b\u5982\u5bf9\u4e8e\u89c6\u89c9\u4efb\u52a1\uff0c\u539f\u59cb\u50cf\u7d20\u7a7a\u95f4\u5305\u542b\u5927\u91cf\u65e0\u5173\u4fe1\u606f\uff0c\u800c\u6f5c\u5728\u7a7a\u95f4\u53ef\u4ee5\u4fdd\u7559\u4efb\u52a1\u76f8\u5173\u7279\u5f81\u3002<\/li>\n\n\n\n<li><strong>\u63d0\u9ad8\u6570\u636e\u6548\u7387<\/strong>\uff1a\u5728\u4f4e\u7ef4\u6f5c\u5728\u7a7a\u95f4\u4e2d\u8fdb\u884c\u5b66\u4e60\u548c\u89c4\u5212\uff0c\u80fd\u591f\u51cf\u5c11\u6570\u636e\u9700\u6c42\uff0c\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u589e\u5f3a\u6cdb\u5316\u80fd\u529b<\/strong>\uff1aPLDM \u5728\u5b9e\u9a8c\u4e2d\u5c55\u73b0\u51fa\u6bd4\u4f20\u7edf\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u66f4\u597d\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u5c24\u5176\u662f\u5728\u65b0\u73af\u5883\u5e03\u5c40\u6216\u6b21\u4f18\u6570\u636e\u4e0a\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>Latent Planning: <a href=\"https:\/\/latent-planning.github.io\/\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/latent-planning.github.io\/<\/a><br>PLDM on GitHub: <a href=\"https:\/\/github.com\/vladisai\/PLDM\" target=\"_blank\" rel=\"noreferrer noopener\">https:\/\/github.com\/vladisai\/PLDM<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Learning from Reward-Free Offline Data: A Case for Pl [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,63,77],"class_list":{"0":"post-5721","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"hentry","6":"category-tech","7":"tag-ai","9":"tag-meta"},"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Learning from Reward-Free Offline Data: A Case for Pl&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5721","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5721"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5721\/revisions"}],"predecessor-version":[{"id":5787,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5721\/revisions\/5787"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5721"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5721"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5721"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}