{"id":10584,"date":"2026-08-01T00:18:42","date_gmt":"2026-07-31T16:18:42","guid":{"rendered":"https:\/\/infernews.com\/blog\/temporal-distance-jepa\/"},"modified":"2026-08-01T00:18:42","modified_gmt":"2026-07-31T16:18:42","slug":"temporal-distance-jepa","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/temporal-distance-jepa\/","title":{"rendered":"Temporal-Distance-JEPA \u8a13\u7df4\u4e16\u754c\u6a21\u578b\u80fd\u61c2\u300c\u6642\u9593\u8ddd\u96e2\u300d"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/08\/pasted-9c92a6e0194d.jpg\" alt=\"Repository image for HKBU-KnowComp\/Temporal-Distance-JEPA\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9999\u6e2f\u6d78\u6703\u5927\u5b78 HKBU KnowComp \u7684 Jiaxin Bai \u516c\u958b\u4e86 Temporal-Distance-JEPA \u8ad6\u6587\u91cd\u73fe\u9805\u76ee\uff0c\u6838\u5fc3\u4e0d\u662f\u518d\u505a\u4e00\u500b\u66f4\u8907\u96dc\u7684\u4e16\u754c\u6a21\u578b\uff0c\u800c\u662f\u4fee\u6b63 JEPA \u898f\u5283\u5e38\u898b\u7684\u843d\u5dee\uff1a\u8a13\u7df4\u6642\u5b78\u77ed\u671f latent prediction\uff0c\u898f\u5283\u6642\u537b\u5f80\u5f80\u76f4\u63a5\u62ff latent Euclidean distance \u7576\u6210\u76ee\u6a19\u9032\u5ea6\u3002\u9019\u500b\u9805\u76ee\u5c6c\u65bc<strong>\u6a21\u578b\u7814\u7a76\u91cd\u73fe\u9805\u76ee<\/strong>\uff0c\u8655\u7406\u7684\u662f\u96e2\u7dda\u793a\u7bc4\u6c92\u6709 reward \u6642\uff0cplanner \u9ede\u6a23\u5224\u65b7\u54ea\u689d\u60f3\u50cf\u8def\u5f91\u66f4\u63a5\u8fd1\u76ee\u6a19\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4f5c\u8005\u4fdd\u7559 LeWM encoder\u2013predictor \u8207 SIGReg backbone\uff0c\u4f46\u4e0d\u518d\u53ea\u4f9d\u8cf4 embedding \u5e7e\u4f55\uff0c\u800c\u662f\u5f9e reward-free demonstration logs \u6316\u51fa directed temporal cost\u3002\u505a\u6cd5\u6cbf\u4f4f\u8ad6\u6587\u908f\u8f2f\u5f88\u6e05\u695a\uff1a\u540c\u4e00\u689d trajectory \u7684\u5148\u5f8c\u6b65\u5e8f\u63d0\u4f9b positive targets\uff0c\u8de8 trajectory \u914d\u5c0d\u5145\u7576 heuristic negatives\uff0c\u518d\u7528 rollout-consistency \u4ee4\u5b78\u5230\u7684\u6210\u672c\u66f4\u8cbc\u8fd1 planner horizon\u3002\u9019\u7a2e framing \u76f4\u63a5\u56de\u61c9\u820a\u7bc4\u5f0f\u628a latent L2 \u7576\u6210\u9032\u5ea6\u4ee3\u7406\u7684\u9650\u5236\uff0c\u56e0\u70ba\u62d3\u64b2\u4e3b\u5c0e\u4efb\u52d9\u672a\u5fc5\u9069\u5408\u53ea\u770b\u5e7e\u4f55\u8ddd\u96e2\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7d50\u679c\u5728 Two-Room\u3001Reacher \u9019\u985e topology-dominated tasks\uff0c\u898f\u5283\u6642\u76f4\u63a5\u90e8\u7f72 mined cost d_psi\uff1b\u5230\u4e86 Push-T\u3001OGB-Cube \u9019\u985e contact-rich tasks\uff0c\u5247\u6539\u7528\u540c\u4e00\u500b temporally trained checkpoint \u4e0a\u7684 latent l2 planning\u3002locked evaluation \u4e0b\uff0cTwo-Room \u6210\u529f\u7387\u63d0\u5347\u5230 100.0%\uff0c\u9ad8\u65bc LeWM \u7684 97.4%\uff1bOGB-Cube \u4ea6\u6bd4 LeWM \u9ad8 14.2 \u5206\uff0c\u4e26\u4e14\u6574\u9ad4\u4e0a\u5c0d LeWM \u8207 RC-aux baseline \u90fd\u80fd\u6301\u5e73\u6216\u66f4\u597d\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5efa\u57fa\u65bc stable-worldmodel\u3001stable-pretraining \u8207 LeWM \u65e2\u6709\u5e03\u5c40\uff0c\u91cd\u9ede\u5728\u8a13\u7df4\u76ee\u6a19\u8207\u898f\u5283\u6210\u672c\u7684\u91cd\u65b0\u5c0d\u9f4a<\/li>\n<li>\u516c\u958b\u5eab\u662f paper-reproduction release\uff0c\u4fdd\u7559 Python CLI\uff0c\u4f46\u523b\u610f\u4e0d\u9644 cluster\uff0fSlurm \u5305\u88dd<\/li>\n<li>\u8cc7\u6599\u4ee5 HDF5 \u5f62\u5f0f\u653e\u5728 STABLEWM_HOME\uff0c\u6db5\u84cb Push-T\u3001Two-Room\u3001Reacher\u3001OGB-Cube<\/li>\n<li>\u63a1\u7528 10 epochs\uff0c\u4e26\u9644\u6709 locked results\u3001cost matrix\u3001diagnostics \u7b49\u7d50\u679c\u8cc7\u6599<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u5eab\u6bd4\u8f03\u50cf\u7d66\u5df2\u6709\u7814\u7a76\u5de5\u4f5c\u6d41\u7684\u4eba\u63a5\u624b\u91cd\u8dd1\uff1a\u6709 Python \u74b0\u5883\u3001requirements\u3001Hydra config\u3001\u8cc7\u6599\u4f4d\u7f6e\u8207 checkpoint \u8f38\u51fa\u65b9\u5f0f\uff0c\u4f46\u6c92\u6709\u66ff\u4e0d\u540c\u6392\u7a0b\u7cfb\u7d71\u6e96\u5099\u73fe\u6210\u5c01\u88dd\u3002\u6b77\u53f2\u4e0a\u7684 contrastive SoftJEPA \u76f8\u5bb9\u7a0b\u5f0f\u4ecd\u53ef\u80fd\u7559\u5728 losses.py \u6216 jepa.py\uff0c\u4e0d\u5c6c\u65bc\u516c\u958b\u914d\u7f6e\uff1b\u63db\u53e5\u8a71\u8aaa\uff0c\u95b1\u8b80\u8207\u91cd\u73fe\u6642\u8981\u4ee5 td_jepa \u8a2d\u5b9a\u70ba\u6e96\u3002\u5c0d\u505a world model\u3001offline RL\u3001latent MPC \u7684\u7814\u7a76\u5718\u968a\u4f86\u8aaa\uff0c\u9019\u500b\u9805\u76ee\u6700\u6709\u50f9\u503c\u7684\u5730\u65b9\uff0c\u662f\u5b83\u628a\u300c\u8868\u793a\u5b78\u5230\u4ec0\u9ebc\u300d\u8207\u300c\u898f\u5283\u8981\u600e\u6a23\u6392\u5e8f\u672a\u4f86\u300d\u91cd\u65b0\u7d81\u5728\u4e00\u8d77\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/HKBU-KnowComp\/Temporal-Distance-JEPA\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.25337\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9999\u6e2f\u6d78\u6703\u5927\u5b78\u5718\u968a\u628a JEPA \u4e16\u754c\u6a21\u578b\u7531\u300c\u6703\u9810\u6e2c\u300d\u63a8\u8fd1\u4e00\u6b65\u5230\u300c\u66f4\u8b58\u898f\u5283\u300d\u3002Temporal-Distance-JEPA \u91dd\u5c0d\u96e2\u7dda\u793a\u7bc4\u7f3a\u5c11\u56de\u5831\u8a0a\u865f\u7684\u75db\u9ede\uff0c\u76f4\u63a5\u5f9e\u8ecc\u8de1\u6b21\u5e8f\u6316\u51fa\u53ef\u7528\u7684\u9032\u5ea6\u6210\u672c\u3002<\/p>\n","protected":false},"author":8,"featured_media":10583,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,163,131,152,127,186,189,208],"tags":[],"class_list":["post-10584","post","type-post","status-publish","format-standard","hentry","category-133","category-163","category-embedding","category-python","category-127","category-186","category-189","category-hkbu"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10584","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10584"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10584\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10583"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10584"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10584"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10584"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}