
{"id":12620,"date":"2026-09-26T00:42:23","date_gmt":"2026-09-25T16:42:23","guid":{"rendered":"https:\/\/infernews.com\/blog\/training-object-permanence-in-world-models-the-codebase\/"},"modified":"2026-09-26T00:42:23","modified_gmt":"2026-09-25T16:42:23","slug":"training-object-permanence-in-world-models-the-codebase","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/training-object-permanence-in-world-models-the-codebase\/","title":{"rendered":"object-permanence\uff1a\u8b93\u4e16\u754c\u6a21\u578b\u5b78\u8b58\u8a18\u4f4f\u88ab\u906e\u64cb\u5605\u7269\u4ef6"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-578fb34307ed.jpg\" alt=\"Training Object Permanence in World Models\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7576\u4e00\u4ef6\u7269\u4ef6\u88ab\u5c4f\u5e55\u6216\u7246\u58c1\u64cb\u4f4f\u518d\u51fa\u73fe\u6642\uff0c\u73fe\u6642\u5605\u4e16\u754c\u6a21\u578b\u5f80\u5f80\u6703\u300c\u5fd8\u8a18\u300d\u4f62\u539f\u672c\u5605\u6a23\u8c8c\u540c\u4f4d\u7f6e\u3002PWM-WROP \u6b63\u6b63\u91dd\u5c0d\u5462\u500b\u9650\u5236\uff1a\u4f62\u4fc2\u4e00\u500b\u57fa\u65bc NVIDIA Cosmos3-Nano \u5fae\u8abf\u5605 16B \u591a\u6a21\u614b\u4e16\u754c\u6a21\u578b\uff08Multimodal World Model\uff09\uff0c\u5c08\u9580\u8a13\u7df4\u7269\u4ef6\u6c38\u7e8c\u6027\uff08object permanence\uff09\u2014\u2014\u5373\u4fc2\u7269\u4ef6\u5373\u4f7f\u88ab\u906e\u64cb\u90fd\u8981\u55ba\u5167\u90e8\u8868\u793a\u4e2d\u4fdd\u7559\u843d\u53bb\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8a13\u7df4\u6578\u64da\u569f\u81ea\u4e00\u500b Blender \u6e32\u67d3\u5605\u6578\u64da\u5de5\u5ee0\uff0c\u5305\u542b 150 \u500b\u4efb\u52d9\u751f\u6210\u5668\uff0c\u6db5\u84cb\u516d\u5927\u8a8d\u77e5\u5bb6\u65cf\uff0c\u4f8b\u5982 turntable_behind_screen_task \u7b49\u3002\u6bcf\u500b\u6a23\u672c\u6703\u8f38\u51fa\u4e94\u4ef6\u6a94\u6848\uff1ainput_video\u3001target_video\u3001prompt\u3001trajectory.npz \u540c metadata.json\uff0c\u65b9\u4fbf\u7528\u569f\u505a video-to-video \u76e3\u7763\u5b78\u7fd2\u3002\u6574\u500b\u8a13\u7df4\u7528\u539f\u751f PyTorch\uff08\u907f\u514d XLA graph tracing\uff09\u55ba AWS Trainium2 \u4e0a\u8dd1\uff0c\u900f\u904e tensor parallel \u00d7 FSDP2 \u8de8 64 \u500b NeuronCore\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u65bc\u5f9e\u4e8b\u7269\u7406\u63a8\u7406\u3001\u6a5f\u5668\u4eba\u6a21\u64ec\u6216\u5f71\u7247\u751f\u6210\u5605\u5718\u968a\u5462\u500b\u9805\u76ee\u7279\u5225\u6709\u7528\u2014\u2014\u4f62\u76f4\u63a5\u63d0\u4f9b\u53ef\u751f\u6210\u5605\u8a13\u7df4\u6578\u64da\u3001\u9810\u8a13\u7df4\u6b0a\u91cd\uff08BF16\uff0c\u7d04 15B \u53c3\u6578\uff09\u540c benchmark\uff08WROP 300 \u984c\uff09\uff0c\u6e1b\u5c11\u7531\u96f6\u5efa\u7acb\u7269\u4ef6\u6c38\u7e8c\u6027\u6e2c\u8a66\u5605\u6210\u672c\u3002<strong>Hugging Face<\/strong> \u4e0a\u9762\u5df2\u6709\u6a21\u578b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u91cd\u9ede\u6458\u8981\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6578\u64da\u898f\u6a21<\/strong>\uff1a150 \u500b Blender \u4efb\u52d9\u751f\u6210\u5668\uff0c\u6bcf\u500b 20 \u689d\u6a23\u672c\uff0c\u5408\u5171\u7d04 150 \u842c\u689d video-to-video \u8a13\u7df4\u6578\u64da<\/li>\n<li><strong>\u6a21\u578b\u898f\u6a21<\/strong>\uff1aPWM-WROP \u70ba 16B \u53c3\u6578\uff08FP16 \u7d04 15B\uff09\uff0c\u57fa\u65bc Cosmos3-Nano \u5fae\u8abf<\/li>\n<li><strong>\u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u8a13\u7df4\u9700\u8981 AWS Trainium2\uff08trn2.48xlarge\uff0c64 NeuronCore\uff09\uff0c\u63a8\u8ad6\u5247\u8996\u4e4e\u90e8\u7f72\u65b9\u5f0f<\/li>\n<li><strong>\u6388\u6b0a\u9650\u5236<\/strong>\uff1a\u63a1\u7528 CC BY-NC 4.0\uff0c\u50c5\u9650\u975e\u5546\u696d\u4f7f\u7528\uff0c\u5546\u7528\u9700\u53e6\u884c\u806f\u7d61\u4f5c\u8005<\/li>\n<li><strong>\u751f\u614b\u6574\u5408<\/strong>\uff1a\u914d\u642d WROP 300 \u984c benchmark\uff0c\u53ef\u4ee5\u91cf\u5316\u5514\u540c\u5f71\u7247\u6a21\u578b\u5605\u7269\u4ef6\u6c38\u7e8c\u6027\u8868\u73fe<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u540c\u4e00\u822c\u53ea\u505a\u5916\u89c0\u9810\u6e2c\u5605\u4e16\u754c\u6a21\u578b\u76f8\u6bd4\uff0cPWM-WROP \u5f37\u8abf\u300c\u88ab\u906e\u64cb\u90fd\u8981\u8a18\u5f97\u300d\uff0c\u4e26\u900f\u904e\u8ecc\u8de1\u6578\u64da\uff08trajectory.npz\uff09\u63d0\u4f9b\u9010\u5e40 ground truth \u505a\u76e3\u7763\uff0c\u5c0d\u65bc\u9700\u8981\u7269\u4ef6\u7d1a\u5225\u4e00\u81f4\u6027\u5605\u61c9\u7528\u5834\u666f\uff08\u4f8b\u5982\u6a5f\u68b0\u81c2\u64cd\u4f5c\u9810\u6f14\u3001\u8996\u983b\u9810\u6e2c\u54c1\u8cea\u63d0\u5347\uff09\u6703\u6709\u660e\u986f\u5e6b\u52a9\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/www.object-permanence.world\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/hokindeng\/object-permanence\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/huggingface.co\/Hokin\/PWM-WROP\" rel=\"noopener noreferrer\"><strong>\u6a21\u578b<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u900f\u904e 150 \u500b Blender \u4efb\u52d9\u540c 150 \u842c\u689d\u8a13\u7df4\u6a23\u672c\uff0c\u8a13\u7df4\u4e00\u500b 16B \u4e16\u754c\u6a21\u578b\u55ba\u7269\u4ef6\u88ab\u906e\u64cb\u6642\u4ecd\u80fd\u4fdd\u6301\u8a18\u61b6\uff0c\u8655\u7406\u8996\u983b\u751f\u6210\u4e2d\u5e38\u898b\u5605\u300c\u7269\u4ef6\u6d88\u5931\u300d\u554f\u984c\u3002<\/p>\n","protected":false},"author":8,"featured_media":12619,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,162,76,119,141,186,179,31,152,199],"tags":[],"class_list":["post-12620","post","type-post","status-publish","format-standard","hentry","category-133","category-ai-productions","category-76","category-119","category-141","category-186","category-nvidia","category-video","category-python","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12620","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=12620"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/12620\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/12619"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=12620"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=12620"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=12620"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}