{"id":10038,"date":"2026-07-08T19:54:47","date_gmt":"2026-07-08T11:54:47","guid":{"rendered":"https:\/\/infernews.com\/blog\/code-for-quot-3d-hamster-enabling-robust-manipulation-in-hierarchical-vision-lan\/"},"modified":"2026-07-08T19:54:47","modified_gmt":"2026-07-08T11:54:47","slug":"code-for-quot-3d-hamster-enabling-robust-manipulation-in-hierarchical-vision-lan","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/code-for-quot-3d-hamster-enabling-robust-manipulation-in-hierarchical-vision-lan\/","title":{"rendered":"3D HAMSTER \u628a\u6a5f\u68b0\u81c2\u898f\u5283\u5e36\u9032\u771f 3D"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/method_figure.jpg\" alt=\"3D HAMSTER architecture: a depth-augmented VLM planner produces metric 3D waypoints that unproject into the point cloud\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a5f\u68b0\u81c2\u6700\u6015\u898f\u5283\u8b1b\u5f97\u901a\uff0c\u843d\u5230\u63a7\u5236\u5c64\u537b\u5c0d\u5514\u6e96\u7a7a\u9593\u4f4d\u7f6e\u30023D HAMSTER\u5c6c\u65bc <strong>Vision-Language-Action \u898f\u5283\u6a21\u578b\uff0f\u6a5f\u68b0\u4eba\u7814\u7a76\u9805\u76ee<\/strong>\uff0c\u91cd\u9ede\u4e0d\u662f\u518d\u7522\u751f\u87a2\u5e55\u4e0a\u7684 2D waypoint\uff0c\u800c\u662f\u6839\u64da\u55ae\u5f35 RGB \u5716\u3001metric depth map \u540c\u6587\u5b57\u6307\u4ee4\uff0c\u76f4\u63a5\u8f38\u51fa\u5e36\u6709\u6df1\u5ea6\u7684 3D end-effector trajectory\uff0c\u540c\u6642\u9644\u4e0a gripper actions\uff0c\u683c\u5f0f\u4ea6\u6574\u7406\u6210 JSON\uff0c\u65b9\u4fbf\u63a5\u53bb\u5f8c\u7e8c\u63a7\u5236\u6d41\u7a0b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b83\u91dd\u5c0d\u7684\u77db\u76fe\u5f88\u660e\u78ba\uff1a\u4e0d\u5c11 hierarchical Vision-Language-Action models \u7531\u9ad8\u5c64 planner \u5148\u9810\u6e2c 2D \u50cf\u7d20\u8ecc\u8de1\uff0c\u518d\u4ea4\u7d66\u5403 point cloud \u7684\u4f4e\u5c64 policy \u57f7\u884c\u3002\u554f\u984c\u5728\u65bc 2D waypoint \u53ea\u6703\u300c\u501f\u7528\u300d\u50cf\u7d20\u4e0b\u65b9\u8868\u9762\u7684\u6df1\u5ea6\uff0c\u5e7e\u4f55\u4e0a\u672a\u5fc5\u771f\u662f\u6a5f\u68b0\u81c2\u61c9\u8d70\u7684\u4f4d\u7f6e\u30023D HAMSTER\u6539\u6210\u5728 metric 3D \u7a7a\u9593\u898f\u5283\uff0c\u7b49\u9ad8\u5c64\u8f38\u51fa\u53ef\u4ee5\u76f4\u63a5\u5c0d\u63a5 point-cloud low-level policy\uff0c\u5c11\u4e86\u4e2d\u9593 2D \u8f49 3D \u7684\u5931\u771f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u6a21\u578b\u9aa8\u5e79\u7528\u7684\u662f <strong>Qwen3-VL-8B<\/strong>\uff0c\u518d\u52a0\u4e0a\u51cd\u7d50\u7684 <strong>LingBot-Depth<\/strong> geometry encoder\uff0c\u7576\u4e2d\u63a1\u7528 <strong>DINOv2 ViT-L\/14<\/strong>\uff0c\u4e26\u914d\u5408 dense depth-reconstruction objective \u53bb\u5f37\u5316\u6df1\u5ea6\u611f\u77e5\u3002\u5132\u5b58\u5eab\u76ee\u524d\u63d0\u4f9b\u7684\u662f <strong>inference-only<\/strong> \u7248\u672c\uff0c\u5373\u662f\u4f60\u53ef\u4ee5\u5b89\u88dd\u5957\u4ef6\u3001\u4e0b\u8f09 checkpoint \u5f8c\u76f4\u63a5\u8dd1\u63a8\u7406\u6216\u7528 Gradio demo \u6e2c\u8a66\uff1b\u4f4e\u5c64 point-cloud policy \u5c6c\u65bc\u8ad6\u6587\u5b8c\u6574\u7cfb\u7d71\u4e00\u90e8\u5206\uff0c\u9019\u500b GitHub \u9805\u76ee\u672a\u6709\u4e00\u4f75\u653e\u51fa\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u76f4\u63a5\u8f38\u51fa <strong>[u, v, depth]<\/strong> waypoint\uff0c\u6df1\u5ea6\u55ae\u4f4d\u662f\u7c73\uff0c\u8f03\u6613\u63a5\u5165\u771f\u5be6\u6a5f\u68b0\u81c2\u6d41\u7a0b<\/li>\n<li>\u8d70 <strong>RGB-D + language instruction<\/strong> \u8def\u7dda\uff0c\u4e0d\u9760\u591a\u8996\u89d2\u89c0\u5bdf<\/li>\n<li>checkpoint \u5df2\u5305\u542b geometry encoder \u6b0a\u91cd\uff0c\u8f09\u5165\u6642\u6bcb\u9808\u518d\u984d\u5916\u6293\u6a21\u578b<\/li>\n<li>\u9805\u76ee\u504f\u5411\u5c55\u793a planner \u80fd\u529b\uff0c\u4e0d\u7b49\u65bc\u5b8c\u6574\u53ef\u90e8\u7f72\u7684\u6a5f\u68b0\u4eba\u63a7\u5236\u5806\u758a<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\uff0c\u5b83\u7684\u5b9a\u4f4d\u5f88\u9069\u5408\u505a\u7814\u7a76\u9a57\u8b49\u3001\u6a5f\u68b0\u64cd\u4f5c\u898f\u5283\u6bd4\u8f03\uff0c\u6216\u8005\u62ff\u4f86\u89c0\u5bdf 2D planner \u8207 3D planner \u5728\u5e7e\u4f55\u4e00\u81f4\u6027\u4e0a\u7684\u5dee\u7570\u3002\u4f5c\u8005\u5728\u9805\u76ee\u9801\u9762\u5f37\u8abf\u9019\u7a2e\u505a\u6cd5\u5c0d\u8a9e\u8a00\u3001\u7a7a\u9593\u8207\u8996\u89ba\u5206\u4f48\u8f49\u79fb\u66f4\u7a69\u5065\uff0c\u4f46\u9019\u500b\u5132\u5b58\u5eab\u672a\u6709\u5b8c\u6574\u516c\u958b\u8a13\u7df4\u6d41\u7a0b\u8207\u4f4e\u5c64\u63a7\u5236\u6a21\u7d44\uff0c\u6240\u4ee5\u8f03\u9069\u5408\u628a\u5b83\u7406\u89e3\u70ba\u4e00\u500b\u6e05\u695a\u5c55\u793a\u65b9\u5411\u7684 3D-native planner\uff0c\u800c\u4e0d\u662f\u5373\u88dd\u5373\u7528\u7684\u5168\u5957\u6a5f\u68b0\u4eba\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/davian-robotics.github.io\/3D_HAMSTER\/\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/DAVIAN-Robotics\/3D_HAMSTER\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2606.31329\" rel=\"noopener noreferrer\" target=\"_blank\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u540c\u6a23\u662f\u770b\u5716\u52a0\u6587\u5b57\u6307\u4ee4\uff0c3D HAMSTER\u5514\u518d\u53ea\u756b2D\u8def\u7dda\uff0c\u800c\u662f\u76f4\u63a5\u4ea4\u51fa\u53ef\u843d\u5730\u76843D\u6a5f\u68b0\u81c2\u8ecc\u8de1\u3002\u5c0d\u6a5f\u68b0\u64cd\u4f5c\u6d41\u7a0b\u4f86\u8b1b\uff0c\u5462\u500b\u6539\u52d5\u7d30\u7bc0\u4e0d\u591a\uff0c\u5f71\u97ff\u537b\u5f88\u5be6\u969b\u3002<\/p>\n","protected":false},"author":8,"featured_media":10037,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,140,132,119,149,184,199,204],"tags":[],"class_list":["post-10038","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-gemini","category-3d","category-119","category-149","category-robotic","category-dataset-","category-visionlanguageaction"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10038","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=10038"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10038\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10037"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10038"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10038"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10038"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}