{"id":92723,"date":"2026-08-11T00:16:44","date_gmt":"2026-08-10T16:16:44","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/92723.html"},"modified":"2026-08-11T00:16:44","modified_gmt":"2026-08-10T16:16:44","slug":"2024-2025%e5%b9%b4%e8%a7%86%e8%a7%89-%e8%af%ad%e8%a8%80-%e5%8a%a8%e4%bd%9c%ef%bc%88vla%ef%bc%89%e6%a8%a1%e5%9e%8b%e5%89%8d%e6%b2%bf%e7%a0%94%e7%a9%b6%e6%8a%a5%e5%91%8a","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/92723.html","title":{"rendered":"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a"},"content":{"rendered":"<h2>1. \u5f15\u8a00&#xff1a;\u5177\u8eab\u667a\u80fd\u7684\u201c\u57fa\u7840\u6a21\u578b\u201d\u65f6\u4ee3<\/h2>\n<h4>1.1 \u4ece\u5355\u4e00\u4efb\u52a1\u5230\u901a\u7528\u7b56\u7565\u7684\u8303\u5f0f\u8f6c\u79fb<\/h4>\n<p>\u57282024\u5e74\u81f32025\u5e74\u521d\u7684\u8fd9\u6bb5\u65f6\u95f4\u91cc&#xff0c;\u673a\u5668\u4eba\u5b66\u4e60&#xff08;Robot Learning&#xff09;\u9886\u57df\u7ecf\u5386\u4e86\u4e00\u573a\u6df1\u523b\u7684\u8303\u5f0f\u8f6c\u79fb&#xff0c;\u5176\u6838\u5fc3\u5728\u4e8e\u4ece\u9488\u5bf9\u7279\u5b9a\u73af\u5883\u3001\u7279\u5b9a\u4efb\u52a1\u7684\u201c\u4e13\u5bb6\u6a21\u578b\u201d\u5411\u5177\u5907\u5e7f\u6cdb\u6cdb\u5316\u80fd\u529b\u7684\u201c\u901a\u624d\u6a21\u578b\u201d&#xff08;Generalist Policies&#xff09;\u6f14\u8fdb\u3002\u8fd9\u4e00\u8f6c\u53d8\u7684\u9a71\u52a8\u529b\u4e3b\u8981\u6765\u81ea\u4e8e\u5927\u8bed\u8a00\u6a21\u578b&#xff08;LLM&#xff09;\u548c\u89c6\u89c9\u8bed\u8a00\u6a21\u578b&#xff08;VLM&#xff09;\u5728\u4e92\u8054\u7f51\u89c4\u6a21\u6570\u636e\u4e0a\u7684\u5de8\u5927\u6210\u529f\u3002\u7814\u7a76\u8005\u4eec\u5f00\u59cb\u601d\u8003&#xff1a;\u5982\u679c\u8bed\u8a00\u548c\u89c6\u89c9\u53ef\u4ee5\u901a\u8fc7\u9884\u6d4b\u4e0b\u4e00\u4e2atoken\u6765\u7406\u89e3\u4e16\u754c&#xff0c;\u90a3\u4e48\u7269\u7406\u4e16\u754c\u7684\u63a7\u5236\u4fe1\u53f7&#xff08;\u52a8\u4f5c&#xff09;\u662f\u5426\u4e5f\u53ef\u4ee5\u88ab\u89c6\u4e3a\u4e00\u79cd\u201c\u8bed\u8a00\u201d&#xff0c;\u4ece\u800c\u901a\u8fc7\u7c7b\u4f3c\u7684\u9884\u8bad\u7ec3\u8303\u5f0f\u6765\u5b9e\u73b0\u901a\u7528\u7684\u7269\u7406\u667a\u80fd&#xff1f;<\/p>\n<p>\u8fd9\u5c31\u50ac\u751f\u4e86\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c&#xff08;Vision-Language-Action, VLA&#xff09;\u6a21\u578b\u7684\u6982\u5ff5\u3002\u4e0e\u4f20\u7edf\u7684\u6a21\u4eff\u5b66\u4e60&#xff08;Imitation Learning&#xff09;\u4e0d\u540c&#xff0c;VLA\u4e0d\u518d\u4ec5\u4ec5\u662f\u7b80\u5355\u7684\u611f\u77e5-\u52a8\u4f5c\u6620\u5c04\u7f51\u7edc&#xff0c;\u800c\u662f\u5efa\u7acb\u5728\u5e9e\u5927\u7684\u8bed\u4e49\u7406\u89e3\u57fa\u7840\u4e4b\u4e0a\u7684\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\u3002\u5b83\u4eec\u4e0d\u4ec5\u80fd\u201c\u770b\u201d\u548c\u201c\u8bfb\u201d&#xff0c;\u8fd8\u80fd\u76f4\u63a5\u8f93\u51fa\u63a7\u5236\u673a\u5668\u4eba\u7684\u5e95\u5c42\u6307\u4ee4\u3002<\/p>\n<h4>1.2 2024-2025\u5e74\u7684\u6280\u672f\u5206\u6c34\u5cad<\/h4>\n<p>2023\u5e74Google DeepMind\u53d1\u5e03\u7684RT-2\u6a21\u578b\u5960\u5b9a\u4e86VLA\u7684\u6982\u5ff5\u57fa\u7840&#xff0c;\u8bc1\u660e\u4e86\u5c06\u52a8\u4f5c\u4f5c\u4e3a\u8bed\u8a00token\u8fdb\u884c\u8054\u5408\u8bad\u7ec3\u662f\u53ef\u884c\u7684\u3002\u7136\u800c&#xff0c;\u8fdb\u51652024\u5e74\u548c2025\u5e74&#xff0c;\u968f\u7740OpenVLA\u3001Octo\u3001$\\\\pi_0$ (Pi-zero) \u4ee5\u53caNVIDIA\u7684GR00T\u7b49\u6a21\u578b\u7684\u6d8c\u73b0&#xff0c;\u8be5\u9886\u57df\u53d1\u751f\u4e86\u663e\u8457\u7684\u6280\u672f\u5206\u5316\u4e0e\u6df1\u5316\u3002<\/p>\n<p>\u5f53\u524d\u7684\u6587\u732e\u663e\u793a\u51fa\u4e24\u6761\u4e3b\u8981\u7684\u6280\u672f\u8def\u7ebf\u4e4b\u4e89&#xff1a;<\/p>\n<li>\n<p>\u79bb\u6563\u5316\u81ea\u56de\u5f52\u8def\u7ebf&#xff08;Autoregressive Discretization&#xff09;&#xff1a; \u4ee5OpenVLA\u4e3a\u4ee3\u8868\u3002\u8fd9\u6761\u8def\u7ebf\u575a\u6301\u201c\u52a8\u4f5c\u5373\u8bed\u8a00\u201d\u7684\u54f2\u5b66&#xff0c;\u901a\u8fc7\u5c06\u8fde\u7eed\u7684\u7269\u7406\u91cf&#xff08;\u5982\u5173\u8282\u89d2\u5ea6&#xff09;\u79bb\u6563\u5316\u4e3a\u8bcd\u8868\u4e2d\u7684token&#xff0c;\u5229\u7528\u6210\u719f\u7684Transformer\u67b6\u6784\u8fdb\u884c\u5e8f\u5217\u9884\u6d4b\u3002\u5176\u4f18\u52bf\u5728\u4e8e\u80fd\u65e0\u7f1d\u7ee7\u627fLLM\u7684\u63a8\u7406\u80fd\u529b&#xff0c;\u4f46\u52a3\u52bf\u5728\u4e8e\u52a8\u4f5c\u7684\u7cbe\u5ea6\u53d7\u9650\u4e8e\u79bb\u6563\u5316\u7684\u7c92\u5ea6&#xff0c;\u4e14\u63a8\u7406\u901f\u5ea6\u8f83\u6162\u3002<\/p>\n<\/li>\n<li>\n<p>\u8fde\u7eed\u751f\u6210\u5f0f\u8def\u7ebf&#xff08;Continuous Generative Modeling&#xff09;&#xff1a; \u4ee5Octo\u3001$\\\\pi_0$\u548cGR00T\u4e3a\u4ee3\u8868\u3002\u8fd9\u6761\u8def\u7ebf\u8ba4\u4e3a\u7269\u7406\u52a8\u4f5c\u672c\u8d28\u4e0a\u662f\u8fde\u7eed\u4e14\u591a\u6a21\u6001\u7684&#xff0c;\u79bb\u6563\u5316\u4f1a\u4e22\u5931\u9ad8\u9891\u63a7\u5236\u6240\u9700\u7684\u7cbe\u7ec6\u5ea6\u3002\u56e0\u6b64&#xff0c;\u5b83\u4eec\u91c7\u7528\u4e86**\u6269\u6563\u6a21\u578b&#xff08;Diffusion Models&#xff09;\u6216\u6d41\u5339\u914d&#xff08;Flow Matching&#xff09;**\u6280\u672f&#xff0c;\u5c06\u52a8\u4f5c\u751f\u6210\u5efa\u6a21\u4e3a\u53bb\u566a\u6216\u5411\u91cf\u573a\u56de\u5f52\u8fc7\u7a0b\u3002\u8fd9\u79cd\u65b9\u6cd5\u80fd\u591f\u5b9e\u73b0\u9ad8\u8fbe50Hz\u4ee5\u4e0a\u7684\u63a7\u5236\u9891\u7387&#xff0c;\u66f4\u9002\u5408\u9ad8\u52a8\u6001\u7684\u673a\u5668\u4eba\u4efb\u52a1\u3002<\/p>\n<\/li>\n<p>\u6b64\u5916&#xff0c;**\u53cc\u7cfb\u7edf\u67b6\u6784&#xff08;Dual-System Architecture&#xff09;**\u7684\u6982\u5ff5\u57282025\u5e74\u9010\u6e10\u6210\u4e3a\u4e3b\u6d41&#xff0c;\u7279\u522b\u662f\u5728\u4eba\u5f62\u673a\u5668\u4eba&#xff08;\u5982GR00T&#xff09;\u7684\u7814\u7a76\u4e2d\u3002\u8fd9\u79cd\u67b6\u6784\u6a21\u4eff\u4eba\u7c7b\u7684\u8ba4\u77e5\u8fc7\u7a0b&#xff0c;\u5c06\u201c\u6162\u601d\u8003\u201d&#xff08;System 2&#xff0c;\u8d1f\u8d23\u9ad8\u5c42\u8bed\u4e49\u89c4\u5212\u7684VLM&#xff09;\u4e0e\u201c\u5feb\u53cd\u5e94\u201d&#xff08;System 1&#xff0c;\u8d1f\u8d23\u5e95\u5c42\u9ad8\u9891\u63a7\u5236\u7684\u6269\u6563\/\u6d41\u5339\u914d\u7f51\u7edc&#xff09;\u89e3\u8026\u53c8\u534f\u540c\u5de5\u4f5c&#xff0c;\u89e3\u51b3\u4e86\u5927\u6a21\u578b\u63a8\u7406\u5ef6\u8fdf\u4e0e\u673a\u5668\u4eba\u5b9e\u65f6\u63a7\u5236\u9700\u6c42\u4e4b\u95f4\u7684\u77db\u76fe\u3002<\/p>\n<p>\u672c\u62a5\u544a\u5c06\u5bf9\u4e0a\u8ff0\u5173\u952e\u6a21\u578b\u8fdb\u884c\u8be6\u5c3d\u7684\u6280\u672f\u62c6\u89e3&#xff0c;\u5206\u6790\u5176\u80cc\u540e\u7684\u6570\u5b66\u539f\u7406\u3001\u6570\u636e\u7b56\u7565\u53ca\u8bad\u7ec3\u8303\u5f0f&#xff0c;\u5e76\u5728\u6700\u540e\u57fa\u4e8e\u8fd9\u4e9b\u9876\u7ea7\u6587\u732e\u7684\u5199\u4f5c\u98ce\u683c&#xff0c;\u603b\u7ed3\u51fa\u64b0\u5199VLA\u8bba\u6587\u201c\u65b9\u6cd5&#xff08;Method&#xff09;\u201d\u90e8\u5206\u7684\u7ed3\u6784\u5316\u6307\u5357\u3002<\/p>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>2. \u6838\u5fc3\u6280\u672f\u8def\u7ebf\u4e00&#xff1a;\u79bb\u6563\u5316\u81ea\u56de\u5f52 VLA \u2014\u2014 \u4ee5 OpenVLA \u4e3a\u4f8b<\/p>\n<h4>2.1 \u67b6\u6784\u8bbe\u8ba1\u54f2\u5b66&#xff1a;\u4e3a\u4e86\u5fae\u8c03\u800c\u751f<\/h4>\n<p>OpenVLA 1 \u662f2024\u5e74\u4e2d\u671f\u53d1\u5e03\u7684\u4e00\u4e2a\u91cc\u7a0b\u7891\u5f0f\u5de5\u4f5c&#xff0c;\u5176\u6838\u5fc3\u76ee\u6807\u662f\u89e3\u51b3\u4ee5\u5f80VLA\u6a21\u578b&#xff08;\u5982RT-2-X&#xff09;\u95ed\u6e90\u4e14\u96be\u4ee5\u9488\u5bf9\u65b0\u673a\u5668\u4eba\u8fdb\u884c\u5fae\u8c03\u7684\u95ee\u9898\u3002OpenVLA\u662f\u4e00\u4e2a70\u4ebf\u53c2\u6570&#xff08;7B&#xff09;\u7684\u5f00\u6e90VLA&#xff0c;\u5b83\u5e76\u6ca1\u6709\u91cd\u65b0\u53d1\u660e\u8f6e\u5b50&#xff0c;\u800c\u662f\u7cbe\u5999\u5730\u6574\u5408\u4e86\u73b0\u6709\u7684\u6700\u5f3a\u7ec4\u4ef6&#xff0c;\u6784\u5efa\u4e86\u4e00\u4e2a\u6781\u5176\u7a33\u5065\u7684\u57fa\u5ea7\u3002<\/p>\n<p>2.1.1 \u89c6\u89c9\u7f16\u7801\u5668\u7684\u878d\u5408\u7b56\u7565<\/p>\n<p>OpenVLA\u5728\u89c6\u89c9\u524d\u7aef\u7684\u8bbe\u8ba1\u4e0a\u4f53\u73b0\u4e86\u6df1\u523b\u7684\u6d1e\u5bdf\u3002\u4f20\u7edf\u7684VLM\u901a\u5e38\u53ea\u4f7f\u7528CLIP\u6216SigLIP\u8fd9\u6837\u7684\u5bf9\u6bd4\u5b66\u4e60\u7279\u5f81&#xff0c;\u8fd9\u4e9b\u7279\u5f81\u64c5\u957f\u8bed\u4e49\u5bf9\u9f50&#xff08;\u4f8b\u5982\u8bc6\u522b\u201c\u8fd9\u662f\u4e00\u4e2a\u82f9\u679c\u201d&#xff09;&#xff0c;\u4f46\u5728\u7a7a\u95f4\u51e0\u4f55\u611f\u77e5\u4e0a\u8f83\u5f31&#xff08;\u4f8b\u5982\u5b9a\u4f4d\u201c\u82f9\u679c\u7684\u8fb9\u7f18\u5728\u54ea\u91cc\u201d&#xff09;\u3002\u673a\u5668\u4eba\u64cd\u4f5c\u4e0d\u4ec5\u9700\u8981\u8bed\u4e49\u7406\u89e3&#xff0c;\u66f4\u9700\u8981\u7cbe\u786e\u7684\u51e0\u4f55\u611f\u77e5\u3002<\/p>\n<p>\u56e0\u6b64&#xff0c;OpenVLA\u91c7\u7528\u4e86**\u53cc\u89c6\u89c9\u7f16\u7801\u5668\u878d\u5408&#xff08;Fused Visual Encoder&#xff09;**\u7b56\u7565&#xff1a;<\/p>\n<ul>\n<li>\n<p>SigLIP (Sigmoid Loss for Language Image Pre-training): \u8d1f\u8d23\u6355\u6349\u56fe\u50cf\u7684\u9ad8\u5c42\u8bed\u4e49\u4fe1\u606f&#xff0c;\u786e\u4fdd\u6a21\u578b\u80fd\u7406\u89e3\u590d\u6742\u7684\u81ea\u7136\u8bed\u8a00\u6307\u4ee4\u4e0e\u89c6\u89c9\u573a\u666f\u7684\u5bf9\u5e94\u5173\u7cfb\u3002<\/p>\n<\/li>\n<li>\n<p>DINOv2 (Self-distillation with NO labels): \u8d1f\u8d23\u63d0\u4f9b\u5bc6\u96c6\u7684\u3001\u50cf\u7d20\u7ea7\u7684\u51e0\u4f55\u7279\u5f81\u3002DINOv2\u5728\u81ea\u76d1\u7763\u8bad\u7ec3\u4e2d\u4e60\u5f97\u7684\u7279\u5f81\u5bf9\u7269\u4f53\u7684\u7a7a\u95f4\u5e03\u5c40\u3001\u7eb9\u7406\u548c\u5f62\u72b6\u6781\u4e3a\u654f\u611f&#xff0c;\u8fd9\u5bf9\u4e8e\u7cbe\u51c6\u7684\u6293\u53d6\u548c\u64cd\u4f5c\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u4e24\u79cd\u7279\u5f81\u5728\u901a\u9053\u7ef4\u5ea6\u6216token\u7ef4\u5ea6\u8fdb\u884c\u878d\u5408\u540e&#xff0c;\u901a\u8fc7\u4e00\u4e2a\u591a\u5c42\u611f\u77e5\u673a&#xff08;MLP&#xff09;\u6295\u5f71\u5668\u6620\u5c04\u5230\u8bed\u8a00\u6a21\u578b\u7684\u5d4c\u5165\u7a7a\u95f4\u3002\u8fd9\u79cd\u8bbe\u8ba1\u76f4\u63a5\u89e3\u51b3\u4e86\u4ee5\u5f80\u6a21\u578b\u201c\u61c2\u4efb\u52a1\u4f46\u624b\u4e0d\u51c6\u201d\u7684\u75db\u70b9\u3002<\/p>\n<p>2.1.2 \u52a8\u4f5c\u7684Token\u5316&#xff08;Action Tokenization&#xff09;<\/p>\n<p>OpenVLA\u6cbf\u7528\u4e86RT-2\u7684\u52a8\u4f5c\u79bb\u6563\u5316\u601d\u8def&#xff0c;\u4f46\u5728\u7ec6\u8282\u4e0a\u8fdb\u884c\u4e86\u4f18\u5316\u3002<\/p>\n<ul>\n<li>\n<p>\u52a8\u4f5c\u7a7a\u95f4\u5b9a\u4e49&#xff1a; \u6a21\u578b\u8f93\u51fa\u7684\u662f7\u81ea\u7531\u5ea6&#xff08;7-DoF&#xff09;\u7684\u672b\u7aef\u6267\u884c\u5668\u52a8\u4f5c&#xff1a;$\\\\Delta x, \\\\Delta y, \\\\Delta z$&#xff08;\u4f4d\u7f6e\u53d8\u5316&#xff09;&#xff0c;$\\\\Delta roll, \\\\Delta pitch, \\\\Delta yaw$&#xff08;\u59ff\u6001\u53d8\u5316&#xff09;&#xff0c;\u4ee5\u53ca\u5939\u722a\u5f00\u5408\u72b6\u6001\u3002<\/p>\n<\/li>\n<li>\n<p>\u5206\u4f4d\u6570\u5206\u6876&#xff08;Quantile Binning&#xff09;&#xff1a; \u4e0e\u7b80\u5355\u7684\u7ebf\u6027\u5206\u6876\u4e0d\u540c&#xff0c;OpenVLA\u91c7\u7528\u4e86\u5206\u4f4d\u6570\u5206\u6876\u7b56\u7565\u3002\u7814\u7a76\u56e2\u961f\u7edf\u8ba1\u4e86Open X-Embodiment\u6570\u636e\u96c6\u4e2d\u6240\u6709\u52a8\u4f5c\u7684\u5206\u5e03&#xff0c;\u6839\u636e\u6570\u636e\u7684\u6982\u7387\u5bc6\u5ea6\u6765\u5212\u5206256\u4e2a\u6876&#xff08;bins&#xff09;\u3002\u8fd9\u610f\u5473\u7740\u5728\u52a8\u4f5c\u6570\u636e\u5bc6\u96c6\u7684\u533a\u57df&#xff08;\u901a\u5e38\u662f\u5fae\u5c0f\u7684\u7cbe\u7ec6\u8c03\u6574\u52a8\u4f5c&#xff09;&#xff0c;\u6876\u7684\u5bbd\u5ea6\u66f4\u7a84&#xff0c;\u5206\u8fa8\u7387\u66f4\u9ad8&#xff1b;\u800c\u5728\u7a00\u758f\u533a\u57df&#xff08;\u5927\u5e45\u5ea6\u7684\u5feb\u901f\u79fb\u52a8&#xff09;&#xff0c;\u6876\u66f4\u5bbd\u3002\u8fd9\u79cd\u975e\u7ebf\u6027\u79bb\u6563\u5316\u6781\u5927\u5730\u63d0\u5347\u4e86\u63a7\u5236\u7cbe\u5ea6\u3002<\/p>\n<\/li>\n<li>\n<p>\u8bcd\u8868\u8986\u5199&#xff1a; \u8fd9256\u4e2a\u52a8\u4f5ctoken\u76f4\u63a5\u8986\u5199\u4e86Llama 2\u8bcd\u8868\u4e2d\u539f\u672c\u4f7f\u7528\u9891\u7387\u6700\u4f4e\u7684256\u4e2atoken\u3002\u8fd9\u6837&#xff0c;Llama 2\u5c31\u88ab\u201c\u6b3a\u9a97\u201d\u7740\u53bb\u9884\u6d4b\u8fd9\u4e9b\u4ee3\u8868\u7269\u7406\u52a8\u4f5c\u7684\u8bcd&#xff0c;\u4ece\u800c\u65e0\u7f1d\u5229\u7528\u4e86\u5176\u5f3a\u5927\u7684\u5e8f\u5217\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>2.2 \u8bad\u7ec3\u914d\u65b9\u4e0e\u53c2\u6570\u9ad8\u6548\u5fae\u8c03&#xff08;PEFT&#xff09;<\/h4>\n<p>OpenVLA\u5728Open X-Embodiment\u6570\u636e\u96c6\u768497\u4e07\u6761\u8f68\u8ff9\u4e0a\u8fdb\u884c\u4e86\u5168\u91cf\u9884\u8bad\u7ec3\u3002\u7136\u800c&#xff0c;\u5176\u771f\u6b63\u7684\u8d21\u732e\u5728\u4e8e\u63a2\u7d22\u4e86\u5982\u4f55\u8ba9\u8fd9\u4e2a\u5e9e\u7136\u5927\u7269\u9002\u5e94\u65b0\u7684\u673a\u5668\u4eba\u3002<\/p>\n<p>\u5728\u673a\u5668\u4eba\u9886\u57df&#xff0c;\u5168\u91cf\u5fae\u8c03&#xff08;Full Fine-Tuning&#xff09;7B\u6a21\u578b\u4e0d\u4ec5\u663e\u5b58\u5f00\u9500\u5de8\u5927&#xff0c;\u800c\u4e14\u5bb9\u6613\u5bfc\u81f4\u201c\u707e\u96be\u6027\u9057\u5fd8\u201d&#xff08;Catastrophic Forgetting&#xff09;&#xff0c;\u5373\u6a21\u578b\u5b66\u4f1a\u4e86\u65b0\u673a\u5668\u4eba\u7684\u64cd\u4f5c&#xff0c;\u5374\u4e27\u5931\u4e86\u539f\u672c\u7684\u901a\u7528\u8bed\u4e49\u80fd\u529b\u3002OpenVLA\u56e2\u961f\u7cfb\u7edf\u6027\u5730\u8bc4\u4f30\u4e86LoRA (Low-Rank Adaptation) \u5728VLA\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n<ul>\n<li>\n<p>LoRA\u7684\u6709\u6548\u6027&#xff1a; \u5b9e\u9a8c\u8868\u660e&#xff0c;\u4ec5\u5bf9Attention\u5c42\u7684\u6743\u91cd\u77e9\u9635\u8fdb\u884c\u4f4e\u79e9\u66f4\u65b0&#xff08;\u4ec5\u5360\u603b\u53c2\u6570\u91cf\u76841.4%&#xff09;&#xff0c;\u5176\u6548\u679c\u751a\u81f3\u4f18\u4e8e\u6216\u6301\u5e73\u4e8e\u5168\u91cf\u5fae\u8c03 4\u3002\u8fd9\u8868\u660e&#xff0c;\u9884\u8bad\u7ec3\u7684VLA\u5df2\u7ecf\u638c\u63e1\u4e86\u7269\u7406\u4ea4\u4e92\u7684\u201c\u8bed\u6cd5\u201d&#xff0c;\u9002\u5e94\u65b0\u73af\u5883\u53ea\u9700\u8981\u8c03\u6574\u5c11\u91cf\u7684\u201c\u8bcd\u6c47\u6620\u5c04\u201d\u3002<\/p>\n<\/li>\n<li>\n<p>\u8bad\u7ec3\u6548\u7387&#xff1a; \u8fd9\u79cd\u65b9\u6cd5\u4f7f\u5f97OpenVLA\u53ef\u4ee5\u5728\u5355\u4e2a\u6d88\u8d39\u7ea7GPU&#xff08;\u5982RTX 3090\u62164090&#xff09;\u4e0a\u8fdb\u884c\u5fae\u8c03&#xff0c;\u6781\u5927\u5730\u964d\u4f4e\u4e86\u5177\u8eab\u667a\u80fd\u7814\u7a76\u7684\u95e8\u69db\u3002\u8fd9\u4e5f\u662fOpenVLA\u5728\u5f00\u6e90\u793e\u533a\u83b7\u5f97\u5e7f\u6cdb\u91c7\u7528\u7684\u5173\u952e\u539f\u56e0\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>2.3 \u5c40\u9650\u6027\u4e0e\u53cd\u601d<\/h4>\n<p>\u5c3d\u7ba1OpenVLA\u5728\u9759\u6001\u64cd\u4f5c\u4efb\u52a1&#xff08;\u5982Pick-and-Place&#xff09;\u4e0a\u8868\u73b0\u51fa\u8272&#xff0c;\u4f46\u5176\u81ea\u56de\u5f52\u7684\u672c\u8d28\u5e26\u6765\u4e86\u4e00\u4e2a\u4e0d\u53ef\u5ffd\u89c6\u7684\u7f3a\u9677&#xff1a;\u63a8\u7406\u5ef6\u8fdf\u3002\u751f\u6210\u4e00\u4e2a\u52a8\u4f5c\u9700\u8981LLM\u4e32\u884c\u9884\u6d4b7\u4e2a\u4ee5\u4e0a\u7684token&#xff0c;\u5bfc\u81f4\u63a7\u5236\u9891\u7387\u901a\u5e38\u57285Hz-10Hz\u5de6\u53f3\u3002\u5bf9\u4e8e\u9700\u8981\u5feb\u901f\u53cd\u5e94\u7684\u52a8\u6001\u4efb\u52a1&#xff08;\u5982\u63a5\u7403\u3001\u751a\u81f3\u5feb\u901f\u5207\u83dc&#xff09;&#xff0c;\u8fd9\u79cd\u9891\u7387\u663e\u5f97\u6349\u895f\u89c1\u8098\u3002\u8fd9\u4e5f\u4e3a\u540e\u7eed\u7684Octo\u548c$\\\\pi_0$\u7559\u51fa\u4e86\u521b\u65b0\u7a7a\u95f4\u3002<\/p>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>3. \u6838\u5fc3\u6280\u672f\u8def\u7ebf\u4e8c&#xff1a;\u8fde\u7eed\u751f\u6210\u5f0f VLA \u2014\u2014 \u4ece Octo \u5230 $\\\\pi_0$<\/p>\n<p>\u4e3a\u4e86\u7a81\u7834\u79bb\u6563\u5316\u5e26\u6765\u7684\u7cbe\u5ea6\u635f\u5931\u548c\u9891\u7387\u9650\u5236&#xff0c;2024-2025\u5e74\u7684\u53e6\u4e00\u6d3e\u7814\u7a76\u8005\u8f6c\u5411\u4e86\u751f\u6210\u5f0f\u6a21\u578b\u3002\u8fd9\u4e00\u6d41\u6d3e\u8ba4\u4e3a&#xff0c;\u52a8\u4f5c\u5206\u5e03\u672c\u8d28\u4e0a\u662f\u591a\u6a21\u6001\u7684&#xff08;\u540c\u4e00\u4e2a\u4efb\u52a1\u53ef\u80fd\u6709\u591a\u79cd\u89e3\u6cd5&#xff09;&#xff0c;\u4e14\u5e94\u5f53\u662f\u8fde\u7eed\u7684\u3002<\/p>\n<h4>3.1 Octo&#xff1a;\u6a21\u5757\u5316\u7684\u6269\u6563\u7b56\u7565\u4e13\u5bb6<\/h4>\n<p>Octo 5 \u662f\u7531Berkeley\u3001Stanford\u548cDeepMind\u8054\u5408\u63a8\u51fa\u7684\u5f00\u6e90\u901a\u7528\u673a\u5668\u4eba\u7b56\u7565\u3002\u5b83\u7684\u6838\u5fc3\u5728\u4e8e\u5c06Transformer\u7684\u4e3b\u5e72\u7f51\u7edc\u4e0e**\u6269\u6563\u6a21\u578b&#xff08;Diffusion Head&#xff09;**\u76f8\u7ed3\u5408\u3002<\/p>\n<p>3.1.1 \u6269\u6563\u7b56\u7565&#xff08;Diffusion Policy&#xff09;\u7684\u6570\u5b66\u539f\u7406<\/p>\n<p>Octo\u91c7\u7528\u4e86\u53bb\u566a\u6269\u6563\u6982\u7387\u6a21\u578b&#xff08;DDPM&#xff09;\u6765\u751f\u6210\u52a8\u4f5c\u3002\u5176\u6570\u5b66\u8868\u8ff0\u4e0d\u518d\u662f\u5206\u7c7b\u95ee\u9898&#xff08;\u9884\u6d4b\u54ea\u4e2abin&#xff09;&#xff0c;\u800c\u662f\u53bb\u566a\u95ee\u9898\u3002<\/p>\n<p>\u7ed9\u5b9a\u89c2\u6d4b o_t \u548c\u8bed\u8a00\u6307\u4ee4 l&#xff0c;\u6a21\u578b\u7684\u76ee\u6807\u662f\u4ece\u9ad8\u65af\u566a\u58f0 x_K \u223c &#x1d4a9;(0, I) \u4e2d\u9010\u6b65\u53bb\u566a&#xff0c;\u6062\u590d\u51fa\u52a8\u4f5c\u5e8f\u5217 x_0\u3002<\/p>\n<p>\u8bad\u7ec3\u65f6\u7684\u635f\u5931\u51fd\u6570\u901a\u5e38\u4e3a\u9884\u6d4b\u566a\u58f0\u7684\u5747\u65b9\u8bef\u5dee&#xff08;MSE&#xff09;&#xff1a;<\/p>\n<p>$$\\\\mathcal{L} &#061; \\\\mathbb{E}_{k, x_0, \\\\epsilon} \\\\left[ \\\\| \\\\epsilon &#8211; \\\\epsilon_\\\\theta(x_k, k, o_t, l) \\\\|^2 \\\\right]$$<\/p>\n<p>\u5176\u4e2d k \u662f\u6269\u6563\u6b65\u6570&#xff0c;\u03b5_\u03b8 \u662f\u795e\u7ecf\u7f51\u7edc\u9884\u6d4b\u7684\u566a\u58f0\u3002<\/p>\n<p>3.1.2 \u6a21\u5757\u5316\u8bbe\u8ba1\u4e0e\u89c2\u5bdf\u5386\u53f2<\/p>\n<p>Octo\u7684\u4e00\u4e2a\u663e\u8457\u7279\u70b9\u662f\u5176\u6a21\u5757\u5316&#xff08;Block-wise&#xff09;\u7684Transformer\u67b6\u6784\u3002\u5b83\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u8f93\u5165&#xff08;\u8155\u90e8\u76f8\u673a\u56fe\u50cf\u3001\u7b2c\u4e09\u4eba\u79f0\u89c6\u89d2\u56fe\u50cf\u3001\u672c\u4f53\u611f\u77e5\u6570\u636e\u3001\u8bed\u8a00\u6307\u4ee4&#xff09;\u5206\u522bToken\u5316&#xff0c;\u7136\u540e\u9001\u5165Transformer\u8fdb\u884c\u6ce8\u610f\u529b\u4ea4\u4e92\u3002 \u4e0eOpenVLA\u4e0d\u540c&#xff0c;Octo\u4e0d\u4ec5\u5173\u6ce8\u5f53\u524d\u5e27&#xff0c;\u8fd8\u663e\u5f0f\u5730\u5904\u7406\u89c2\u5bdf\u5386\u53f2&#xff08;Observation History&#xff09;\u3002\u8fd9\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u63a8\u65ad\u7269\u4f53\u7684\u901f\u5ea6\u548c\u52a0\u901f\u5ea6&#xff0c;\u5bf9\u4e8e\u52a8\u6001\u4efb\u52a1\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>3.1.3 \u52a8\u4f5c\u5206\u5757&#xff08;Action Chunking&#xff09;<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u63a8\u7406\u901f\u5ea6\u95ee\u9898\u5e76\u4fdd\u8bc1\u52a8\u4f5c\u7684\u5e73\u6ed1\u6027&#xff0c;Octo\u91c7\u7528\u4e86\u52a8\u4f5c\u5206\u5757\u6280\u672f\u3002\u6a21\u578b\u4e00\u6b21\u63a8\u7406\u4e0d\u662f\u9884\u6d4b\u8fd9\u4e00\u79d2\u7684\u52a8\u4f5c&#xff0c;\u800c\u662f\u9884\u6d4b\u672a\u6765 $H$ \u6b65&#xff08;\u4f8b\u598216\u6b65\u621632\u6b65&#xff09;\u7684\u52a8\u4f5c\u5e8f\u5217\u3002\u5728\u6267\u884c\u65f6&#xff0c;\u673a\u5668\u4eba\u53ef\u4ee5\u4f9d\u6b21\u6267\u884c\u8fd9\u4e00\u4e32\u52a8\u4f5c&#xff0c;\u6216\u8005\u901a\u8fc7\u201c\u91cd\u6258\u6c34\u5e73\u201d&#xff08;Receding Horizon&#xff09;\u63a7\u5236&#xff0c;\u6bcf\u6b65\u90fd\u91cd\u65b0\u9884\u6d4b\u5e76\u4ec5\u6267\u884c\u7b2c\u4e00\u6b65\u3002\u8fd9\u79cd\u673a\u5236\u7ed3\u5408\u6269\u6563\u6a21\u578b\u7684\u751f\u6210\u80fd\u529b&#xff0c;\u4f7f\u5f97Octo\u80fd\u591f\u4ea7\u751f\u6781\u5176\u5e73\u6ed1\u3001\u8fde\u8d2f\u7684\u8f68\u8ff9&#xff0c;\u907f\u514d\u4e86\u79bb\u6563\u5316\u6a21\u578b\u5e38\u89c1\u7684\u201c\u6296\u52a8\u201d\u73b0\u8c61\u3002<\/p>\n<h4>3.2 $\\\\pi_0$ (Pi-zero)&#xff1a;\u6d41\u5339\u914d\u5e26\u6765\u7684\u901f\u5ea6\u9769\u547d<\/h4>\n<p>\u5982\u679c\u8bf4Octo\u8bc1\u660e\u4e86\u6269\u6563\u6a21\u578b\u5728VLA\u4e2d\u7684\u6f5c\u529b&#xff0c;\u90a3\u4e48Physical Intelligence\u516c\u53f8\u63a8\u51fa\u7684 $\\\\pi_0$ 8 \u5219\u662f\u5c06\u8fd9\u4e00\u8def\u7ebf\u63a8\u5411\u4e86\u6781\u81f4&#xff0c;\u7279\u522b\u662f\u9488\u5bf9\u9ad8\u9891\u63a7\u5236\u7684\u4f18\u5316\u3002<\/p>\n<p>3.2.1 \u4ece\u6269\u6563\u5230\u6d41\u5339\u914d&#xff08;Flow Matching&#xff09;<\/p>\n<p>\u6269\u6563\u6a21\u578b\u867d\u7136\u751f\u6210\u8d28\u91cf\u9ad8&#xff0c;\u4f46\u5f80\u5f80\u9700\u8981\u51e0\u5341\u6b65\u7684\u53bb\u566a\u8fed\u4ee3&#xff0c;\u5bfc\u81f4\u63a8\u7406\u5ef6\u8fdf\u9ad8\u3002\u03c0\u2080 \u5f15\u5165\u4e86\u6d41\u5339\u914d&#xff08;Flow Matching&#xff09;\u6280\u672f\u3002\u6d41\u5339\u914d\u662f\u4e00\u79cd\u57fa\u4e8e\u5e38\u5fae\u5206\u65b9\u7a0b&#xff08;ODE&#xff09;\u7684\u751f\u6210\u6a21\u578b\u3002\u5b83\u4e0d\u662f\u53bb\u5b66\u4e60\u5982\u4f55\u53bb\u9664\u566a\u58f0&#xff0c;\u800c\u662f\u5b66\u4e60\u4e00\u4e2a\u5411\u91cf\u573a&#xff08;Vector Field&#xff09;v_t(x)&#xff0c;\u8be5\u5411\u91cf\u573a\u5b9a\u4e49\u4e86\u4ece\u566a\u58f0\u5206\u5e03\u76f4\u63a5&#034;\u6d41&#034;\u5411\u6570\u636e\u5206\u5e03\u7684\u6700\u4f73\u8def\u5f84&#xff08;\u901a\u5e38\u662f\u76f4\u7ebf\u8def\u5f84&#xff09;\u3002<\/p>\n<p>\u5176\u8bad\u7ec3\u76ee\u6807\u662f\u56de\u5f52\u8fd9\u4e2a\u5411\u91cf\u573a&#xff1a;<\/p>\n<p>$$\\\\mathcal{L}_{FM} &#061; \\\\mathbb{E}_{t, x_1, x_0} \\\\left[ \\\\| v_t(\\\\phi_t(x_1)) &#8211; (x_1 &#8211; x_0) \\\\|^2 \\\\right]$$<\/p>\n<p>\u8fd9\u91cc x_1 \u662f\u6570\u636e&#xff0c;x_0 \u662f\u566a\u58f0&#xff0c;\u03c6_t \u662f\u63d2\u503c\u8def\u5f84\u3002<\/p>\n<p>\u7531\u4e8e\u6d41\u5339\u914d\u53ef\u4ee5\u76f4\u63a5\u5b66\u4e60\u4ece\u566a\u58f0\u5230\u6570\u636e\u7684\u76f4\u7ebf\u8f68\u8ff9&#xff0c;\u63a8\u7406\u65f6\u53ea\u9700\u8981\u6781\u5c11\u7684\u6b65\u6570&#xff08;\u4f8b\u59821-10\u6b65&#xff09;\u5373\u53ef\u901a\u8fc7ODE\u6c42\u89e3\u5668\u751f\u6210\u9ad8\u8d28\u91cf\u52a8\u4f5c\u3002\u8fd9\u4f7f\u5f97 \u03c0\u2080 \u80fd\u591f\u5728\u4fdd\u6301\u5f3a\u5927\u751f\u6210\u80fd\u529b\u7684\u540c\u65f6&#xff0c;\u5b9e\u73b0\u9ad8\u8fbe50Hz\u7684\u63a7\u5236\u9891\u7387 8\u3002<\/p>\n<p>3.2.2 \u8de8\u5177\u8eab&#xff08;Cross-Embodiment&#xff09;\u4e0e\u534f\u540c\u8bad\u7ec3<\/p>\n<p>$\\\\pi_0$ \u7684\u53e6\u4e00\u4e2a\u6838\u5fc3\u8d21\u732e\u662f\u5176\u5927\u89c4\u6a21\u7684\u8de8\u5177\u8eab\u8bad\u7ec3\u3002\u5b83\u4e0d\u4ec5\u5728\u5355\u4e00\u7c7b\u578b\u7684\u673a\u5668\u4eba\u4e0a\u8bad\u7ec3&#xff0c;\u800c\u662f\u5728\u5305\u62ec\u5355\u81c2\u3001\u53cc\u81c2\u3001\u79fb\u52a8\u5e95\u76d8\u7b498\u79cd\u4e0d\u540c\u6784\u578b\u7684\u673a\u5668\u4eba\u6570\u636e\u4e0a\u8054\u5408\u8bad\u7ec3\u3002 \u66f4\u8fdb\u4e00\u6b65&#xff0c;\u5176\u5347\u7ea7\u7248 $\\\\pi_{0.5}$ \u5f15\u5165\u4e86**\u534f\u540c\u8bad\u7ec3&#xff08;Co-training&#xff09;**\u7b56\u7565 11\u3002\u6a21\u578b\u4e0d\u4ec5\u5b66\u4e60\u8f93\u51fa\u52a8\u4f5c&#xff0c;\u8fd8\u540c\u65f6\u5b66\u4e60\u7eaf\u6587\u672c\u548c\u7eaf\u89c6\u89c9\u4efb\u52a1&#xff08;\u5982VQA&#xff09;\u3002<\/p>\n<ul>\n<li>\n<p>\u601d\u7ef4\u94fe&#xff08;Chain of Thought&#xff09;\u5728\u673a\u5668\u4eba\u4e2d\u7684\u5e94\u7528&#xff1a; $\\\\pi_{0.5}$ \u5728\u6267\u884c\u4efb\u52a1\u65f6&#xff0c;\u4f1a\u5148\u5728\u5185\u90e8\u751f\u6210\u4e00\u4e2a\u9ad8\u5c42\u7684\u8bed\u4e49\u89c4\u5212&#xff08;High-level Action&#xff0c;\u6587\u672c\u5f62\u5f0f&#xff09;&#xff0c;\u4f8b\u5982\u201c\u6211\u73b0\u5728\u9700\u8981\u6298\u53e0\u886c\u886b\u7684\u5de6\u8896\u5b50\u201d&#xff0c;\u7136\u540e\u518d\u901a\u8fc7\u6d41\u5339\u914d\u5934&#xff08;Flow Matching Head&#xff09;\u751f\u6210\u5bf9\u5e94\u7684\u5e95\u5c42\u5173\u8282\u89d2\u5ea6\u3002\u8fd9\u79cd\u663e\u5f0f\u7684\u8bed\u4e49\u5f15\u5bfc\u6781\u5927\u5730\u589e\u5f3a\u4e86\u6a21\u578b\u5728\u957f\u7a0b\u590d\u6742\u4efb\u52a1&#xff08;\u5982\u6574\u7406\u623f\u95f4\u3001\u6298\u53e0\u8863\u7269&#xff09;\u4e2d\u7684\u9c81\u68d2\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>4. \u524d\u6cbf\u67b6\u6784\u4e09&#xff1a;\u53cc\u7cfb\u7edf\u4eba\u5f62\u673a\u5668\u4eba\u6a21\u578b \u2014\u2014 GR00T<\/p>\n<p>NVIDIA\u57282025\u5e74\u53d1\u5e03\u7684GR00T N1 9 \u4ee3\u8868\u4e86VLA\u5728\u4eba\u5f62\u673a\u5668\u4eba&#xff08;Humanoid&#xff09;\u9886\u57df\u7684\u6700\u9ad8\u6c34\u5e73\u3002\u4eba\u5f62\u673a\u5668\u4eba\u7684\u63a7\u5236\u96be\u5ea6\u8fdc\u8d85\u673a\u68b0\u81c2&#xff0c;\u56e0\u4e3a\u5b83\u4eec\u9700\u8981\u540c\u65f6\u5904\u7406\u5e73\u8861\u3001\u884c\u8d70\u548c\u53cc\u81c2\u64cd\u4f5c&#xff0c;\u81ea\u7531\u5ea6\u6781\u9ad8\u3002<\/p>\n<h4>4.1 System 1 \u4e0e System 2 \u7684\u8ba4\u77e5\u67b6\u6784<\/h4>\n<p>GR00T\u7684\u8bbe\u8ba1\u7075\u611f\u76f4\u63a5\u6765\u6e90\u4e8e\u8ba4\u77e5\u5fc3\u7406\u5b66\u4e2d\u7684\u53cc\u7cfb\u7edf\u7406\u8bba&#xff1a;<\/p>\n<ul>\n<li>\n<p>System 2&#xff08;\u6162\u601d\u8003&#xff09;&#xff1a; \u7531\u4e00\u4e2a\u5f3a\u5927\u7684VLM&#xff08;\u57fa\u4e8eNVIDIA Eagle-2&#xff09;\u6784\u6210\u3002\u5b83\u8fd0\u884c\u9891\u7387\u8f83\u4f4e&#xff08;\u598210Hz&#xff09;&#xff0c;\u8d1f\u8d23\u201c\u770b\u201d\u548c\u201c\u60f3\u201d\u3002\u5b83\u89e3\u6790\u590d\u6742\u7684\u73af\u5883\u8bed\u4e49&#xff0c;\u7406\u89e3\u6a21\u7cca\u7684\u8bed\u8a00\u6307\u4ee4&#xff0c;\u5e76\u751f\u6210\u4e00\u4e2a\u6f5c\u5728\u7684\u610f\u56fe\u5d4c\u5165&#xff08;Intent Embedding&#xff09;\u3002<\/p>\n<\/li>\n<li>\n<p>System 1&#xff08;\u5feb\u53cd\u5e94&#xff09;&#xff1a; \u7531\u4e00\u4e2a\u57fa\u4e8e**DiT&#xff08;Diffusion Transformer&#xff09;**\u7684\u52a8\u4f5c\u751f\u6210\u7f51\u7edc\u6784\u6210\u3002\u5b83\u8fd0\u884c\u9891\u7387\u6781\u9ad8&#xff08;\u598250-100Hz&#xff09;&#xff0c;\u8d1f\u8d23\u201c\u52a8\u201d\u3002\u5b83\u63a5\u6536System 2\u4f20\u6765\u7684\u610f\u56fe\u5d4c\u5165&#xff0c;\u7ed3\u5408\u5f53\u524d\u7684\u672c\u4f53\u611f\u77e5&#xff08;\u5173\u8282\u4f4d\u7f6e\u3001\u901f\u5ea6\u3001IMU\u6570\u636e&#xff09;&#xff0c;\u901a\u8fc7\u53bb\u566a\u751f\u6210\u5b9e\u65f6\u7684\u7535\u673a\u63a7\u5236\u4fe1\u53f7\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u79cd\u89e3\u8026\u8bbe\u8ba1\u5b8c\u7f8e\u89e3\u51b3\u4e86VLM\u63a8\u7406\u6162\u4e0e\u673a\u5668\u4eba\u63a7\u5236\u5feb\u4e4b\u95f4\u7684\u77db\u76fe\u3002System 2 \u5c31\u50cf\u5927\u8111\u76ae\u5c42&#xff0c;\u8d1f\u8d23\u89c4\u5212&#xff1b;System 1 \u5c31\u50cf\u5c0f\u8111\u548c\u810a\u9ad3&#xff0c;\u8d1f\u8d23\u5177\u4f53\u7684\u808c\u8089\u534f\u8c03\u3002<\/p>\n<h4>4.2 \u6570\u636e\u91d1\u5b57\u5854&#xff08;Data Pyramid&#xff09;\u4e0e\u4eff\u771f<\/h4>\n<p>GR00T\u7684\u6210\u529f\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5f52\u529f\u4e8e\u5176\u6570\u636e\u7b56\u7565\u3002\u7531\u4e8e\u771f\u5b9e\u4e16\u754c\u7684\u4eba\u5f62\u673a\u5668\u4eba\u6570\u636e\u6781\u5176\u7a00\u7f3a&#xff0c;NVIDIA\u6784\u5efa\u4e86\u4e00\u4e2a\u5e9e\u5927\u7684\u6570\u636e\u91d1\u5b57\u5854&#xff1a;<\/p>\n<ul>\n<li>\n<p>\u5854\u57fa&#xff08;Web Data&#xff09;&#xff1a; \u6570\u4ee5\u4ebf\u8ba1\u7684\u4e92\u8054\u7f51\u56fe\u6587\u6570\u636e&#xff0c;\u7528\u4e8e\u8bad\u7ec3System 2\u7684\u901a\u7528\u8bed\u4e49\u7406\u89e3\u3002<\/p>\n<\/li>\n<li>\n<p>\u5854\u8eab&#xff08;Simulation Data&#xff09;&#xff1a; \u5229\u7528Isaac Lab\u548cOSMO\u7f16\u6392\u7cfb\u7edf&#xff0c;\u5728\u7269\u7406\u4eff\u771f\u4e2d\u5e76\u884c\u751f\u6210\u6570\u767e\u4e07\u6761\u673a\u5668\u4eba\u64cd\u4f5c\u8f68\u8ff9\u3002\u901a\u8fc7\u57df\u968f\u673a\u5316&#xff08;Domain Randomization&#xff09;&#xff0c;\u4eff\u771f\u6570\u636e\u4e3aSystem 1\u63d0\u4f9b\u4e86\u6781\u5176\u4e30\u5bcc\u7684\u7269\u7406\u4ea4\u4e92\u5148\u9a8c\u3002<\/p>\n<\/li>\n<li>\n<p>\u5854\u5c16&#xff08;Real World Data&#xff09;&#xff1a; \u5c11\u91cf\u7684\u9ad8\u8d28\u91cf\u9065\u64cd\u4f5c\u6570\u636e&#xff0c;\u7528\u4e8e\u5c06\u6a21\u578b\u5bf9\u9f50\u5230\u771f\u5b9e\u7684\u7269\u7406\u4e16\u754c\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8fd9\u79cd\u201cSim-to-Real\u201d\u7684\u5927\u89c4\u6a21\u5e94\u7528&#xff0c;\u914d\u5408\u53cc\u7cfb\u7edf\u67b6\u6784&#xff0c;\u4f7f\u5f97GR00T\u80fd\u591f\u5728\u6ca1\u6709\u89c1\u8fc7\u7684\u573a\u666f\u4e2d\u5c55\u73b0\u51fa\u60ca\u4eba\u7684\u96f6\u6837\u672c&#xff08;Zero-shot&#xff09;\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>5. \u5176\u4ed6\u65b0\u5174\u6a21\u578b\u4e0e\u6280\u672f\u6982\u89c8<\/p>\n<p>\u9664\u4e86\u4e0a\u8ff0\u4e09\u5927\u5de8\u5934&#xff0c;2024-2025\u5e74\u8fd8\u6709\u591a\u4e2a\u503c\u5f97\u5173\u6ce8\u7684\u6a21\u578b\u586b\u8865\u4e86\u751f\u6001\u4f4d\u7684\u7a7a\u7f3a&#xff1a;<\/p>\n<table>\n<tbody>\n<tr>\n<td>\n<p>\u6a21\u578b<\/p>\n<\/td>\n<td>\n<p>\u6838\u5fc3\u7279\u70b9<\/p>\n<\/td>\n<td>\n<p>\u9002\u7528\u573a\u666f<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>SmolVLA<\/p>\n<\/td>\n<td>\n<p>\u8f7b\u91cf\u5316\u3002\u9488\u5bf9\u7aef\u4fa7\u90e8\u7f72\u4f18\u5316&#xff0c;\u53c2\u6570\u91cf\u6781\u5c0f&#xff0c;\u65e8\u5728\u5728\u7b97\u529b\u53d7\u9650\u7684\u5d4c\u5165\u5f0f\u8bbe\u5907\u4e0a\u8fd0\u884c\u3002<\/p>\n<\/td>\n<td>\n<p>\u79fb\u52a8\u673a\u5668\u4eba\u3001\u4f4e\u6210\u672c\u673a\u68b0\u81c2<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>Helix<\/p>\n<\/td>\n<td>\n<p>\u53cc\u7cfb\u7edf\u67b6\u6784\u5148\u9a71\u3002\u7531Figure AI\u5f00\u53d1&#xff0c;\u540c\u6837\u91c7\u7528LLM\u89c4\u5212&#043;\u7b56\u7565\u7f51\u7edc\u6267\u884c\u7684\u67b6\u6784&#xff0c;\u5f3a\u8c03\u7aef\u5230\u7aef\u7684\u795e\u7ecf\u7f51\u7edc\u63a7\u5236\u800c\u975e\u4f20\u7edf\u7684MPC\u3002<\/p>\n<\/td>\n<td>\n<p>Figure 01\/02 \u4eba\u5f62\u673a\u5668\u4eba<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>Gemini Robotics<\/p>\n<\/td>\n<td>\n<p>\u4e91\u7aef\u534f\u540c\u3002\u5229\u7528Gemini 2.0\u7684\u8d85\u957f\u4e0a\u4e0b\u6587\u548c\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\u5728\u4e91\u7aef\u505a\u89c4\u5212&#xff0c;\u914d\u5408\u7aef\u4fa7\u8f7b\u91cf\u7ea7\u7b56\u7565\u6267\u884c&#xff0c;\u89e3\u51b3\u4e86\u957f\u7a0b\u4efb\u52a1\u8bb0\u5fc6\u95ee\u9898\u3002<\/p>\n<\/td>\n<td>\n<p>\u590d\u6742\u5bb6\u5ead\u670d\u52a1\u4efb\u52a1<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>RDT-1B<\/p>\n<\/td>\n<td>\n<p>\u53cc\u81c2\u534f\u540c\u3002\u4e13\u95e8\u9488\u5bf9\u53cc\u81c2\u64cd\u4f5c\u4f18\u5316\u7684\u6269\u6563Transformer&#xff0c;\u89e3\u51b3\u4e86\u53cc\u81c2\u534f\u8c03\u4e2d\u7684\u9ad8\u7ef4\u7a7a\u95f4\u63a2\u7d22\u96be\u9898\u3002<\/p>\n<\/td>\n<td>\n<p>\u53cc\u81c2\u7ec4\u88c5\u3001\u7cbe\u7ec6\u64cd\u4f5c<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>6. \u6570\u636e&#xff1a;VLA\u80fd\u529b\u7684\u6e90\u6cc9<\/p>\n<h4>6.1 Open X-Embodiment \u6570\u636e\u96c6<\/h4>\n<p>\u4e0a\u8ff0\u6240\u6709\u6a21\u578b\u7684\u6210\u529f\u90fd\u79bb\u4e0d\u5f00Open X-Embodiment\u6570\u636e\u96c6 6\u3002\u8fd9\u662f\u4e00\u4e2a\u7531\u5168\u740334\u4e2a\u673a\u5668\u4eba\u5b9e\u9a8c\u5ba4\u8054\u5408\u6784\u5efa\u7684\u6570\u636e\u96c6&#xff0c;\u5305\u542b\u4e86&#xff1a;<\/p>\n<ul>\n<li>\n<p>22\u79cd\u673a\u5668\u4eba\u6784\u578b&#xff1a; \u4eceFranka\u3001UR5\u7b49\u5de5\u4e1a\u81c2&#xff0c;\u5230WidowX\u7b49\u4f4e\u6210\u672c\u81c2&#xff0c;\u518d\u5230\u5404\u79cd\u79fb\u52a8\u5e95\u76d8\u3002<\/p>\n<\/li>\n<li>\n<p>100\u4e07&#043; \u8f68\u8ff9&#xff1a; \u6db5\u76d6\u4e86\u4ece\u7b80\u5355\u7684\u62fe\u53d6\u653e\u7f6e\u5230\u590d\u6742\u7684\u7269\u4f53\u64cd\u4f5c\u3002<\/p>\n<\/li>\n<li>\n<p>\u6807\u51c6\u5316\u683c\u5f0f&#xff08;RLDS&#xff09;&#xff1a; \u7edf\u4e00\u4e86\u6570\u636e\u683c\u5f0f&#xff0c;\u4f7f\u5f97\u8de8\u5177\u8eab\u8bad\u7ec3\u6210\u4e3a\u53ef\u80fd\u3002<\/p>\n<\/li>\n<\/ul>\n<h4>6.2 \u6570\u636e\u7684\u201c\u5f02\u8d28\u6027\u201d\u6311\u6218\u4e0e\u5e94\u5bf9<\/h4>\n<p>\u5728\u8bad\u7ec3VLA\u65f6&#xff0c;\u4e0d\u540c\u6765\u6e90\u7684\u6570\u636e\u5f80\u5f80\u5177\u6709\u4e0d\u540c\u7684\u6a21\u6001&#xff08;\u6709\u7684\u6709\u8bed\u8a00\u6807\u6ce8&#xff0c;\u6709\u7684\u6ca1\u6709&#xff09;\u3001\u4e0d\u540c\u7684\u89c6\u89d2&#xff08;\u8155\u90e8\u76f8\u673a vs \u5168\u5c40\u76f8\u673a&#xff09;\u548c\u4e0d\u540c\u7684\u63a7\u5236\u9891\u7387\u3002<\/p>\n<ul>\n<li>\n<p>\u6df7\u5408\u7b56\u7565&#xff1a; OpenVLA\u548cOcto\u90fd\u91c7\u7528\u4e86\u7279\u5b9a\u7684\u6570\u636e\u6df7\u5408\u6bd4\u4f8b&#xff08;Mixture Ratios&#xff09;&#xff0c;\u901a\u5e38\u4f1a\u7ed9\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u6027\u5f3a\u7684\u5b50\u6570\u636e\u96c6&#xff08;\u5982Bridge Data V2&#xff09;\u66f4\u9ad8\u7684\u91c7\u6837\u6743\u91cd\u3002<\/p>\n<\/li>\n<li>\n<p>\u52a8\u4f5c\u91cd\u6807\u6ce8&#xff1a; \u4e3a\u4e86\u7edf\u4e00\u52a8\u4f5c\u7a7a\u95f4&#xff0c;\u7814\u7a76\u8005\u901a\u5e38\u4f1a\u5c06\u4e0d\u540c\u673a\u5668\u4eba\u7684\u52a8\u4f5c\u7edf\u4e00\u6620\u5c04\u5230\u4ece\u672b\u7aef\u6267\u884c\u5668\u5750\u6807\u7cfb\u4e0b\u76846-DoF\u4f4d\u59ff\u53d8\u5316 &#043; \u5939\u722a\u72b6\u6001\u3002<\/p>\n<\/li>\n<\/ul>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>7. \u5b66\u672f\u8bba\u6587\u5199\u4f5c\u6307\u5357&#xff1a;\u5982\u4f55\u64b0\u5199\u201c\u65b9\u6cd5&#xff08;Method&#xff09;\u201d\u90e8\u5206<\/p>\n<p>\u57fa\u4e8e\u5bf9OpenVLA\u3001Octo\u3001$\\\\pi_0$\u7b49\u9876\u7ea7\u8bba\u6587\u7684\u6df1\u5ea6\u89e3\u6784&#xff0c;\u672c\u8282\u4e3a\u7814\u7a76\u4eba\u5458\u63d0\u4f9b\u4e00\u4efd\u5173\u4e8e\u5982\u4f55\u64b0\u5199VLA\u8bba\u6587\u201c\u65b9\u6cd5\u201d\u90e8\u5206\u7684\u7ed3\u6784\u5316\u6307\u5357\u3002\u8be5\u90e8\u5206\u662f\u8bba\u6587\u7684\u6838\u5fc3&#xff0c;\u65e8\u5728\u5efa\u7acb\u6570\u5b66\u5f62\u5f0f\u5316\u3001\u89e3\u91ca\u6280\u672f\u51b3\u7b56\u5e76\u4fdd\u8bc1\u53ef\u590d\u73b0\u6027\u3002<\/p>\n<h4>7.1 \u6838\u5fc3\u5199\u4f5c\u539f\u5219<\/h4>\n<li>\n<p>\u6570\u5b66\u5f62\u5f0f\u5316\u5148\u884c&#xff08;Formalization First&#xff09;&#xff1a; \u4e0d\u8981\u4e00\u4e0a\u6765\u5c31\u5806\u780c\u7f51\u7edc\u7ed3\u6784\u56fe\u3002\u9996\u5148\u8981\u5b9a\u4e49\u4f60\u7684\u63a7\u5236\u95ee\u9898\u3002\u662f\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b&#xff08;MDP&#xff09;&#xff1f;\u8fd8\u662f\u90e8\u5206\u53ef\u89c2\u6d4b&#xff08;POMDP&#xff09;&#xff1f;<\/p>\n<\/li>\n<li>\n<p>\u8f93\u5165-\u5904\u7406-\u8f93\u51fa&#xff08;Input-Process-Output&#xff09;\u903b\u8f91&#xff1a; \u6309\u7167\u6570\u636e\u6d41\u52a8\u7684\u65b9\u5411\u7ec4\u7ec7\u7ae0\u8282\u3002\u611f\u77e5 -&gt; \u63a8\u7406\/\u878d\u5408 -&gt; \u751f\u6210\u3002<\/p>\n<\/li>\n<li>\n<p>\u4e3a\u6bcf\u4e00\u4e2a\u8bbe\u8ba1\u51b3\u7b56\u8fa9\u62a4&#xff08;Justify Every Choice&#xff09;&#xff1a; \u5ba1\u7a3f\u4eba\u4e0d\u4ec5\u60f3\u77e5\u9053\u4f60\u505a\u4e86\u4ec0\u4e48&#xff0c;\u66f4\u60f3\u77e5\u9053\u4e3a\u4ec0\u4e48\u3002\u4e3a\u4ec0\u4e48\u9009SigLIP\u800c\u4e0d\u662fCLIP&#xff1f;\u4e3a\u4ec0\u4e48\u9009\u6d41\u5339\u914d\u800c\u4e0d\u662f\u6269\u6563&#xff1f;&#xff08;\u4f8b\u5982&#xff1a;\u201c\u4e3a\u4e86\u89e3\u51b3\u6269\u6563\u6a21\u578b\u7684\u63a8\u7406\u5ef6\u8fdf\u95ee\u9898\u2026\u2026\u201d&#xff09;\u3002<\/p>\n<\/li>\n<h4>7.2 \u63a8\u8350\u7684\u7ed3\u6784\u5927\u7eb2\u4e0e\u5185\u5bb9\u7ec6\u8282<\/h4>\n<p>7.2.1 \u9884\u5907\u77e5\u8bc6\u4e0e\u95ee\u9898\u5b9a\u4e49 (Preliminaries &amp; Problem Formulation)<\/p>\n<ul>\n<li>\n<p>\u5185\u5bb9&#xff1a;<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u5b9a\u4e49\u7b56\u7565 \u03c0_\u03b8(a_t | o_t, l)\u3002<\/p>\n<\/li>\n<li>\n<p>\u89c2\u6d4b\u7a7a\u95f4 O&#xff1a;\u660e\u786e\u5b9a\u4e49 o_t \u5305\u542b\u4ec0\u4e48\u3002\u4f8b\u5982&#xff1a;o_t &#061; {I_wrist, I_third, p_prop}&#xff0c;\u5176\u4e2d I \u2208 \u211d^{H\u00d7W\u00d73}\u3002<\/p>\n<\/li>\n<li>\n<p>\u52a8\u4f5c\u7a7a\u95f4 A&#xff1a;\u660e\u786e\u5b9a\u4e49 a_t\u3002\u662f\u5173\u8282\u89d2\u5ea6&#xff08;Joint Positions&#xff09;\u8fd8\u662f\u672b\u7aef\u4f4d\u59ff&#xff08;EEF Pose&#xff09;&#xff1f;\u662f\u7edd\u5bf9\u5750\u6807\u8fd8\u662f\u76f8\u5bf9\u589e\u91cf&#xff08;Delta&#xff09;&#xff1f;<\/p>\n<\/li>\n<li>\n<p>\u793a\u4f8b\u53e5\u5f0f&#xff1a;&#034;We formulate the vision-language-action task as learning a policy \u03c0 that maps a sequence of observations and a natural language instruction to a sequence of future actions.&#034;<\/p>\n<\/li>\n<\/ul>\n<p>7.2.2 \u6a21\u578b\u67b6\u6784 (Model Architecture)<\/p>\n<p>\u8fd9\u662f\u6700\u957f\u3001\u6700\u8be6\u7ec6\u7684\u5c0f\u8282\u3002\u5efa\u8bae\u914d\u5408\u4e00\u5f20\u6e05\u6670\u7684\u7cfb\u7edf\u6846\u56fe&#xff08;System Overview&#xff09;\u6765\u5199\u3002<\/p>\n<ul>\n<li>\n<p>\u89c6\u89c9\u7f16\u7801\u5668 (Visual Encoders)&#xff1a;<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u63cf\u8ff0\u4f7f\u7528\u7684Backbone&#xff08;\u5982SigLIP-SO400M&#xff09;\u3002<\/p>\n<\/li>\n<li>\n<p>\u89e3\u91ca\u7279\u5f81\u63d0\u53d6\u7684\u7ec6\u8282&#xff08;\u5982&#xff1a;\u63d0\u53d6\u5012\u6570\u7b2c\u4e8c\u5c42\u7684Patch Features&#xff09;\u3002<\/p>\n<\/li>\n<li>\n<p>\u5173\u952e\u70b9&#xff1a;\u5982\u679c\u4f7f\u7528\u4e86\u591a\u7f16\u7801\u5668&#xff08;\u5982OpenVLA&#xff09;&#xff0c;\u5fc5\u987b\u89e3\u91ca\u878d\u5408\u673a\u5236&#xff08;Concatenation, Cross-Attention, or MLP Projection&#xff09;\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>VLM\/Transformer\u4e3b\u5e72 (Backbone)&#xff1a;<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u6307\u5b9aLLM\u578b\u53f7&#xff08;\u5982Llama 2 7B&#xff09;\u3002<\/p>\n<\/li>\n<li>\n<p>\u63cf\u8ff0\u591a\u6a21\u6001\u5bf9\u9f50\u65b9\u6cd5&#xff08;Projector\u7684\u8bbe\u8ba1&#xff09;\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u52a8\u4f5c\u5934\/\u89e3\u7801\u5668 (Action Head\/Decoder)&#xff1a;\u8fd9\u662fVLA\u8bba\u6587\u7684\u7075\u9b42\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u5982\u679c\u662f\u81ea\u56de\u5f52\u6a21\u578b&#xff08;OpenVLA&#xff09;&#xff1a;\u5fc5\u987b\u8be6\u7ec6\u63cf\u8ff0\u5206\u8bcd&#xff08;Tokenization&#xff09;\u8fc7\u7a0b\u3002\u7ed9\u51fa\u5206\u4f4d\u6570\u5206\u6876\u7684\u516c\u5f0f&#xff0c;\u89e3\u91ca\u8bcd\u8868\u6269\u5145\u7684\u7ec6\u8282\u3002<\/p>\n<\/li>\n<li>\n<p>\u5982\u679c\u662f\u751f\u6210\u5f0f\u6a21\u578b&#xff08;Octo\/\u03c0\u2080&#xff09;&#xff1a;\u5fc5\u987b\u5199\u51fa\u751f\u6210\u8fc7\u7a0b\u7684\u6570\u5b66\u516c\u5f0f\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u6269\u6563\u516c\u5f0f&#xff1a;\u2112_Diff &#061; ||\u03b5 &#8211; \u03b5_\u03b8(x_t, t, cond)||\u00b2<\/p>\n<\/li>\n<li>\n<p>\u6d41\u5339\u914d\u516c\u5f0f&#xff1a;\u5b9a\u4e49\u5411\u91cf\u573a\u56de\u5f52\u76ee\u6807\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u5173\u952e\u70b9&#xff1a;\u89e3\u91ca\u52a8\u4f5c\u5206\u5757&#xff08;Chunking&#xff09;\u7b56\u7565\u3002\u4e3a\u4ec0\u4e48\u4e00\u6b21\u9884\u6d4b H \u6b65&#xff1f;\u8fd9\u5bf9\u5e73\u6ed1\u6027\u6709\u4f55\u5f71\u54cd&#xff1f;<\/p>\n<\/li>\n<\/ul>\n<p>7.2.3 \u8bad\u7ec3\u65b9\u6cd5 (Training Methodology)<\/p>\n<ul>\n<li>\n<p>\u8bad\u7ec3\u76ee\u6807 (Training Objective)&#xff1a;\u5199\u51fa\u603b\u7684\u635f\u5931\u51fd\u6570\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u4f8b\u5982&#xff1a;\u2112_total &#061; \u2112_action &#043; \u03bb\u2112_auxiliary<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u4f8b\u5982&#xff1a;$\\\\mathcal{L}_{total} &#061; \\\\mathcal{L}_{action} &#043; \\\\lambda \\\\mathcal{L}_{auxiliary}$\u3002<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u5fae\u8c03\u7b56\u7565 (Fine-tuning Strategy)&#xff1a;<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u8be6\u7ec6\u63cf\u8ff0\u53c2\u6570\u51bb\u7ed3\u60c5\u51b5\u3002\u89c6\u89c9\u7f16\u7801\u5668\u51bb\u7ed3\u4e86\u5417&#xff1f;LLM\u662f\u5168\u91cf\u5fae\u8c03\u8fd8\u662fLoRA&#xff1f;<\/p>\n<\/li>\n<li>\n<p>\u5982\u679c\u662fLoRA&#xff0c;\u79e9&#xff08;Rank&#xff09;\u662f\u591a\u5c11&#xff1f;\u5e94\u7528\u5728\u54ea\u4e9b\u6a21\u5757&#xff08;q, k, v, o&#xff09;&#xff1f;<\/p>\n<\/li>\n<\/ul>\n<p>7.2.4 \u63a8\u7406\u4e0e\u90e8\u7f72 (Inference &amp; Deployment)<\/p>\n<p>\u5f88\u591a\u8bba\u6587\u5bb9\u6613\u5ffd\u7565\u8fd9\u4e00\u90e8\u5206&#xff0c;\u4f46\u5b83\u5bf9\u590d\u73b0\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<ul>\n<li>\n<p>\u5185\u5bb9&#xff1a;<\/p>\n<\/li>\n<\/ul>\n<ul>\n<li>\n<p>\u65f6\u5e8f\u96c6\u6210 (Temporal Ensembling)&#xff1a; \u5982\u679c\u4f7f\u7528\u4e86\u52a8\u4f5c\u5206\u5757&#xff0c;\u5982\u4f55\u5904\u7406\u91cd\u53e0\u7684\u52a8\u4f5c\u5e8f\u5217&#xff1f;\u662f\u53ea\u53d6\u7b2c\u4e00\u6b65&#xff0c;\u8fd8\u662f\u52a0\u6743\u5e73\u5747&#xff1f;<\/p>\n<\/li>\n<li>\n<p>\u63a7\u5236\u9891\u7387 (Control Frequency)&#xff1a; \u6a21\u578b\u7684\u63a8\u7406\u8017\u65f6\u662f\u591a\u5c11&#xff1f;\u5b9e\u9645\u4e0b\u53d1\u7ed9\u673a\u5668\u4eba\u7684\u9891\u7387\u662f\u591a\u5c11&#xff1f;<\/p>\n<\/li>\n<li>\n<p>\u786c\u4ef6\u7ec6\u8282&#xff1a; \u8fd9\u4e00\u90e8\u5206\u53ef\u4ee5\u653e\u5728\u5b9e\u9a8c\u8bbe\u7f6e\u4e2d&#xff0c;\u4f46\u5982\u679c\u6d89\u53ca\u7aef\u4fa7\u4f18\u5316&#xff08;\u5982\u91cf\u5316&#xff09;&#xff0c;\u5e94\u5728\u6b64\u5904\u8bf4\u660e\u3002<\/p>\n<\/li>\n<\/ul>\n<h3>\u00a0<\/h3>\n<hr \/>\n<p>8. \u7ed3\u8bba\u4e0e\u5c55\u671b<\/p>\n<p>2024-2025\u5e74\u662f\u5177\u8eab\u667a\u80fd\u4ece\u201c\u4f5c\u574a\u5f0f\u201d\u8d70\u5411\u201c\u5de5\u4e1a\u5316\u201d\u7684\u5173\u952e\u65f6\u671f\u3002<\/p>\n<ul>\n<li>\n<p>\u6a21\u578b\u5c42\u9762&#xff1a; \u6211\u4eec\u89c1\u8bc1\u4e86\u4ece\u7b80\u5355\u7684\u81ea\u56de\u5f52Token\u9884\u6d4b&#xff08;OpenVLA&#xff09;\u5411\u66f4\u7b26\u5408\u7269\u7406\u89c4\u5f8b\u7684\u8fde\u7eed\u751f\u6210\u6a21\u578b&#xff08;$\\\\pi_0$, GR00T&#xff09;\u7684\u6f14\u8fdb\u3002<\/p>\n<\/li>\n<li>\n<p>\u67b6\u6784\u5c42\u9762&#xff1a; \u53cc\u7cfb\u7edf&#xff08;System 1\/2&#xff09;\u67b6\u6784\u7684\u6210\u719f&#xff0c;\u6807\u5fd7\u7740\u673a\u5668\u4eba\u5f00\u59cb\u62e5\u6709\u7c7b\u4f3c\u4eba\u7c7b\u7684\u201c\u601d\u8003\u201d\u4e0e\u201c\u672c\u80fd\u201d\u5206\u79bb\u7684\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p>\u6570\u636e\u5c42\u9762&#xff1a; Open X-Embodiment\u548c\u5408\u6210\u6570\u636e\u91d1\u5b57\u5854\u7684\u5efa\u7acb&#xff0c;\u521d\u6b65\u7f13\u89e3\u4e86\u56f0\u6270\u673a\u5668\u4eba\u9886\u57df\u51e0\u5341\u5e74\u7684\u6570\u636e\u9965\u6e34\u95ee\u9898\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u5bf9\u4e8e\u672a\u6765\u7684\u7814\u7a76\u8005\u800c\u8a00&#xff0c;\u5355\u4e00\u7684\u6a21\u578b\u67b6\u6784\u521b\u65b0\u53ef\u80fd\u4e0d\u518d\u662f\u552f\u4e00\u7684\u51fa\u8def\u3002\u5982\u4f55\u9ad8\u6548\u5730\u5229\u7528\u5f02\u6784\u6570\u636e\u3001\u5982\u4f55\u5b9e\u73b0\u66f4\u5f3a\u7684Sim-to-Real\u8fc1\u79fb\u3001\u4ee5\u53ca\u5982\u4f55\u8ba9VLA\u6a21\u578b\u5177\u5907\u957f\u7a0b\u7684\u56e0\u679c\u63a8\u7406\u80fd\u529b&#xff08;\u800c\u4e0d\u4ec5\u4ec5\u662f\u52a8\u4f5c\u514b\u9686&#xff09;&#xff0c;\u5c06\u662f\u63a5\u4e0b\u6765\u7684\u6838\u5fc3\u547d\u9898\u3002\u64b0\u5199\u9ad8\u8d28\u91cf\u7684\u8bba\u6587&#xff0c;\u4e0d\u4ec5\u9700\u8981\u624e\u5b9e\u7684\u5b9e\u9a8c&#xff0c;\u66f4\u9700\u8981\u6e05\u6670\u3001\u89c4\u8303\u4e14\u5177\u6709\u6d1e\u5bdf\u529b\u7684\u65b9\u6cd5\u8bba\u8868\u8ff0&#xff0c;\u5e0c\u671b\u672c\u62a5\u544a\u7684\u5199\u4f5c\u6307\u5357\u80fd\u4e3a\u6b64\u63d0\u4f9b\u6709\u76ca\u7684\u53c2\u8003\u3002<\/p>\n<h4>\u6570\u636e\u5f15\u7528\u8868\u683c&#xff1a;\u4e3b\u6d41\u6a21\u578b\u53c2\u6570\u5bf9\u6bd4<\/h4>\n<table>\n<tbody>\n<tr>\n<td>\n<p>\u7279\u6027<\/p>\n<\/td>\n<td>\n<p>OpenVLA<\/p>\n<\/td>\n<td>\n<p>Octo<\/p>\n<\/td>\n<td>\n<p>\u03c00\u200b (Pi-zero)<\/p>\n<\/td>\n<td>\n<p>GR00T N1<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u6838\u5fc3\u67b6\u6784<\/p>\n<\/td>\n<td>\n<p>Prismatic VLM (Llama 2)<\/p>\n<\/td>\n<td>\n<p>Transformer &#043; Diffusion<\/p>\n<\/td>\n<td>\n<p>VLM &#043; Flow Matching<\/p>\n<\/td>\n<td>\n<p>System 2 (VLM) &#043; System 1 (DiT)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u52a8\u4f5c\u8868\u793a<\/p>\n<\/td>\n<td>\n<p>\u79bb\u6563 Token (256 bins)<\/p>\n<\/td>\n<td>\n<p>\u8fde\u7eed\u5411\u91cf (Continuous)<\/p>\n<\/td>\n<td>\n<p>\u8fde\u7eed\u5411\u91cf (Continuous)<\/p>\n<\/td>\n<td>\n<p>\u8fde\u7eed\u5411\u91cf (Continuous)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u751f\u6210\u673a\u5236<\/p>\n<\/td>\n<td>\n<p>\u81ea\u56de\u5f52 (Autoregressive)<\/p>\n<\/td>\n<td>\n<p>\u6269\u6563\u53bb\u566a (Denoising)<\/p>\n<\/td>\n<td>\n<p>\u5411\u91cf\u573a\u56de\u5f52 (ODE)<\/p>\n<\/td>\n<td>\n<p>\u53bb\u566a\/\u6d41\u5339\u914d<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u63a8\u7406\u9891\u7387<\/p>\n<\/td>\n<td>\n<p>\u4f4e (&lt;10Hz)<\/p>\n<\/td>\n<td>\n<p>\u4e2d<\/p>\n<\/td>\n<td>\n<p>\u9ad8 (50Hz&#043;)<\/p>\n<\/td>\n<td>\n<p>\u6781\u9ad8 (System 1: 100Hz)<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u5fae\u8c03\u65b9\u5f0f<\/p>\n<\/td>\n<td>\n<p>LoRA (PEFT)<\/p>\n<\/td>\n<td>\n<p>Full \/ Head Only<\/p>\n<\/td>\n<td>\n<p>Full \/ Co-training<\/p>\n<\/td>\n<td>\n<p>Domain Adaptation<\/p>\n<\/td>\n<\/tr>\n<tr>\n<td>\n<p>\u4e3b\u8981\u6570\u636e\u96c6<\/p>\n<\/td>\n<td>\n<p>Open X-Embodiment<\/p>\n<\/td>\n<td>\n<p>Open X-Embodiment<\/p>\n<\/td>\n<td>\n<p>Proprietary &#043; Open Mix<\/p>\n<\/td>\n<td>\n<p>Data Pyramid (Web&#043;Sim&#043;Real)<\/p>\n<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>(\u6ce8&#xff1a;\u672c\u62a5\u544a\u4e2d\u6d89\u53ca\u7684\u5177\u4f53\u53c2\u6570\u548c\u6027\u80fd\u6307\u6807\u5747\u57fa\u4e8e2024-2025\u5e74\u516c\u5f00\u53d1\u8868\u7684\u8bba\u6587\u53ca\u6280\u672f\u62a5\u544a\u6574\u7406\u3002)<\/p>\n<p>\u5f15\u7528\u7684\u8457\u4f5c<\/p>\n<li>\n<p>OpenVLA: An Open-Source Vision-Language-Action Model &#8211; OpenReview, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; Verifying your browser | OpenReview<\/p>\n<\/li>\n<li>\n<p>OpenVLA: An Open-Source Vision-Language-Action Model &#8211; arXiv, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/arxiv.org\/html\/2406.09246v1<\/p>\n<\/li>\n<li>\n<p>[2406.09246] OpenVLA: An Open-Source Vision-Language-Action Model &#8211; arXiv, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; [2406.09246] OpenVLA: An Open-Source Vision-Language-Action Model<\/p>\n<\/li>\n<li>\n<p>OpenVLA: An Open-Source Vision-Language-Action Model, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; OpenVLA: An Open-Source Vision-Language-Action Model<\/p>\n<\/li>\n<li>\n<p>&#x1f419; Octo: An Open-Source Generalist Robot Policy, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; &#x1f419; Octo: An Open-Source Generalist Robot Policy<\/p>\n<\/li>\n<li>\n<p>An Open-Source Generalist Robot Policy &#8211; Octo, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/octo-models.github.io\/paper.pdf<\/p>\n<\/li>\n<li>\n<p>[2405.12213] Octo: An Open-Source Generalist Robot Policy &#8211; arXiv, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; [2405.12213] Octo: An Open-Source Generalist Robot Policy<\/p>\n<\/li>\n<li>\n<p>\u03c0 0 : Our First Generalist Policy &#8211; Physical Intelligence, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; Our First Generalist Policy<\/p>\n<\/li>\n<li>\n<p>Vision-language-action model &#8211; Wikipedia, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/en.wikipedia.org\/wiki\/Vision-language-action_model<\/p>\n<\/li>\n<li>\n<p>[Paper Review] Pi0, Pi0.5, Pi0-FAST &#8211; Tracing the Path of Physical Intelligence (PI), \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/bequiet-log.vercel.app\/pi-review<\/p>\n<\/li>\n<li>\n<p>A VLA with Open-World Generalization &#8211; Physical Intelligence, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; A VLA with Open-World Generalization<\/p>\n<\/li>\n<li>\n<p>GR00T N1: An Open Foundation Model for Generalist Humanoid Robots &#8211; arXiv, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; [2503.14734] GR00T N1: An Open Foundation Model for Generalist Humanoid Robots<\/p>\n<\/li>\n<li>\n<p>GR00T N1: An Open Foundation Model for Generalist Humanoid Robots &#8211; arXiv, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/arxiv.org\/pdf\/2503.14734<\/p>\n<\/li>\n<li>\n<p>Paper notes: OpenVLA. With the large-scale accessibility of\u2026 | by Jay Vakil | Toward Humanoids | Medium, \u8bbf\u95ee\u65f6\u95f4\u4e3a \u4e8c\u6708 1, 2026&#xff0c; https:\/\/medium.com\/correll-lab\/paper-notes-openvla-17540381187e<\/p>\n<\/li>\n<p>\u00a0<\/p>\n","protected":false},"excerpt":{"rendered":"<p>1. \u5f15\u8a00&#xff1a;\u5177\u8eab\u667a\u80fd\u7684\u201c\u57fa\u7840\u6a21\u578b\u201d\u65f6\u4ee3<br \/>\n1.1 \u4ece\u5355\u4e00\u4efb\u52a1\u5230\u901a\u7528\u7b56\u7565\u7684\u8303\u5f0f\u8f6c\u79fb<br \/>\n\u57282024\u5e74\u81f32025\u5e74\u521d\u7684\u8fd9\u6bb5\u65f6\u95f4\u91cc&#xff0c;\u673a\u5668\u4eba\u5b66\u4e60&#xff08;Robot Learning&#xff09;\u9886\u57df\u7ecf\u5386\u4e86\u4e00\u573a\u6df1\u523b\u7684\u8303\u5f0f\u8f6c\u79fb&#xff0c;\u5176\u6838\u5fc3\u5728\u4e8e\u4ece\u9488\u5bf9\u7279\u5b9a\u73af\u5883\u3001\u7279\u5b9a\u4efb\u52a1\u7684\u201c\u4e13\u5bb6\u6a21\u578b\u201d\u5411\u5177\u5907\u5e7f\u6cdb\u6cdb\u5316\u80fd\u529b\u7684\u201c\u901a\u624d\u6a21\u578b\u201d&#xff08;Generalist Policies&#xff09;\u6f14\u8fdb\u3002\u8fd9\u4e00\u8f6c\u53d8\u7684\u9a71\u52a8\u529b\u4e3b\u8981\u6765\u81ea\u4e8e\u5927\u8bed\u8a00\u6a21\u578b&#xff08;LLM<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[2681,188,523,51],"topic":[],"class_list":["post-92723","post","type-post","status-publish","format-standard","hentry","category-server","tag-vllm","tag-188","tag-523","tag-51"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \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\/92723.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"1. \u5f15\u8a00&#xff1a;\u5177\u8eab\u667a\u80fd\u7684\u201c\u57fa\u7840\u6a21\u578b\u201d\u65f6\u4ee3 1.1 \u4ece\u5355\u4e00\u4efb\u52a1\u5230\u901a\u7528\u7b56\u7565\u7684\u8303\u5f0f\u8f6c\u79fb \u57282024\u5e74\u81f32025\u5e74\u521d\u7684\u8fd9\u6bb5\u65f6\u95f4\u91cc&#xff0c;\u673a\u5668\u4eba\u5b66\u4e60&#xff08;Robot Learning&#xff09;\u9886\u57df\u7ecf\u5386\u4e86\u4e00\u573a\u6df1\u523b\u7684\u8303\u5f0f\u8f6c\u79fb&#xff0c;\u5176\u6838\u5fc3\u5728\u4e8e\u4ece\u9488\u5bf9\u7279\u5b9a\u73af\u5883\u3001\u7279\u5b9a\u4efb\u52a1\u7684\u201c\u4e13\u5bb6\u6a21\u578b\u201d\u5411\u5177\u5907\u5e7f\u6cdb\u6cdb\u5316\u80fd\u529b\u7684\u201c\u901a\u624d\u6a21\u578b\u201d&#xff08;Generalist Policies&#xff09;\u6f14\u8fdb\u3002\u8fd9\u4e00\u8f6c\u53d8\u7684\u9a71\u52a8\u529b\u4e3b\u8981\u6765\u81ea\u4e8e\u5927\u8bed\u8a00\u6a21\u578b&#xff08;LLM\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/92723.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-08-10T16:16:44+00:00\" \/>\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=\"7 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/92723.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/92723.html\",\"name\":\"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-08-10T16:16:44+00:00\",\"dateModified\":\"2026-08-10T16:16:44+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/92723.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/92723.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/92723.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a\"}]},{\"@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":"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \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\/92723.html","og_locale":"zh_CN","og_type":"article","og_title":"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"1. \u5f15\u8a00&#xff1a;\u5177\u8eab\u667a\u80fd\u7684\u201c\u57fa\u7840\u6a21\u578b\u201d\u65f6\u4ee3 1.1 \u4ece\u5355\u4e00\u4efb\u52a1\u5230\u901a\u7528\u7b56\u7565\u7684\u8303\u5f0f\u8f6c\u79fb \u57282024\u5e74\u81f32025\u5e74\u521d\u7684\u8fd9\u6bb5\u65f6\u95f4\u91cc&#xff0c;\u673a\u5668\u4eba\u5b66\u4e60&#xff08;Robot Learning&#xff09;\u9886\u57df\u7ecf\u5386\u4e86\u4e00\u573a\u6df1\u523b\u7684\u8303\u5f0f\u8f6c\u79fb&#xff0c;\u5176\u6838\u5fc3\u5728\u4e8e\u4ece\u9488\u5bf9\u7279\u5b9a\u73af\u5883\u3001\u7279\u5b9a\u4efb\u52a1\u7684\u201c\u4e13\u5bb6\u6a21\u578b\u201d\u5411\u5177\u5907\u5e7f\u6cdb\u6cdb\u5316\u80fd\u529b\u7684\u201c\u901a\u624d\u6a21\u578b\u201d&#xff08;Generalist Policies&#xff09;\u6f14\u8fdb\u3002\u8fd9\u4e00\u8f6c\u53d8\u7684\u9a71\u52a8\u529b\u4e3b\u8981\u6765\u81ea\u4e8e\u5927\u8bed\u8a00\u6a21\u578b&#xff08;LLM","og_url":"https:\/\/www.wsisp.com\/helps\/92723.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-08-10T16:16:44+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"7 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/92723.html","url":"https:\/\/www.wsisp.com\/helps\/92723.html","name":"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-08-10T16:16:44+00:00","dateModified":"2026-08-10T16:16:44+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/92723.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/92723.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/92723.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"2024-2025\u5e74\u89c6\u89c9-\u8bed\u8a00-\u52a8\u4f5c\uff08VLA\uff09\u6a21\u578b\u524d\u6cbf\u7814\u7a76\u62a5\u544a"}]},{"@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\/92723","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=92723"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/92723\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=92723"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=92723"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=92723"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=92723"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}