{"id":3095,"date":"2025-05-14T16:22:42","date_gmt":"2025-05-14T08:22:42","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=3095"},"modified":"2025-05-14T16:22:42","modified_gmt":"2025-05-14T08:22:42","slug":"202505-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-drivegpt4interpretable-end-to-end-autonomous-driving-via-large-language-model","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202505-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-drivegpt4interpretable-end-to-end-autonomous-driving-via-large-language-model\/","title":{"rendered":"202505 \u8bba\u6587\u7814\u8bfb-DriveGPT4:Interpretable End-to-End Autonomous Driving Via Large Language Model"},"content":{"rendered":"\n

\u4f5c\u8005\uff1a<\/a>Zhenhua Xu<\/a><\/a><\/a>;\u00a0Yujia Zhang<\/a>
\u5355\u4f4d\uff1a\u9999\u6e2f\u5927\u5b66\u3001\u6d59\u6c5f\u5927\u5b66
\u6765\u6e90\uff1a\u00a0<\/strong>IEEE Robotics and Automation Letters<\/a>
\u65f6\u95f4\uff1a2024.08.07<\/p>\n\n\n\n

\u4e00\u3001\u7814\u7a76\u80cc\u666f<\/strong><\/h2>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u81ea\u52a8\u9a7e\u9a76\u7cfb\u7edf\u4e3b\u8981\u5206\u4e3a\u6a21\u5757\u5316\u6d41\u6c34\u7ebf<\/strong>\uff08\u611f\u77e5\u2192\u89c4\u5212\u2192\u63a7\u5236\uff09\u548c\u7aef\u5230\u7aef\u5b66\u4e60<\/strong>\u4e24\u7c7b\u3002\u524d\u8005\u53ef\u89e3\u91ca\u6027\u8f83\u597d\u4f46\u96be\u4ee5\u6574\u4f53\u4f18\u5316\uff0c\u540e\u8005\u867d\u7136\u7701\u53bb\u4e2d\u95f4\u6b65\u9aa4\u5374\u5f80\u5f80\u662f\u9ed1\u76d2\u51b3\u7b56\uff0c\u5e26\u6765\u5b89\u5168\u3001\u6cd5\u5f8b\u548c\u4fe1\u4efb\u65b9\u9762\u7684\u6311\u6218\u3002\u4e3a\u4e86\u89e3\u51b3\u7aef\u5230\u7aef\u65b9\u6cd5\u7684\u53ef\u89e3\u91ca\u6027\u95ee\u9898\uff0c\u4e00\u4e9b\u5de5\u4f5c\u4f7f\u7528\u5c0f\u578b\u8bed\u8a00\u6a21\u578b\u751f\u6210\u6709\u9650\u7684\u6587\u5b57\u8bf4\u660e\uff0c\u4f46\u56e0\u6a21\u578b\u5bb9\u91cf\u53d7\u9650\uff0c\u5176\u7b54\u590d\u50f5\u5316\u4e14\u53ea\u80fd\u5904\u7406\u9884\u8bbe\u95ee\u9898\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08\u5982 ChatGPT\u3001LLaMA\uff09\u5728\u89c6\u89c9\u7406\u89e3\u4e0e\u63a8\u7406\u65b9\u9762\u5c55\u73b0\u51fa\u5f3a\u5927\u80fd\u529b\uff0c\u4e3a\u5c06\u53ef\u89e3\u91ca\u6027\u5f15\u5165\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\u63d0\u4f9b\u4e86\u65b0\u5951\u673a\u3002<\/p>\n\n\n\n

\u4e8c\u3001\u6838\u5fc3\u5185\u5bb9<\/strong><\/h2>\n\n\n\n

1.\u4efb\u52a1\u5b9a\u4e49<\/strong><\/p>\n\n\n\n

    \n
  • \u8f93\u5165\uff1a\u524d\u89c6\u5355\u76ee\u89c6\u9891\u5e8f\u5217 + \u7528\u6237\u81ea\u7136\u8bed\u8a00\u67e5\u8be2\uff08\u5982\u201c\u6b64\u523b\u8f66\u8f86\u5728\u505a\u4ec0\u4e48\uff1f\u201d\u3001\u201c\u9884\u6d4b\u4e0b\u5e27\u901f\u5ea6\u548c\u8f6c\u5411\u89d2\u201d\u6216\u66f4\u4e30\u5bcc\u7684\u591a\u8f6e\u5bf9\u8bdd\uff09<\/li>\n<\/ul>\n\n\n\n
      \n
    • \u8f93\u51fa\uff1a\u6587\u672c\u5316\u7684\u52a8\u4f5c\u63cf\u8ff0<\/strong>\u4e0e\u884c\u4e3a\u7406\u7531<\/strong>\uff0c\u4ee5\u53ca\u4f4e\u7ea7\u63a7\u5236\u4fe1\u53f7<\/strong>\uff08\u901f\u5ea6 v\uff0c\u4e0b\u4e00\u4e2a\u8f6c\u5411\u89d2 \u0394\uff09 \u3002<\/li>\n<\/ul>\n\n\n\n

      2.\u6307\u4ee4\u8c03\u4f18\u6570\u636e\u96c6\u6784\u5efa<\/strong><\/p>\n\n\n\n

        \n
      • BDD-X QA<\/strong>\uff1a\u4ece BDD-X \u63d0\u4f9b\u7684\u52a8\u4f5c\u63cf\u8ff0\u3001\u884c\u4e3a\u7406\u7531\u3001\u63a7\u5236\u4fe1\u53f7\u6807\u7b7e\uff0c\u5206\u522b\u6784\u9020\u4e09\u7c7b\u540c\u4e49\u95ee\u53e5\u96c6\uff0c\u5171 16 K \u6761 \u3002<\/li>\n<\/ul>\n\n\n\n
          \n
        • ChatGPT \u751f\u6210 QA<\/strong>\uff1a\u5229\u7528 YOLOv8 \u63a2\u6d4b\u5e27\u5185\u76ee\u6807\uff0c\u5e76\u5c06\u68c0\u6d4b\u6846\u3001\u771f\u503c\u63a7\u5236\u3001\u539f\u59cb BDD-X \u6587\u672c\u4f5c\u4e3a\u201cprivileged info\u201d\u9001\u5165 ChatGPT\uff0c\u751f\u6210 40 K \u6761\u591a\u8f6e\u3001\u591a\u6837\u5316\u7684\u9a7e\u9a76\u573a\u666f\u5bf9\u8bdd\u5f0f\u6307\u4ee4 \u3002<\/li>\n<\/ul>\n\n\n\n
            \n
          • \u6700\u7ec8\u5f97\u5230 56 K<\/strong> \u6761\u81ea\u52a8\u9a7e\u9a76\u4e13\u5c5e\u6307\u4ee4\uff0c\u518d\u4e0e 223 K<\/strong> \u6761\u901a\u7528\u89c6\u89c9\u6307\u4ee4\u6df7\u5408\u7528\u4e8e\u5fae\u8c03 \u3002<\/li>\n<\/ul>\n\n\n\n

            \u4e09\u3001\u6838\u5fc3\u6846\u67b6\u8d21\u732e<\/strong><\/h2>\n\n\n\n
            \"\"<\/figure>\n\n\n\n

            1.\u591a\u6a21\u6001\u7edf\u4e00 Token \u5316<\/strong><\/p>\n\n\n\n

              \n
            • \u89c6\u9891 Tokenizer<\/strong>\uff1a\u7528\u9884\u8bad\u7ec3 CLIP \u63d0\u53d6\u6bcf\u5e27\u5168\u5c40\u4e0e patch \u7279\u5f81\uff0c\u5e76\u901a\u8fc7\u8f7b\u91cf\u6295\u5f71\u5668\u6620\u5c04\u5230\u4e0e\u6587\u672c Token \u540c\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n\n\n\n
                \n
              • \u7edf\u4e00 De-tokenizer<\/strong>\uff1a\u5c06\u63a7\u5236\u4fe1\u53f7\u4e5f\u5f53\u4f5c\u201c\u6587\u672c\u201d\u5904\u7406\uff0c\u501f\u9274 RT-2 \u7684\u601d\u8def\uff0c\u5c06 v\u3001\u0394 \u5e8f\u5217\u5316\u4e3a Token\uff0c\u7531 LLaMA2 \u76f4\u63a5\u9884\u6d4b\uff0c\u518d\u89e3\u7801\u56de\u6570\u503c \u3002<\/li>\n<\/ul>\n\n\n\n

                2.Mix-Finetune \u8bad\u7ec3\u7b56\u7565<\/strong><\/p>\n\n\n\n

                  \n
                • \u9884\u8bad\u7ec3\u9636\u6bb5<\/strong>\uff1a\u4ec5\u8bad\u7ec3\u6295\u5f71\u5668\uff0c\u5728 CC3M\uff08593 K \u56fe\u6587\u5bf9\uff09\u548c WebVid-2M\uff08703 K \u89c6\u9891\u6587\u5bf9\uff09\u4e0a\u5bf9\u9f50\u89c6\u89c9\u4e0e\u6587\u672c\uff0c\u4fdd\u6301 CLIP \u548c LLM \u53c2\u6570\u56fa\u5b9a\u3002<\/li>\n<\/ul>\n\n\n\n
                    \n
                  • \u6df7\u5408\u5fae\u8c03\u9636\u6bb5<\/strong>\uff1a\u5148\u7528 223 K \u901a\u7528\u89c6\u89c9\u6307\u4ee4\u6570\u636e\u5fae\u8c03\uff0c\u518d\u4ee5 56 K \u81ea\u52a8\u9a7e\u9a76\u6307\u4ee4\u6570\u636e\u8fdb\u4e00\u6b65\u7cbe\u8c03\uff0c\u4f7f\u6a21\u578b\u65e2\u5177\u5907\u901a\u7528\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff0c\u53c8\u80fd\u7cbe\u51c6\u6267\u884c\u53ef\u89e3\u91ca\u9a7e\u9a76\u4efb\u52a1 \u3002<\/li>\n<\/ul>\n\n\n\n

                    3.\u53ef\u89e3\u91ca\u4e0e\u53ef\u63a7\u53cc\u91cd\u8f93\u51fa<\/strong><\/p>\n\n\n\n

                      \n
                    • DriveGPT4 \u4e0d\u4ec5\u80fd\u5728\u81ea\u7136\u8bed\u8a00\u5c42\u9762\u7ed9\u51fa\u52a8\u4f5c\u4e0e\u7406\u7531\uff0c\u8fd8\u80fd\u540c\u6b65\u8f93\u51fa\u53ef\u76f4\u63a5\u6267\u884c\u7684\u4f4e\u7ea7\u63a7\u5236\u547d\u4ee4\uff0c\u5b9e\u73b0\u201c\u770b\u5f97\u89c1\u3001\u95ee\u5f97\u660e\u3001\u63a7\u5f97\u51c6\u201d\u7684\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76\u3002<\/li>\n<\/ul>\n\n\n\n

                      \u56db\u3001\u5b9e\u9a8c\u90e8\u5206<\/strong><\/h2>\n\n\n\n
                      \"\"<\/figure>\n\n\n\n

                      \u89e3\u91ca\u6027\u8bc4\u4f30\uff08BDD-X QA\uff09<\/strong><\/p>\n\n\n\n

                        \n
                      • \u5728\u52a8\u4f5c\u63cf\u8ff0\u3001\u884c\u4e3a\u7406\u7531\u4e09\u4e2a\u56fa\u5b9a\u683c\u5f0f\u4efb\u52a1\u4e0a\uff0cDriveGPT4 \u5728 CIDEr\u3001BLEU-4\u3001ROUGE-L \u7b49\u6307\u6807\u4e0a\u5747\u8d85\u8d8a ADAPT \u7b49 SOTA\uff0c\u5c24\u5176\u5728\u201cHard\u201d\u573a\u666f\u63d0\u5347\u663e\u8457\u3002<\/li>\n\n\n\n
                      • \u5bf9\u968f\u673a\u62bd\u53d6\u7684 100 \u6761\u591a\u6837\u5316\u95ee\u7b54\uff08Additional QA\uff09\u6d4b\u8bd5\uff0cDriveGPT4 \u53ef\u4ee5\u56de\u7b54\u5404\u7c7b\u8f66\u8f86\u72b6\u6001\u3001\u5468\u8fb9\u73af\u5883\u63d0\u95ee\uff0c\u800c\u5176\u4ed6\u65b9\u6cd5\u591a\u534a\u4e0d\u80fd \u3002<\/li>\n<\/ul>\n\n\n\n

                        \u7aef\u5230\u7aef\u63a7\u5236\u4fe1\u53f7\u9884\u6d4b<\/strong><\/p>\n\n\n\n

                          \n
                        • \u5728 BDD-X \u5168\u91cf\u6d4b\u8bd5\u96c6\u4e0a\uff0c\u4ee5 RMSE<\/strong> \u548c \u591a\u9608\u503c\u51c6\u786e\u7387 A\u03c4<\/strong>\uff08\u03c4\u2208{0.1,0.5,1.0,5.0}\uff09\u8bc4\u4f30\uff0cDriveGPT4 \u5728\u901f\u5ea6\u548c\u8f6c\u5411\u89d2\u9884\u6d4b\u4e0a\u5747\u4f18\u4e8e ADAPT \u7b49\u57fa\u7ebf \u3002<\/li>\n<\/ul>\n\n\n\n

                          \u96f6\u6837\u672c\u6cdb\u5316\u80fd\u529b<\/strong><\/p>\n\n\n\n

                            \n
                          • \u5c06\u6a21\u578b\u76f4\u63a5\u5e94\u7528\u4e8e nuScenes \u6570\u636e\u96c6\u548c\u7535\u5b50\u6e38\u620f\u753b\u9762\uff0cDriveGPT4 \u5728\u95ee\u7b54\u548c\u63a7\u5236\u9884\u6d4b\u4e0a\u5747\u7ed9\u51fa\u5408\u7406\u7ed3\u679c\uff0c\u5c55\u793a\u51fa\u8de8\u57df\u9002\u5e94\u6027 \u3002<\/li>\n<\/ul>\n\n\n\n

                            \u6d88\u878d\u7814\u7a76<\/strong><\/p>\n\n\n\n

                              \n
                            • \u53bb\u9664 BDD-X QA\u3001ChatGPT QA\u3001\u6216\u901a\u7528\u89c6\u89c9\u6307\u4ee4\u4e2d\u7684\u4efb\u610f\u4e00\u90e8\u5206\uff0c\u5747\u4f1a\u5bfc\u81f4\u5404\u9879\u4efb\u52a1\u6027\u80fd\u660e\u663e\u4e0b\u964d\uff0c\u9a8c\u8bc1\u4e86\u5404\u7c7b\u6307\u4ee4\u6570\u636e\u4e0e\u6df7\u5408\u5fae\u8c03\u7b56\u7565\u7684\u5fc5\u8981\u6027 \u3002<\/li>\n<\/ul>\n\n\n\n

                              \u4e94\u3001\u603b\u7ed3<\/strong><\/h2>\n\n\n\n

                              DriveGPT4 \u662f\u9996\u4e2a\u5c06\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u5b8c\u6574\u5e94\u7528\u4e8e\u53ef\u89e3\u91ca\u7aef\u5230\u7aef\u81ea\u52a8\u9a7e\u9a76<\/strong>\u7684\u7cfb\u7edf\uff0c\u901a\u8fc7\u89c6\u89c9\u2013\u6587\u672c\u2013\u63a7\u5236\u7edf\u4e00 Token \u5316\u548c\u5927\u89c4\u6a21\u6df7\u5408\u6307\u4ee4\u5fae\u8c03\uff0c\u5b9e\u73b0\u4e86\u540c\u65f6\u751f\u6210\u81ea\u7136\u8bed\u8a00\u89e3\u91ca\u4e0e\u4f4e\u7ea7\u63a7\u5236\u547d\u4ee4\uff0c\u5e76\u5728\u591a\u9879\u57fa\u51c6\u4efb\u52a1\u4e0a\u5237\u65b0\u4e86 SOTA\u3002\u672a\u6765\u5de5\u4f5c\u5c06\u9762\u5411\u95ed\u73af\u63a7\u5236<\/strong>\u4e0e\u5728\u7ebf\u6570\u636e\u6536\u96c6<\/strong>\u8bbe\u8ba1\uff0c\u8fdb\u4e00\u6b65\u89e3\u51b3\u6a21\u4eff\u5b66\u4e60\u6f02\u79fb\u95ee\u9898\u5e76\u63d0\u5347\u7cfb\u7edf\u7684\u5b89\u5168\u6027\u548c\u53ef\u9760\u6027\u3002<\/p>\n\n\n\n

                              <\/p>\n","protected":false},"excerpt":{"rendered":"

                              \u4f5c\u8005\uff1aZhenhua Xu;\u00a0Yujia Zhang\u5355\u4f4d\uff1a\u9999\u6e2f\u5927\u5b66\u3001\u6d59\u6c5f\u5927\u5b66\u6765\u6e90\uff1a\u00a0IEEE Robotics […]<\/p>\n","protected":false},"author":63,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-3095","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3095","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/63"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=3095"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3095\/revisions"}],"predecessor-version":[{"id":3098,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/3095\/revisions\/3098"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=3095"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=3095"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=3095"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}