[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fxvKs8e7Sr79RjrG3BNGenhbdpOqzgq56TwWsrRS5cTg":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1211,"别盯数据了！宇树王兴兴：模型才是机器人真瓶颈","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">2025年世界机器人大会上，宇树科技创始人王兴兴抛出一个反共识观点：限制机器人产业爆发的核心不是数据不足，而是模型架构的落后——这犹如一盆冷水，浇在疯狂收集数据的机器人厂商头上。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">北京亦庄，2025世界机器人大会主论坛现场，宇树科技CEO王兴兴站在演讲台前，面对全球机器人产业精英，抛出了一个观点：“目前全球范围内，大家对机器人数据这个问题关注度有点太高了。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在他看来：“现在最大的问题其实是模型问题，而不是数据问题。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">会场内一阵低声议论。在OpenAI成功范式影响下，全球机器人产业正疯狂收集数据，各地机器人数采中心如雨后春笋般涌现。王兴兴的论断犹如一盆冷水，泼在了整个行业的发展思路上。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">当硬件够用，Ai掉队\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">2025年，机器人产业正迎来前所未有的繁荣景象。先是智元机器人以大手笔入主上纬新材，引发“借壳上市”的猜测，股价连续涨停10次，每次涨幅达20%，刷新了A股市场的纪录；随后，宇树科技年内四次传出IPO消息，直至中国证监会公布其上市辅导备案报告，IPO的悬念终于尘埃落定。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在近日开幕的世界机器人大会展厅内，灵活舞动的人形机器人随处可见，展台前观众络绎不绝。超过150款人形机器人同台亮相，创下了国内有史以来最大规模的人形机器人集体展示纪录。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F1319fe96306e4763a7fd83f0bfe4322d\u002FAA1KgOze.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">王兴兴在演讲中透露\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">，今年上半年，机器人行业增速惊人，整机及零部件厂商的平均增长率达到了50%-100%\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">，这一增速在行业历史上极为罕见。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在这场行业狂奔的背后，是资本热潮的涌动。据盖世汽车不完全统计，截至8月7日，今年该领域已发生超过百起融资事件，累计融资金额接近300亿元（未披露的融资额未纳入统计）。相比之下，去年全年共发生72起融资事件，累计融资金额约为130.23亿元。今年以来的投资数量和金额已远超去年全年数据。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">然而繁荣背后暗藏隐忧。王兴兴指出一个关键矛盾：“机器人硬件性能虽然还不够好，但目前是够用的。目前最大的挑战还是具身智能的AI完全不够用。这也是限制人形机器人大规模应用的最大点。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这一判断得到产业数据的佐证。今年以来，从机器人首登央视春晚跳舞，到全球首个机器人半程马拉松“洋相百出”、即将开幕的全球首届人形机器人运动会，再到本届世界机器人大会和全国大学生机器人大赛等密集活动，人形机器人在表演、格斗等娱乐场景大放异彩，但在真正需要“干活”的工业场景，应用深度仍然有限。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F8dddeb4d2fe349008eaa9017a8f6a521\u002FAA1Kgqt0.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“机器人的应用场景究竟为何？没有找准这一定位，就很难针对性的做技术突破。”在与某位业内技术大拿沟通时，他如是向盖世汽车坦言，是单纯陪伴，还是家务助手，抑或需要更精密操作，这些都需要不同的技术实现。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">而在王兴兴看来，很多技术的进步是需要时间的，当下马上让一个机器人去家里干点有实际价值的活还不太现实，如果只是做个Demo（演示）或者示例是没问题的。“我们去年就跟汽车工厂合作，在工厂落地部署机器人，但是真正让机器人产生比较大的价值，当下是不太现实的。”正如早期电脑诞生时，也同样不具备普适性、实用性。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“当前这个时点有点像ChatGPT出来之前的1到3年，”王兴兴如此比喻机器人大模型的发展阶段，“整个业界已经发现了类似的方向以及技术路线，但是没人把它做出来。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">具身智能大模型的滞后已成为制约机器人真正“干活”的核心挑战。当硬件准备就绪，AI大脑却跟不上，这场产业革命的步伐因此被拖慢。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">VLA遇冷，世界模型崛起\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">面对具身智能的模型瓶颈，全球科研团队正探索不同技术路径。其中VLA+RL（视觉-语言-动作模型+强化学习）路线获得了众多顶尖机构的青睐，被视为通向通用机器人智能的可行之路。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">那么什么是VLA呢？简单来说，这类模型旨在将视觉感知、语言理解和物理动作融为一体，让机器人能够听懂人的指令（“把桌子上的苹果拿给我”），看懂当前的环境（识别出哪个是苹果、哪个是桌子），并自主生成一系列动作来完成任务。可以说，VLA正是未来通用机器人的“大脑”。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">2023年7月，谷歌DeepMind就曾基于VLA架构推出RT-2模型，通过整合大语言模型与多模态数据训练，赋予机器人执行复杂任务的能力。其任务准确率较初代模型提升近一倍（从32%至62%），突破性地实现了垃圾分类等场景的零样本学习。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">随后，VLA的理念很快被汽车公司关注，快速应用于汽车智能驾驶领域，如果说2024年“端到端”是智能驾驶领域最火的词汇，那么2025年非“VLA“莫属。小鹏汽车、理想汽车等公司都发布了各自的VLA方案。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">但相较于智能驾驶汽车动辄百亿参数、近千TOPS算力的海量数据，仍处于量产初期的机器人训练数据集的参数量也大多只有100万至300万之间。更遑论机器人应用场景的多模态感知更丰富、执行动作更复杂、传感器数据更微观。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F24fee231d87e4d17a56d99d3bbcaa1de\u002FAA1KgOzj.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“我个人感觉，包括我们公司目前尝试下来，VLA+RL还是不够的。”王兴兴在大会上直言：\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">他点出关键问题：VLA模型在对真实世界交互时数据质量不足，即使在强化学习的加持下，该模型架构仍需继续升级优化。这一判断揭示了当前机器人学习效率低下的核心原因——用错误的方法处理再多数据也难有突破。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在VLA路线遭遇挑战的同时，另一种技术路径正在崭露头角：“世界模型”作为通向通用人工智能（AGI）的重要阶梯，正吸引越来越多研究者的目光。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">这种模型通过学习环境的时空动态，不仅能预测未来状态，还能评估自身行动的后果。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">谷歌DeepMind于8月6日发布的第三代通用世界模型Genie 3成为大会热议话题。该系统能为机器人提供低成本虚拟训练环境，支持复杂任务的长时程模拟。王兴兴特别指出：“世界模型可能会比VLA模型更快落地，这一路线值得关注。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">世界模型的优势在于大幅降低训练成本。传统机器人训练需要大量物理试错，既耗时又耗资源；而高质量的世界模型能构建逼真的虚拟环境，使机器人在进入真实场景前已积累丰富“经验”。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">另一个关键挑战是知识复用问题。王兴兴指出：“机器人学习新技能需从头训练，无法复用旧知识，亟需实现类似大模型的持续学习能力。” 这导致当前机器人技能学习效率低下，与人类举一反三的学习能力形成鲜明对比。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">具身智能领域的“小样本高泛化”新范式成为突破方向。业内专家正研发通过少量数据训练出高算法能力的模型，而非依赖纯数据驱动，这将大幅提升机器人的适应能力和学习效率。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">何时跨越“可用”到“好用”\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">面对未来2-5年的关键发展期，王兴兴清晰勾勒出智能机器人技术的三大重心：统一端到端智能机器人大模型、低成本高寿命硬件及超大批量制造、低成本大规模算力。这三大支柱将共同支撑机器人产业实现从“可用”到“好用”的关键跨越。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在硬件领域，降本增效已取得显著进展。宇树科技通过材料创新与模块化设计，已将人形机器人核心部件成本降低40%；同时，通过仿生关节设计，将关键部件寿命提升至3万小时以上。这些突破为人形机器人规模化商用奠定基础。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">分布式算力网络将成为破解机器人算力瓶颈的关键。王兴兴指出，人形机器人本体算力有限，峰值功耗约100瓦，仅相当于几部手机的算力。“通过联邦学习实现多机器人数据共享，我们预计2027年将单台机器人训练成本降低至万元级。” 这一预测意味着机器人训练成本将迎来断崖式下降，为大规模部署扫清障碍。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F31f8cbb3de4f454ba9e3aac8dc09e84a\u002FAA1KgOzj.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">王兴兴预测：“未来几年，全行业人形机器人出货量每年翻番都是有保证的。如果有更大的技术突破，甚至可能未来2-3年突然一年出货几十万台，甚至上百万台也有可能。” 这一爆发式增长将首先在工业场景显现，而汽车制造业有望成为最大受益者。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">行业生态建设也在加速。杭州具身智能应用中试基地近期启动，整合宇树科技“最优本体”与华为、阿里云等“最强大脑”，打造 “算力+数据+模型+场景应用”生态，加速具身智能在工业领域的规模化落地。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">机器人产业的“ChatGPT时刻”何时到来？王兴兴给出的预测是：最快未来1-2年，最慢3-5年。那个临界点的标志简单而直观——当一个人形机器人被带到陌生会场，能听懂“把这瓶水带给某位观众”的随机指令，并流畅完成任务，便宣告了机器人智能时代的真正开启。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】盖世汽车 \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002F%E5%88%AB%E7%9B%AF%E6%95%B0%E6%8D%AE%E4%BA%86-%E5%AE%87%E6%A0%91%E7%8E%8B%E5%85%B4%E5%85%B4-%E6%A8%A1%E5%9E%8B%E6%89%8D%E6%98%AF%E6%9C%BA%E5%99%A8%E4%BA%BA%E7%9C%9F%E7%93%B6%E9%A2%88\u002Far-AA1KgR1j?ocid=msedgntphdr&cvid=67764a46dd014287b627707a75cbb518&ei=25\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FSFMrN\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F3a6e69d4874844c0bd052417279c9b0b\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002Fthumbs\u002F3a6e69d4874844c0bd052417279c9b0b\u002FAI领域.jpg",0,1,221,"2025-08-12 18:26",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3Ab4edd37e-e0a3-48cb-9464-31e1967ddf85%3A0.wav?Expires=1754998800&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=ciEXIyJ1%2B4cjEXm%2Blb84YT3dChA%3D",16960778,"b4edd37e-e0a3-48cb-9464-31e1967ddf85","2025-08-12 18:22","Stop staring at the data! Wang Xingxing, founder of Unitech, says: the model is the real bottleneck for robots.","\u003Cp>\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">At the 2025 World Robot Conference, Wang Xingxing, founder of Unitech, put forward a controversial view: the core issue limiting the explosive growth of the robot industry is not insufficient data, but outdated model architecture — this is like a bucket of cold water poured over the robot manufacturers who are obsessed with collecting data.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In Yizhuang, Beijing, during the main forum of the 2025 World Robot Conference, Wang Xingxing, CEO of Unitech, stood on the podium facing global leaders in the robot industry and put forward an idea: \"At present, people around the world pay too much attention to the issue of robot data.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In his view: \"The biggest problem right now is the model problem, not the data problem.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">There was a low murmur in the hall. Under the influence of OpenAI's successful model, the global robot industry is collecting data obsessively, and robot data collection centers have sprung up everywhere. Wang Xingxing's statement is like a bucket of cold water, pouring over the entire industry's development thinking.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">When hardware is sufficient, AI falls behind\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In 2025, the robot industry is experiencing unprecedented prosperity. First, Zhiyuan Robotics made a big move to take over Shangwei New Materials, triggering speculation about \"acquisition listing,\" and the stock price rose by 20% for ten consecutive days, breaking the record in the A-share market; then, Unitech announced four IPO messages within the year, until the China Securities Regulatory Commission announced its listing guidance filing report, and the IPO suspense finally came to a close.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the exhibition hall of the recently opened World Robot Conference, humanoid robots dancing flexibly can be seen everywhere, with audiences constantly coming and going at the booths. More than 150 humanoid robots were displayed together, setting a record for the largest scale of humanoid robot display in Chinese history.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F1319fe96306e4763a7fd83f0bfe4322d\u002FAA1KgOze.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Wang Xingxing revealed in his speech\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">that this year, the robotics industry has grown rapidly, with an average growth rate of 50%-100% for complete machines and parts manufacturers\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">, which is extremely rare in the history of the industry.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Behind this industry sprint is the surge of capital enthusiasm. According to Guashigas, as of August 7th, there have been more than 100 financing events in this field this year, with cumulative financing amount approaching 30 billion yuan (unrevealed financing amounts are not included in the statistics). Compared with last year, which had 72 financing events and a total financing amount of approximately 13.023 billion yuan, the number and amount of investments this year have far exceeded last year's figures.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">However, under the prosperity lies hidden concerns. Wang Xingxing pointed out a key contradiction: \"Although the performance of robot hardware is still not good enough, it is currently sufficient. The biggest challenge is that embodied AI is completely insufficient. This is also the biggest point limiting the large-scale application of humanoid robots.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This judgment is supported by industry data. Since the beginning of this year, from robots first appearing on CCTV's Spring Festival Gala to dance, to the global first robot half marathon \"making a mess\", to the upcoming global first humanoid robot sports meeting, and to this World Robot Conference and the National College Students Robot Competition, etc., humanoid robots have shone brightly in entertainment scenarios such as performance and combat, but their application depth in industrial scenarios that really require \"working\" is still limited.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F8dddeb4d2fe349008eaa9017a8f6a521\u002FAA1Kgqt0.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“What is the application scenario of robots? Without finding this positioning, it is difficult to make targeted technological breakthroughs.” When communicating with a technical expert in the industry, he frankly told Guashigas, whether it is simple companionship, household assistant, or requires more precise operations, these all need different technical implementations.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In Wang Xingxing's view, many technological advances take time. It is not realistic to have a robot do some practical value work at home right now. If it is just a Demo (demonstration) or example, it is okay. “We cooperated with an automobile factory last year and deployed robots at the factory, but it is not realistic to have robots generate significant value at this time.” Just like when early computers were born, they also lacked universality and practicality.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“This moment is a bit like the 1 to 3 years before ChatGPT came out,” Wang Xingxing used this analogy to describe the development stage of robot large models, “the entire industry has already found similar directions and technology routes, but no one has made it.”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The lag of embodied intelligence large models has become the core challenge restricting robots from truly \"working.\" When the hardware is ready, the AI brain is not catching up, so the pace of this industrial revolution is slowed down.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">VLA cools down, world models rise\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Faced with the model bottleneck of embodied intelligence, global research teams are exploring different technical paths. Among them, the VLA+RL (Vision-Language-Action model + Reinforcement Learning) path has been favored by many top institutions, regarded as a feasible way to achieve general robot intelligence.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">So what is VLA? Simply put, this type of model aims to integrate visual perception, language understanding, and physical actions, allowing robots to understand human instructions (\"Bring me the apple on the table\"), understand the current environment (identify which is the apple and which is the table), and autonomously generate a series of actions to complete the task. In short, VLA is the \"brain\" of future general robots.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In July 2023, Google DeepMind launched the RT-2 model based on the VLA architecture, which through integrating large language models and multimodal data training, endowed robots with the ability to perform complex tasks. Its task accuracy increased nearly twice (from 32% to 62%), achieving zero-shot learning in scenarios such as garbage classification for the first time.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Subsequently, the concept of VLA was quickly noticed by car companies and rapidly applied to the field of intelligent driving. If \"end-to-end\" was the hottest word in the intelligent driving field in 2024, then in 2025 it would be \"VLA\". Companies such as XPeng and Li Auto have released their own VLA solutions.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">But compared to the hundreds of billions of parameters and thousands of TOPS computing power of smart driving cars, the parameter count of robot training datasets, which are still in the initial production phase, is mostly only between 1 million and 3 million. Let alone the fact that the robot application scenarios have richer multimodal perception, more complex execution actions, and more microscopic sensor data.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F24fee231d87e4d17a56d99d3bbcaa1de\u002FAA1KgOzj.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">“Personally, I feel that including our company’s current attempts, VLA+RL is still not enough,” Wang Xingxing said directly at the conference:\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">He pointed out the key issue: VLA models have insufficient data quality when interacting with the real world, and even with the support of reinforcement learning, the model architecture still needs to continue upgrading and optimizing. This judgment reveals the core reason for the low efficiency of current robot learning - using the wrong method to process more data will not lead to breakthroughs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">At the same time that the VLA route encountered challenges, another technical path is emerging: \"World Model\" as an important step towards general artificial intelligence (AGI) is attracting more and more researchers' attention.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">This model learns the spatiotemporal dynamics of the environment, not only predicting future states but also evaluating the consequences of its own actions.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Google DeepMind's third-generation general world model Genie 3, released on August 6th, became a hot topic at the conference. This system provides a low-cost virtual training environment for robots, supporting long-term simulation of complex tasks. Wang Xingxing specifically pointed out: \"The world model may be implemented faster than the VLA model, and this route is worth paying attention to.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The advantage of the world model lies in significantly reducing training costs. Traditional robot training requires a lot of physical trial and error, which is time-consuming and resource-intensive; while high-quality world models can build realistic virtual environments, allowing robots to accumulate rich \"experience\" before entering real scenes.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Another key challenge is the knowledge reuse issue. Wang Xingxing pointed out: \"Robots need to train from scratch to learn new skills and cannot reuse old knowledge, and it is urgently needed to achieve continuous learning capabilities similar to large models.\" This leads to the current low efficiency of robot skill learning, forming a sharp contrast with human learning abilities that can apply knowledge to similar situations.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The new paradigm of \"few samples and high generalization\" in embodied intelligence has become a breakthrough direction. Industry experts are developing models that can train with a small amount of data and achieve high algorithmic capabilities, rather than relying solely on data-driven approaches, which will greatly improve the adaptability and learning efficiency of robots.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">When will the transition from \"usable\" to \"useful\" happen?\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Facing the critical development period of 2-5 years, Wang Xingxing clearly outlined three key focuses of intelligent robot technology: unified end-to-end intelligent robot large model, low-cost and long-life hardware and large-scale manufacturing, and low-cost large-scale computing power. These three pillars will jointly support the robot industry's key transition from \"usable\" to \"useful.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the hardware field, cost reduction and efficiency improvement have achieved significant progress. Unitech has reduced the cost of core components of humanoid robots by 40% through material innovation and modular design; meanwhile, by using bionic joints, the life of key components has been extended to more than 30,000 hours. These breakthroughs lay the foundation for the commercialization of humanoid robots at scale.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Distributed computing networks will become the key to solving the robot computing power bottleneck. Wang Xingxing pointed out that the computing power of humanoid robots themselves is limited, with peak power consumption of about 100 watts, equivalent to the computing power of a few mobile phones. \"Through federated learning to realize multi-robot data sharing, we expect to reduce the training cost of a single robot to the level of ten thousand yuan by 2027.\" This prediction means that the training cost of robots will experience a drastic decline, clearing the way for large-scale deployment.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F08\u002F31f8cbb3de4f454ba9e3aac8dc09e84a\u002FAA1KgOzj.jpg\" width=\"undefined\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Wang Xingxing predicted: \"In the next few years, the annual shipment of humanoid robots across the industry will double every year. If there are greater technological breakthroughs, it is even possible that in the next 2-3 years, the shipment could suddenly reach tens of thousands of units in a single year, or even millions.\" This explosive growth will first appear in industrial scenarios, and the automotive manufacturing industry is expected to be the biggest beneficiary.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Industry ecosystem construction is also accelerating. The Hangzhou Embodied Intelligence Application Pilot Base was recently launched, integrating Unitech's \"best body\" with the \"most powerful brain\" of Huawei and Alibaba Cloud, building an \"computing power + data + model + scenario application\" ecosystem, accelerating the large-scale implementation of embodied intelligence in the industrial field.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">When will the \"ChatGPT moment\" come for the robot industry? Wang Xingxing gave a forecast: the fastest is 1-2 years in the future, and the slowest is 3-5 years. The sign of this critical point is simple and straightforward - when a humanoid robot is brought to an unfamiliar venue, it can understand the random instruction \"bring this bottle of water to a certain audience\" and complete the task smoothly, which signals the true start of the era of robot intelligence.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News Source】Guashigas \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.msn.cn\u002Fzh-cn\u002Fnews\u002Fother\u002F%E5%88%AB%E7%9B%AF%E6%95%B0%E6%8D%AE%E4%BA%86-%E5%AE%81%E6%A0%91%E7%8E%8B%E5%85%B4%E5%85%B4-%E6%A8%A1%E5%9E%8B%E6%89%8D%E6%98%AF%E6%9C%BA%E5%99%A8%E4%BA%BA%E7%9C%9F%E7%93%B6%E9%A2%88\u002Far-AA1KgR1j?ocid=msedgntphdr&cvid=67764a46dd014287b627707a75cbb518&ei=25\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">http:\u002F\u002Fu5a.cn\u002FSFMrN\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is reprinted by the network to provide readers with more information and news. The content involved does not constitute investment or consumption advice. If there are any doubts about the facts of the article, please verify with the relevant parties. The views of the article are not the views of the network, and are for reference only.)\u003C\u002Fspan>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A61083b60-10d4-44fa-9936-b83a55598c45%3A0.wav?Expires=1774838498&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=oUwzmt8hhpOvaKekuasFCFNMIY4%3D","61083b60-10d4-44fa-9936-b83a55598c45",17114286]