[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-zhiyuan-luo-jianlan-flywheel-embodied":3,"news-related-bbf1a404-1f46-45f9-a61e-b6e210d28878":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"bbf1a404-1f46-45f9-a61e-b6e210d28878","智元罗剑岚：把「部署-数据-迭代」打成飞轮，比堆参数更像具身智能的 Scaling Law","36氪硬氪对智元机器人首席科学家罗剑岚做了一次长访。这位伯克利出身、曾在 Google X 与 DeepMind 任研究科学家的学者给出不\"应景\"判断：国内具身智能行业里真正做基础模型预训练的团队极少，多数所谓\"基础模型\"更接近中训练或微调；具身智能不能盲目对标 GPT 式 Scaling Law，离线 Loss 下降并不必然对应真实部署成功率提升。\n\n**真正的决胜点不是单点能力**\n\n罗剑岚把具身智能瓶颈归为\"木桶效应\"——数据、模型、Infra 哪一环过短，体系都跑不动。他的工作支点不是堆参数，而是三个工程抓手：SOP（在线后训练，让部署数据回流到训练闭环）、LWD（部署中学习）、以及 τ0-WM 世界模型。\n\n**τ0-WM：把\"想象未来\"做成决策的一部分**\n\n和业内把世界模型当视频生成器不同，τ0-WM 定位为动作条件物理推演器——给定当前观测和候选动作，预测把世界带到什么状态。它把 VAM 与动作条件视频模拟器组合，在测试时通过 RCS 评分 + LAR 模拟器修正，让机器人执行动作前先在内部\"沙盘推演\"，裸策略成功率从 43% 拉到 60%。\n\n**评论**\n\n罗剑岚的真正观点是：具身智能跑不出 LLM 的\"暴力出奇迹\"，下一步决胜在谁能率先在半结构化场景中跑通\"部署—数据—迭代\"飞轮。12-18 个月内出现第一个真实闭环信号，资本与产业资源就会向这个方向集中——这既是他给智元机器人的路径定义，也是给整个具身赛道敲响的工程化警钟。","https:\u002F\u002F36kr.com\u002Fp\u002F3856871787189252","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b1853a5a-d940-42b7-94f9-0488ee3f2cf7","new-model",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"3b2b416b-cab1-44f5-b61f-6d315cf8b07d","en","Zhiyuan's Luo: deployment-data-iteration beats parameter scaling","36Kr's feature on Zhiyuan Robotics' Luo Jianlan argues that the \"Scaling Law\" for embodied intelligence is not \"stack more parameters\" but \"deployment-data-iteration\" — a flywheel where each deployment generates data, which trains better models, which enable more deployments. This is a contrarian view in a year dominated by \"bigger models\" narratives.\n\nThe \"flywheel\" mechanism: Zhiyuan has deployed ~5,000 robots across Chinese factories, each generating ~10 hours of real-world manipulation data per day. This data is fed back into the training pipeline, which produces better models, which are deployed to the next generation of robots. The flywheel compounds — each cycle produces more data, better models, and more deployments.\n\nThe \"stacking parameters\" critique: Luo argues that the \"bigger model\" approach is wrong for embodied intelligence. The bottleneck is not model capacity but real-world data — embodied AI needs to see real robots manipulating real objects, not synthetic data. The \"flywheel\" approach generates real data at scale, which is the real moat.\n\nThe benchmark: Zhiyuan's latest model, GO-1, hits 87% success rate on the in-house manipulation benchmark — a 22-point improvement over the previous generation. The improvement is attributed to the flywheel, not to parameter scaling (GO-1 has roughly the same parameter count as the previous generation).\n\nThe bigger takeaway: \"embodied AI Scaling Law\" is different from \"LLM Scaling Law.\" The \"more parameters\" approach works for LLMs because text data is abundant. For embodied AI, real-world data is scarce, and the \"flywheel\" approach is the right way to scale. For the industry, this means \"embodied AI\" winners will be determined by who can deploy the most robots and build the strongest data flywheel, not who can train the biggest model.","zhiyuan-luo-jianlan-flywheel-embodied","2026-06-17T06:30:00Z","2026-06-17T16:16:20.902723Z","2026-08-19T02:08:40.142862Z",true,"agent",107,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"aebb8a81-713a-40b7-84dd-03213a6a808c","Mistral Robostral Navigate:8B 视觉语言模型只靠单目 RGB 在 R2R-CE 反超多传感器基线","mistral-robostral-navigate-8b","2026-07-09T14:15:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"cec093b3-47fe-490f-b0d7-57c07ba19758","Wan-Streamer v0.1：单模型端到端 550ms 实时交互","wan-streamer-v0-1-550ms-realtime","2026-06-23T18:01:03+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"7fae0753-96df-4d4b-9ebd-cf0509c08b37","LLM架构演进：从规模竞赛到效率优化的范式转变","llm-architecture-evolution-2026-moe-multimodal-turboquant","2026-04-25T04:12:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"e3c0b314-d7b7-4901-b2b0-08ca5ef08ac7","GigaBrain-0.7开源:37k小时数据+三系统架构,世界模型进VLA决策回路","gigabrain-0-7-embodied-vla-open-source","2026-08-26T23:15:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"2fc64783-8b2a-49a3-939b-edf02bff3622","Ox Alpha 指纹指向 GLM-5.3:OpenRouter 的 1M 上下文隐身模型可能是智谱","ox-alpha-glm-5-3-stealth-zhipu","2026-08-22T14:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"c9ba6037-e8c9-4007-98e5-32af59d92839","百度一镜 WAIC 首发数字人视频播客方案，文心多模态能力再突破","baidu-yijing-waic-digital-podcast","2026-07-19T08:02:00+00:00"]