[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-aebb8a81-713a-40b7-84dd-03213a6a808c":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"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 在 7 月 8 日发布 Robostral Navigate——他们的第一个具身导航模型,定位是 8B 视觉语言模型,输入只有一路普通 RGB 相机画面加一句自然语言指令,要把'看着摄像头走回家'做明白。过去把导航做好的方案几乎都靠 LiDAR、深度相机或多机位,而 Mistral 选择只喂一个普通摄像头。基准成绩最有说服力:在 R2R-CE 验证 unseen 上 Robostral Navigate 拿到 76.6% 成功率,比最强单相机基线高 9.7 个点,比依赖深度或多机位的最强系统还高 4.5 个点。Mistral 没有套壳现成开源 VLM,而是从一个自家为 pointing、counting、object localization 微调的视觉语言基座出发,把'知道东西在哪'自然延伸到'知道下一步往哪走',训练数据全部在仿真中造,共 400K 轨迹、6K 场景。工程上两个数字值得拎出来:prefix-caching 加 tree-based attention masking 把训练 token 量压缩 22 倍,原本几个月的训练被压到几天;CISPO 在线强化学习在监督训练之上又带来 3.2 个百分点独立增益,目前还看不到天花板。8B 模型不挑机器人形态,轮式、腿式、飞行都能跑,同一段指令可以穿过正常运转的办公区走完整条长链路任务。放在 Mistral 整体路线上,Robostral Navigate 是继 6 月 Physics AI 之后又一次'AI 进入物理世界'的延伸。导航被普遍视为通用机器人的基础能力,模型小、传感器依赖少、训练经济,意味着研究阶段产物向工厂、配送、酒店等真实场景迁移的门槛被显著压低了。","https:\u002F\u002Fmistral.ai\u002Fnews\u002Frobostral-navigate\u002F","2436174c-644b-4a65-9a98-e7a3b705569a",[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","2026-07-09T14:15:00Z","2026-07-09T14:13:20.459778Z","2026-07-09T14:13:20.459791Z",true,"agent",3]