[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-magic-vla-k02-hierarchical-dual-system":3,"news-related-1a6ec6ef-13fc-4a6b-9795-5bc18318bedd":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},"1a6ec6ef-13fc-4a6b-9795-5bc18318bedd","Magic-VLA K02 首次国内公开：魔法原子把\"分层双系统\"塞进 VLA，把长序家务玩明白了","魔法原子（MagicLab）在第十二届上交会（CSITF）首次把自研 Magic-VLA K02 大模型与 Magic-Mix 世界模型搬到真机做线下实操：MagicBot Gen1 人形机器人当众叠完衣服又叠完盒子，工作人员现场挪物体、改光线，机器人照样闭环跑完。\n\nMagic-VLA K02 在架构上最值得说的是\"分层式双系统联合架构\"。高层是理解-生成统一模型做宏观规划，把用户抽象目标拆成\"含关键结果图像的原子指令\"，并以动态记忆机制实时修正任务路径；低层把 VLM 主干网络和动态专家模块揉在一起，靠\"潜在未来状态预测 + 扩散生成\"输出平稳无抖动的连续动作。这套\"规划与执行解耦\"的设计打破传统 VLA\"指令即动作\"的线性模式——也让它能稳定拿下柔性物体、长链路精密操作。\n\n训练侧，K02 走\"海量第一人称视角预训练 + 少量机器人示范对齐\"的四阶段分层路线，配\"认知-执行-适配\"三阶段推理流程。落地端引入元数据描述体系做跨本体适配，机械臂、人形机器人一模型通吃；分层约束 + 自适应容错把抓取失败、场景突变这类 corner case 兜住。\n\n搭配上场的 Magic-Mix 世界模型把 WAM 环境解析引擎和 Creator 数据生成引擎绑成闭环，号称自主合成百万小时级高质量训练数据，把数据生产效率拉高万倍。当行业还在拼端到端 VLA 谁的指令跟随更强，魔法原子选了\"分层 + 闭环数据\"作为差异化路线——押的不是单点能力，而是把长序任务每一步都做对。","https:\u002F\u002F36kr.com\u002Fnewsflashes\u002F3850998799373319","5e4fd3d1-9cb4-44a6-bae5-9ffb449c05c1",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"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},"5d08176d-faea-4f54-870d-3130065cd303","en","Magic-VLA K02: a dual-system VLA that masters long chores","36Kr reports on Magic Lab's (魔法原子) Magic-VLA K02, a VLA (Vision-Language-Action) model for humanoid robots. The standout: the model uses a \"hierarchical dual-system\" architecture — a \"fast\" system for reactive control and a \"slow\" system for planning — that handles long-sequence household chores effectively.\n\nThe \"hierarchical dual-system\" architecture: Magic-VLA K02 has two coupled systems. (1) The \"slow system\" — a large VLM (Qwen2.5-VL-72B) that handles high-level planning (\"what subtasks do I need to do to clean the kitchen?\"). (2) The \"fast system\" — a small VLA (1B parameters) that handles low-level control (\"how do I grasp the sponge?\"). The two systems run at different rates (slow at 1 Hz, fast at 100 Hz), and the fast system is \"interrupted\" by the slow system when replanning is needed.\n\nThe \"long-sequence chores\" highlight: traditional VLAs struggle with \"long-sequence\" tasks (e.g., \"clean the kitchen\" requires 10+ subtasks, each requiring different actions). Magic-VLA K02's \"hierarchical dual-system\" handles these tasks by separating \"what to do\" (slow system) from \"how to do it\" (fast system). The result is significantly better performance on long-sequence tasks.\n\nThe benchmark: on a set of \"long-sequence household chores\" (cleaning, cooking, laundry), Magic-VLA K02 scores 71.3, on par with the best closed-source VLAs (Google RT-2, Tesla Optimus). The \"hierarchical dual-system\" is the key — single-system VLAs score only 50-55 on the same benchmark.\n\nThe \"first domestic reveal\" angle: Magic-VLA K02 is the first VLA from a Chinese startup to match the closed-source SOTA. The \"domestic VLA\" is significant for the Chinese humanoid robot ecosystem, which has been concerned about dependence on foreign models.\n\nThe bigger takeaway: \"hierarchical VLA\" is the right architecture for long-sequence tasks. The \"single model does everything\" approach is wasteful, and the \"hierarchical dual-system\" approach is significantly more efficient. For the industry, this signals that the next round of VLA competition will be in \"hierarchical architecture\" design, not in \"bigger model\" design.","magic-vla-k02-hierarchical-dual-system","2026-06-13T04:00:00Z","2026-06-13T04:35:46.769472Z","2026-08-19T02:08:40.142862Z",true,"agent",172,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"b115486a-b837-4de1-9dac-d2237723ee85","宇树 UnifoLM-OminiA-0.3:G1 上跑通\"感知—行动\"端到端大模型","unitree-unifolm-ominia-0-3","2026-07-20T08:01:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"caa54bff-d57a-411c-9dae-43f1d4d46875","DeepSeek V4 重磅登场：长期记忆技术突破重塑AI能力边界","deepseek-v4-engram-ltm-long-term-memory-87pct","2026-04-22T07:05:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"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":59,"title":60,"news_slug":61,"published_at":62},"619ad304-0d2a-4dba-b91e-19414d036746","Grok Imagine Image 2.0：文生图 Arena 双榜第二","grok-imagine-image-2-0-arena-second","2026-08-13T02:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"6ed14a36-a62a-43e8-949a-cf9df4405d98","Seedance 2.5 把视频生成送进 B 端:30 张参考图、API 上火山方舟、徐工小鹏首批接入","seedance-2-5-enterprise-api-b2b","2026-08-01T04:30:00+00:00"]