具身智能在家庭场景的落地正从概念验证走向产品化探索，但如何在非结构化环境中实现长时序任务、多模态交互与持续成长，仍是行业共同面对的难题。
本文整理自乐享科技联席 CTO、穹明智能总经理李元庆在 AICon 全球人工智能开发与应用大会 2026 深圳站的分享《轻量化异构双臂 × VLA/世界模型：家庭服务具身机器人的落地实践》。他系统梳理了家庭机器人的形态定义、轻量化本体设计、分层软件架构以及仿真到真机的数据闭环实践，重点讨论了 rule-based 与 modular-based 策略的协同关系、端到端方法在现实约束下的差距、软硬一体开发中的自动化代码生成，以及机器人自我进化系统 RSI 的构建思路。
以下是演讲实录（经 InfoQ 进行不改变原意的编辑整理）。
背景与挑战：家庭机器人离真实需求还有多远
坦率地说，目前大家对 to C 具身智能是否到了落地的节点仍有疑虑，但家庭场景中的真实需求是存在的。每个人在家里都希望有机器人能够帮你做越来越多的事情，我们对于厨师、保姆、阿姨这些角色的期待，本质上就是希望从大量重复性家务中解放出来。这几年工作日里大家越来越忙碌，很多人选择点外卖而不是自己做饭，那么周末能不能把时间还给自己，这些事情在 to C 场景里大有可为。
在具身智能这个赛道，数据、本体、场景和硬件就像一只手的五个手指头，只有握在一起才能成为一个拳头。这个比喻非常形象。对于现在做机器人的团队来说，这几个维度目前还没有真正做到清晰的解耦。做硬件的团队、做模型的团队、做软件的团队，彼此之间仍然需要非常多的打磨。原因很简单，做 AI 的人都会围绕数据和落地的应用成功率展开工作，场景决定了数据的结构，数据结构又决定了模型的参数量和形态。而一旦在真机中引入 physical 本体，本体产生的数据中，on policy 强化学习和 post training 那一部分依然是至关重要的。这说明硬件本体、数据采集和策略训练之间并没有完全分开，它们天然地绑定在一起。
家庭机器人的理想与现实之间还存在明显的鸿沟。to C 有一个好处是数据的 diversity，它可能让 AI 的整个基线跑得更快。而且在应用端，我们可以通过 IP、情绪价值和营销策略让用户更早地接受产品，这是 to C 与 to B 不一样的地方。但与此同时，to C 场景中的本体往往还比较笨重，能不能做得更轻、更小，是一个非常现实的问题。举个例子，帮我们拿一杯水这个看似简单的任务，在复杂家庭环境中其实是一个长时序任务。机器人需要先去找水，找到水的过程中可能需要先打开一扇门，走到冰箱旁边，打开冰箱门，再在冰箱内部找到那瓶水并把它拿出来。整个过程是一个 long context，也是一个 long horizon 的任务。交互的多模态与情感化同样重要。当用户说“帮我拿一杯水”时，如果机器人只是一个多模态模型，它怎么知道哪一杯水属于用户？这需要前面的 memory，也需要非常准确的 reasoning 来完成移动。
非结构化的家庭环境是另一个挑战。这几年一些团队在做酒店、康养等半结构化场景，但家庭场景始终是最后一个复杂、高动态、高灵活度的地方。此外还有隐形的意图与长短记忆问题。用户说出的一句话携带的信息量其实非常少，机器人想把任务做对的难度极大。“出厂即精英”的假设也不符合实际需求。我们在调研中感受到，机器人走进家里的时候更像一个笨笨的小朋友，它需要像阿姨一样熟悉用户的生活习惯。比如用户说“帮我拿杯水”，早上九点拿的可能是一杯白水，下午三点拿的可能是一杯冰可乐。这些上下文习惯必须被长期记忆和学习，这件事才有可能真正做成。
产品形态与轻量化本体设计
基于 to C 的思考，我们采用了一个比较务实的产品策略。当机器人出现在家庭中，它首先面临的问题是自己到底是什么：是一个工具、一个家电，还是一个家庭成员？我们把产品身份定义为三种形态：pet、partner、assistant。
三者的区别非常清晰。Pet 提供纯粹的情绪价值和功能价值，更像日本之前 LOVOT 那种笨笨的机器人，用户感受到的是多巴胺和催产素的分泌，购买决策由营销和情感驱动。Partner 的核心能力是养成系能力：它走进家里时一半是情绪价值，能在交互和简单移动操作中给出情感上的理解，同时能完成基础任务，并在这个家庭中持续成长。Assistant 则纯粹是功能价值，像钢铁侠里的贾维斯，或者那个笨笨机器人帮你拿灭火器去喷着火的工作台。我们认为这三类形态是未来具身智能走进家庭时最真实的产品形态。
在这些形态背后，我们在整个公司体系下做了一些前沿探索。技术侧我们自研了高扭矩密度电机。之所以要自研，是因为前两年我们在国内供应链中寻找能把机器人做轻做小的方案时遇到了限制。如果希望机器人保持同样的末端负载和同样的工作空间，就必须把电机尽可能做小，这样才有机会把上半身做轻，底下也做轻。这是轻量化最本质的路径。另外在情绪识别上，我们发现语言里包含的信息太少，能不能把语音的语调、语速都作为参考输入，把人的表情、说话时间和前面的上下文联系在一起，这是我们智能大脑的一部分工作。机械臂、灵巧手和 VLA 模型共同组成智能大脑，这一点后面会在软件架构上展开。
我们最近发布了一款轻量化的家庭协作机器人。当前整机重量是 30 千克，我们还在继续压重量。它的底盘很小，直径只有 45 厘米，比家里的扫地机器人大一圈，接近一个大号花盆。升降高度从 82 厘米到 147 厘米，82 厘米，机器人蹲在那里像家里一个大一点的桶，147 厘米则接近身材矮小女生在操作时胳膊的高度。
我们做了三个版本，大家在看到产品时可能会注意到两个机械臂有的一样、有的不一样。两个七自由度臂的方案，是这几年大家都在讲 data driven，通过 common sense 策略，用像人胳膊一样的自由度去完成同样的事情。但这对机器本体的要求非常高。另一个方案是一个四自由度和一个七自由度配合。这里的设计逻辑是：人是分左手和右手的。在做穿针引线这类动作时，左手往往负责固定的握持和约束，右手执行精细化操作。比如拧瓶盖时，人其实是用两只手协作，但如果放到标准的笛卡尔空间里考虑，左手可以固定住，只要把握持动作做好，右手来做精细操作。因此我们设计了四加七的结构。四自由度这边整条臂可以做得非常轻，配合拖链和型材结构，完成高处的动作时前后摩擦力需要对抗，但抬起的动作非常轻，这样整个策略都依赖于一个稳定的约束。如果只看这一半，它是一个标准的笛卡尔坐标系，对 AI 算法来说，做 manipulation、动态视觉追踪和策略都轻松很多。底盘方面我们提供了全向轮和差速轮两个版本，都是基于成本和实际业务中的发现来选择的。比如机器人会缠到头发、需要越过小的门槛缝、用户对柔性化灵活度有要求，这些我们都做了大量测试。最终推出目前这三个产品版本，是希望用户在用的时候能感受到工程师在设计时的认真。
我们发布了一支产品视频，画面是温馨的艺术化渲染。视频里展示了一个理想场景：家里不只是一间房子，有人提前为你把生活的杂乱捋清，每句叮嘱都有回应，小恩协作机器人会把你顾不上的小事一件一件放在心上，出门时提醒你带上口袋里的爱心，疲惫时轻轻接住你。虽然视频是艺术渲染的结果，但大家可以从中感受一下，如果这个机器人纯靠遥操作，它能完成的工作空间和想象空间到底在哪里。接下来我把它的技术底座和实际做到的东西跟大家分享。
软件架构：rule-based 与 modular-based 的分层协同
软件架构上，我们的设计思路是分层的。这几年大家听到 agent 比较多，前两年听到端到端比较多。现在大家听到 rule-based 这个词可能觉得陌生，但它其实在这个赛道已经存在很久了。状态机、behavior tree 都是 rule base 的典型代表，本质上就是明确地遇到一个情况然后产生一个行为，类似 switch case 或者 if else。扫地机里很早就积累了大量这样的经验：卷到窗帘就退出来，卷到地毯就退出来，被门挤了一下就退出来。这些都是之前工程师积累的 corner case 和长尾问题经验。家里并不是从来没有过智能机器人，扫地机器人曾经走进过家里，在移动导航和避障上积累的经验非常有价值。我们能不能把这些高解释性、高可靠性、低成本的策略作为机器人的兜底？这是我们的第一层，它也像无人驾驶里的 questions 和 checker 一样，承担安全兜底和基础行为保障的角色。
第二层是 modular-based。这个阶段大家做的策略其实很像 agentic 架构：把 perception、decision making、reasoning、memory 以及后面的 action 和各种 skills 都拆开去做。这件事很有意思，因为通过 modular-based 策略，能把一个机器人的软件体系先完整地连在一起。它看起来可能很慢，泛化能力也不够强，但它能够把前面各个可解释性的环节连通起来。它主要解决了一个困惑行业很久的问题：可控性的 AI 到底怎么做？后来我们发现，可控性的 AI 可以总结为一句话：让每一个模块都完全可控。这句话听起来抽象，但核心意思是不能让所有模块细节都变成黑盒，中间必须有一个阶段是拆分的。拆分之后，我们能看到显式的结构特征。在早期可能是专家网络或稠密结构，通过这些显式结构特征，机器人可以快速学会一个东西，实现 few shot，从而替代掉端到端在现阶段难以落地的部分。
端到端仍然是未来要努力去攻克的方向。目前在 modular-based 架构下，perception 环节需要反复看一个东西，调用摄像头的次数很多。端到端则是像看一个事情就推一个 chunk，摄像头调用次数大幅减少，能耗会压缩，模型大小会变小，API 调用次数也会减少。只要考虑到一天的落地运行，稳定性和能耗就是必须面对的问题，端到端一定要做。而且端到端的信息利用效率是最高的。modular-based 阶段我们做分块，是因为自然语义文本和环境信息里包含的内容太少，我们必须在合适的节点把最合适的数据交给它，否则有些地方数据是浪费的，有些地方数据又实在太少。这一页内容做得比较早，后面论文更新会进一步展开。
把软硬一体这件事展开来看，我其实很早以前就开始做机器人，那时候 state machine 和 behavior tree 比较多。一个最真实的感觉是，这几年大家讲 AI 讲得有点乱了，总说 AI 和 data driven 体系需要重新做一遍。但如果真的去看，很多东西其实没有那么大的变化。研发策略和研发效率需要在 artifact 这一层做明确的拆解。硬件这一层，我们通过自动化的策略来把问题解决的速度提高。这几年确实出现了很多新名词，但本质上这些是我最开始做传统机器人赛道时就在用的东西。蓝色标注的部分都是现在新加的。AI 对机器人开发者的意义在于，它把能力边界扩大了，让性能有机会更高，这些蓝色空间为机型带来了更多能力。AI 在 robotics 能力上的赋能确实非常大。
软硬一体的另一个进展是，我们前面已经能在一些成熟的硬件本体上完成 general 的物体放置、刚性物体和均匀尺寸物体的操作，当时成功率已经可以做到百分之百。这类似于大家前面用 grasp net 识别抓取位姿，后面用 IK 去执行，已经很成熟了。今年我们在做的事情是，在自由度高、约束复杂的情况下，把 WBC 的整个 control 做得更强，让它完成更强的跟随能力、更强的避障能力，以及模式层中更好的能力。更让我自豪的是，这一块代码已经不再是人写的了。我们在某个仿真的小 simulation 闭环里，机器人可以自己去做 evaluation，然后 AI 自己去改代码。人在里面大部分参与的环节，只是在决策链条中有些 checklist 选择走哪条路径可能收敛更快。最终我们看到的结果是，硬件本体在那里做的事情成功率越来越高。
从 modular-based 到端到端的工程演进与仿真闭环
在整个 infra 上，这几年大家做 VLA 或者 VAM，确实看到了一些不错的东西。但回到本质，主要还是 sim to real，也就是强化学习那一套和后面的 behavior cloning 几条路线。这里想跟大家分享的是机器人赛道所谓的 RSI 可能是什么。简单来说，就是把 target model 构建、system model 构建中间的 development、deployment、verification、evaluation 串起来。设定一个目标之后，能不能把 system 友好地整合在一起；这个 system 包含 robot 本体、hardware、software 和整个架构。development 是开发过程中你开发了什么，我们做的是前面的 modular-based，然后往端到端再往上一层走。
deployment 的难度被普遍低估了。做一台 deployment 很简单，但做到三十台到一百台的时候，你会发现非常依赖云端厂商的边端推理能力，需要非常稳定的数据传输和非常稳定的信息链路。再往下是 verification，你必须清晰地完成任务验证，搞清楚什么成功了、什么失败了、哪里失败、怎么失败的，把它清晰定义出来再做评估。这里有一个比喻：做一台特别好的机器人，就像一把刀和一把尺的关系。你必须有一把非常强的尺，才有机会做出一把更好的刀；再把精细度高的刀配上这把尺，才能形成闭环，把事情越做越好。所以 RSI 的底层就是一个自动化系统，只有这样才能看到 evolving 或者 evolution。
在演示中，我们连续询问机器人状态：“小恩小恩，录制得怎么样了？”“小恩小恩，结束录制。”“小恩小恩，构建的怎么样了？”它回复：“构建进度 60%、空间定位完成。”接下来是物体更换为挂饰、物体是个逻辑等一系列指令。用户继续问：“小恩小恩，你的记忆里有什么？”它回答有视频、遥控器、仪表盘、鞋子、小包裹盒、显示器和音响。然后用户说：“小恩小恩，帮我拿瓶可乐。”机器人回答要去看一下，确定目标物品并执行抓取。这是一个非常真实的 agentic 架构演示，或者说是我们称之为 modular-based 架构的真实演示。这里的细节是，整个过程一点 hard code 都没有，所有东西都靠系统一点一点叠出来。前面我分享过，在 simulation 里面我们已经把 AI 自己的验证功能跑通了。现在在 real world 下，我们也完成了一部分验证，能看到的效果就是视频中展示的那样。
机器人的智慧不像人类那么强，所以我们的 reasoning 做得比较复杂。人类的 recognition 能力比想象中强很多，而机器人在看东西时很像我这种近视眼：最开始视野模糊，先构建整个语义，用长链条去大概知道每个房间有什么、在哪个位置。在家庭场景中，它需要知道自己在厨房、卧室还是客厅，需要知道不会变的东西是家具，比如桌子、椅子、柜台，甚至门从房子建立第一天起就不会变。在门之上，我们用类似八叉树的策略构建上面所有物品的关系。找东西的过程就像近视眼找东西：先看到一个模糊的位置，看看哪个东西疑似目标，然后做 template 匹配、多模态识别、槽位和微辨识，大致知道东西在哪里之后，再“戴上眼镜”仔细靠近，找到它并带回来。大家看到的任务链条时间很长，但成功率提高了特别多。这比之前纯粹的端到端在现阶段效果好很多，同时它也是一个真正具备泛化性、用有限能力完成任务的方法。
在这些 modular-based 的基础上，我们也在看怎么在小的 skills 上、在 action experts 这一层做得更好一点。大家之前可能比较熟悉 digger 或者 husserl，以及强化学习之后的很多东西。我们做的事情是把策略里的各种数据策略都走了一遍。比如我们做了自己的 UMI 夹爪，虽然没有产品化，但在实验室里一直在用；一臂的策略也做了，centridig 的策略、后面 in the loop 的策略全部过了一遍，并且自己搭建了整套 simulation evaluation 的 loop。我们尝试在采集数据的过程中，把数据清洗和上传链条在云端完整搭建出来。实际的效果是，我们用各种各样的数据做了各种各样的训练，把真实实验室中的流程搭建了出来，去构建小的 VLA、小的 VAM、小的 skills。我达到的结论是，之前做的有些东西很值得参考，七比二比一是我们目前拿到的一个比较真实的数据结果。
今年上半年我们开始看 human centric 的工作。我们的做法是拿人的第一视角，用手去做一些简单的操作，然后把 assumption 拿出来，下一件事开始给它做，把不需要的部分去掉，现在加上仿真的隐形机械臂，把机械臂的工程、IK、动力学和运动学都模拟出来。最早做这个是为了提高整个 human centric 对有遮挡物体的想象能力。这一块用的就是李飞飞老师现在比较强调的 real to sim to real 的事情。我们这个链条做了很多，希望大家能一起看到未来。
在 demo 这一层我们做得更多一些，早期我们用 oppa max 机械臂去做放木块的任务。任务中一旦发现中间某些链条不够成熟、成功率比较低、会走错方向，人就会介入。我们需要专家在小空间里引入意见，看看怎么把成功率提得更高，然后人工介入纠正机器人的策略。所谓力反馈或者主动服务，就是在同样空间里人工介入纠正策略。我们在 demo 体系下做了很多工作，把它磨成一个内部使用的实验平台。这样采集二十条、三十条数据之后，我们明确看到，小物块这种精细化操作，之前成功率在仿真上只有 30%左右，后面在这种操作下大概能提到 60%。后来我们又做了一些更优秀的操作，把成功率提到了 90%。
从整个探索过程来看，我们其实是在从 demo 往产品化一步步走。前面提到的很多工程细节，本质上都不是在做 simulation demo，而是在打磨更可靠的产品。我们从工作堆砌转向场景深耕，从技术驱动转向数据驱动。这是这几年最真实的感受：不再需要那么多工程师去疯狂写 corner case，不需要针对每一个长尾问题都花大量时间。AI 在 simulation 和 real world 中确实展现出了很强的效果。一个数据拿回来之后，当你建立了比较好的数据采集方式，并且增强了 verification 和 evaluation 能力之后，AI 能够理解自己到底做了什么，能够理解自己做错了什么，并且从中找到怎么能做得更好。这就是数据驱动带来的质变。
单机智能转向家庭共生是另一个重要的转变。机器人慢慢会成为一个 member，而成为 member 需要用户和机器人一起成长。我们希望最终能做出出现在家里的 pet、partner、assistant，这是真实的期许。
在工程实践中，自我进化系统的构建是一个持续迭代的过程。我们在仿真中建立了自动化验证闭环，让机器人自己去做 evaluation 和改代码，这个闭环不仅提升了开发效率，也让人从大量重复性工作中解脱出来。人参与的部分越来越少，主要的决策集中在选择哪条收敛路径更快。这个体系从仿真延伸到真机，虽然在真实世界中还有很多细节要处理，但已经完成了一部分验证。可以预见的是，随着 verification 和 evaluation 能力的不断增强，机器人在环境中的表现会越来越稳定，从能完成任务走到能可靠地完成任务。
从更宏观的视角看，机器人赛道的 RSI 本质上是在构建一个能够自我进化的系统。目标模型、系统模型、development、deployment、verification、evaluation 形成闭环，只有让这个闭环转起来，机器人才有可能不断进化。这个系统不能只停留在仿真阶段，必须和真实环境产生连接。仿真的成功率可以做到百分之百，但真实环境中的长尾问题依然大量存在。我们的策略是先用 rule-based 做高可靠性的兜底，用 modular-based 实现可控的智能，再逐步向端到端演进，这样既保证了当前阶段产品可落地的稳定性，也为后续的技术升级留出了路径。
做机器人是一件很浪漫的事情，也是一件让我很有使命感的事情。公司的很多小伙伴都认可这一点。我们想做的是一个伟大的机器人，希望大家能用它去做伟大的事情，这是我们对机器人最真实的期许。希望用户能看到的是一个稳定、好用的平台，希望机器人能实现自进化——仿真界的自进化，在真实世界里的自进化，在用户手上完成真正的自进化。从工作堆砌到场景深耕，从技术驱动到数据驱动，从单机智能到家庭共生，这三条路径就是我们正在走的方向。我们希望和伙伴、开发者一起构建软件的进化生态，让每一台机器人都能成为这个系统上的贡献。
热度延续：12 月 AICon 北京站见
12 月 18–19 日，AICon 全球人工智能开发与应用大会将在北京举办。本届大会以探索 A 应用边界为主题，设置 13 大技术专题，覆盖大模型后训练、推理优化、Agent 工程、知识工程、智能体评测与安全、自进化系统、AI 软件交付、原生架构及组织变革等核心方向，汇聚一线技术专家，深入分享 AI 技术突破与企业规模化落地的真实实践。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intelligence in the home setting is moving from conceptual validation to product-based exploration, but long time-series tasks, multi-modular interactions and sustained growth in a non-structured environment remain common challenges for the industry.
The paper is based on the joint CTO, General Director of Dome Wise, Lee Yuanqing at the AICon Global Congress on Artificial Intelligence Development and Application in Shenzhen Station 2026. His system combed the morphological definition of family robots, the light quantization prototype design, the layered software architecture and the simulation-to-real data closed-ring practice, focusing on the synergy between rule-based and modular-based strategies, the gap between end-to-end approaches and the constraints of reality, the creation of automated codes in soft and hard development, and the idea of robotic self-engineered systems RSI.
The following is a transcript of the presentation (a non-changed editorial by InfoQ).
Background and challenges: How far is a family robot from real needs?
Frankly, there are still doubts about whether to C's intelligence has reached the drop point, but the real needs in the family scene exist. Everyone at home wants robots to do more and more for you, and our expectations of the roles of cooks, nannys, aunts are essentially to be freed from a lot of repetitive household chores. People are getting more and more busy on working days over the years, and a lot of people have chosen to order takeout instead of cooking themselves, so can they give themselves back some weekends?
On a smart track, data, physical, scene and hardware are like five fingers in one hand, and only a fist can be held together. This metaphor is very graphic. For the team that is now a robot, these dimensions are not really clear. Hardware teams, model teams and software teams still need a lot of grinding. The simple reason is that all those who do AI work around data and land application success rate, the scene determines the structure of the data, and the structure of the data determines the model parameter count and the form. And once the physical body is introduced into the real machine, the data generated by the body, the part that's still critical is on policy and post training. This means that there is not a complete separation between the hardware matrix, data collection and strategic training, which are naturally bound together.
There is a clear gap between the ideals and reality of family robots. To C has the advantage of data diversity, which could make the whole baseline of AI run faster. And at the application end, we can use IP, emotional values and marketing strategies to get users to accept the product earlier, which is to C and to B. But at the same time, the nature of the to C scene is often rather heavy, and it is a very real question whether it can be done lighter and smaller. For example, the seemingly simple task of getting us a glass of water is in fact a long-term task in a complex family environment. The robot needs to look for water first, and in the process of finding water, it may need to open a door, go to the fridge, open the fridge door, find the bottle of water inside the fridge and take it out. The whole process is a long context and a long mandate. Interactive polymorphism is as important as emotionalization. When the user says, "Get me a glass of water," if the robot is a multi-modular model, how does it know which glass of water belongs to the user? It takes memory in front, and it also needs to be very accurate to complete the move.
The unstructured family environment is another challenge. Over the years, some teams have been working on semi-structured scenes such as hotels, rehabilitation, etc., but the family scene has always been the last complex, dynamic and flexible place. There is also the question of hidden intentions and long memory. It's very difficult for a robot to get a job right. Nor does the “employer-employee” assumption correspond to real needs. In our research, we felt that robots were more of a dumb little friend when they walked into the house, and they needed to be as familiar with the user habits as aunts. For example, the user said, "Get me a glass of water," maybe a glass of white water at 9:00 a.m. and an ice Coke at 3:00 p.m. These context habits must be studied and studied by long-term memory, which will make it possible.
Product form and light quantization prototype
Based on our thinking to C, we have adopted a more pragmatic product strategy. When a robot appears in the family, what is the first problem it faces: a tool, a household electricity, or a family member? We define product identity as three forms: Pete, Partner, Asistant.
The difference between the three is very clear. Pet provides purely emotional and functional value, more like the former Japanese LOVOT-like robotics, where users feel dopamine and oxytocin secretions, and purchasing decisions are driven by marketing and emotion. The core competencies of Partners are those of developing systems: Half of the time it enters the home is of emotional value, giving emotional understanding in interactive and simple mobile operations, while fulfilling basic tasks and growing up in the family. Assistant is purely functional, like Jarvis in Iron Man, or the stupid robot that helped you get a fire extinguisher to spray the fire. We consider these three types to be the most real product patterns of future intelligent entry into the family.
Behind these patterns, we've done some frontier explorations under the entire corporate system. On the technical side, we've studied the high-twirl-density power. The reason for this is that the last two years have been limited in our search for small-scale solutions for robots in domestic supply chains. If robots are to maintain the same end load and the same workspace, they must be kept as small as possible so that they have the opportunity to lighten the upper half and lower the bottom. This is the most essential path of light quantization. And in terms of emotional recognition, we find that the language contains too little information to use the tone and speed of the voice as a reference input, linking the human expression, the speaking time and the front context as part of our intellectual brain. The mechanical arm, the smart hand and the VLA model together form the intelligent brain, and this is followed by the software architecture.
We've recently got a light quantization family collaboration robot. The current overall weight is 30 kg, and we continue to press. Its chassis is small, 45 centimetres in diameter, and it's a circle bigger than a home-sweeping robot, approaching a large flower pot. Up and down from 82 centimetres to 147 centimetres, 82 centimetres, robots crouched there like a bigger barrel at home, 147 centimetres near the height of a small, short-sized girl's arm during the operation.
We made three versions, and when you look at the product, you might notice that the two arms are the same and some different. The two seven-free arm programmes are the ones that have been spoken of in the past few years, using the common sense strategy to do the same thing with the same freedom as a human arm. But this is very demanding for the machine. Another option is four degrees of freedom and seven degrees of freedom. The design logic here is that people are left-handed and right-handed. The left hand is often responsible for holding and restraining, and the right hand is performing fine-tuning operations in the case of actions such as needle penetrating. For example, when you twist a bottle cover, you actually work with two hands, but if you put it in the standard cartesian space, you can hold it on your left hand, and you can do it with your right hand. So we designed four plus seven structures. The whole arm of the four degrees of freedom can be very light, coupled with a tow chain and model structure, and the front- and back-to-back frictions are required to complete the highs, but the lifts are very light, so that the whole strategy depends on a stable constraint. If you look at only one and a half, it's a standard Cartesian coordinate system, and it's easier for AI algorithms to do manipulation, dynamic visual tracking and strategy. With respect to the chassis, we have provided two versions of the full-way and differential wheels, which are selected on the basis of costs and findings in actual operations. For example, we've all done a lot of tests on robotics that are entangled in their hair, that need to cross a small threshold, and that users need to be flexible. The final launch of the current three product versions is a hope that the user will feel serious about the engineer ' s design while using it.
We've got a video of publish, and it's a sweet artistic rendering. The video shows an ideal scene: the house is not just a house, and someone's got your life in front of you. Qing, there's a response to every tinkle, and the cyborg takes care of every little thing you can't take care of, reminds you to bring the love in your pocket and catch you when you're tired. While the video is the result of artistic rendering, you can see from it that if the robot is purely remotely operated, it can complete the workspace and imagine where it is. And I'm going to share it with you on the technical base and what it actually does.
Software architecture: rule-based and modular-based layers
In the software architecture, we have a layered design. You've heard more agents over the years, more end-to-ends over the last two years. Now you hear the word rule-based may be strange, but it's actually been on this track for a long time. The status machine, behavior tree, are typical of the rule case, essentially to encounter a situation clearly and then produce an act, similar to a case or if otherwise. A great deal of experience has been accumulated in the ground sweeper at an early stage: rolls to the curtains, rolls to the carpets, pulls out of the doors. These are the experiences of the former engineers with the Corner case and the long end. The home has not never been an intelligent robot, and a ground sweeper has walked into the home and has accumulated valuable experience in mobile navigation and shelter. Can we use these high-interpretative, high-reliability, low-cost strategies as the back of robots? It's our first level, and it's like the questions and checkers in the drones, taking on the role of a safety loop and basic behavioral security.
The second floor is modular-based. This is a phase in which you're doing a strategy that looks like the structure of anatomy: tearing down the structure, decision making, reasoning, memory, and later action and skills. It's interesting, because through the Modular-based strategy, you can connect a robot's software system in its entirety. It may appear to be slow and less extensive, but it can connect the explanatory links ahead. It solves a long-standing problem of confusion in the industry: how does AI do it? And then we found that controllability AI can sum up in one sentence: make every module completely manageable. It sounds abstract, but the central meaning is that all module details should not be turned into black boxes, and that there must be a phase in the middle that is split. After split, we can see the obvious structural features. At an early stage, it may be an expert network or dense structure, through which robots can quickly learn something and achieve a few shot, thus replacing the tip-to-end that is hard to land at this stage.
The end-to-end is still the way to go. At present, under the Modular-based structure, the perception link requires a double look at something and calls a lot of cameras. The end-to-end is like looking at one thing pushing a chunk, with fewer cameras called, less energy consumption, smaller models, less API calls. As long as one day is taken into account, stability and energy consumption are problems that must be faced and must be done from the end to the end. And end-to-end information is most efficient. The Modular-based phase, where we split up, is because the natural syntax text and environmental information contain too little, and we have to give it the most appropriate data at the right node, otherwise there are places where data are wasted and there are places where data are too few. The contents of this page are relatively early, and the latter will be further updated.
I started to be a robot a long time ago, when I started being a robot. One of the most real feelings is that over the years, people have been saying that AI is a little confused, that the AI and the Data DVD system need to be done again. But if you really look, a lot of things don't really change that much. R & D strategies and R & D efficiency need to be clearly dismantled at the level of artifact. The hardware layer, we solve problems more quickly through automated strategies. There have been many new terms in the past few years, but these are, in essence, the things that I used when I first started playing the traditional robot track. The blue labels are all new now. AI's significance to robot developers is that it expands the boundaries of capabilities and gives higher performance opportunities, and these blue spaces bring more power to the machine. AI is really very powerful at the ability of robotics.
Another advance in soft and hardness is that we have been able to complete the operation of success rate in a number of mature hardware prototypes for the placement, rigidity and even size of objects. It's like you're in front of the grasps net to recognize the grab position, and then IK to execute it, and it's mature. What we're doing this year is, in a context of great freedom and complexity, making the whole of the WBC more capable of following, avoiding, and better capabilities in the model. And what makes me proud is that this code is no longer written. We're in a real little simulation ring, and robots can do it themselves, and AI can change the code itself. Most of the people involved in it are just a couple of checklists in the decision-making chain that choose which path to follow might be more rapid. In the end, we see that what the hardware owner is doing there is getting higher and higher.
Project evolution and simulation closed loop from Modular-based to end-to-end
On the whole infra, these years people do VLA or VAM, actually see something nice. But back to the essence, it's mostly Sim to real, which is the kind of route that strengthens learning and the back. What we want to share here is what the robot track might call the RSI. In short, it's a string of taget model construction, system model construction, development, development, verification, evaluation. Sets a goal that can integrate systems amicably; this system contains the robot body, Hardware, software and the whole architecture. Development is what you've developed in the development process, and what we've done is we've got a modular-based front, and then we're going to end to end and go up.
The difficulty of deproyment is generally underestimated. It's easy to do a deloyment, but when you get to 30 to 100, you'll find that you're very dependent on cloud-end manufacturers for marginal reasoning, for very stable data transmission and very stable information links. Going down, you have to be clear about what worked, what failed, where failed, how failed, and define it for evaluation. Here's a metaphor: make a very good robot, like a knife and a ruler. You've got to have a very strong ruler to get a better knife, and then you've got to put a fine, tall knife on this ruler, and you've got to close the rings and do the better. So the bottom of RSI is an automated system, which is the only way to see evolving or evollution.
In the demonstration, we asked the robotic state: "How's it going? How's it going?" "Aun's, how's it going?" "Aun's, how's it going?" And it said, "60% construction, space positioning." This is followed by a series of instructions, such as the replacement of the object with a label, the object as a logic. The user goes on to say, "What do you remember?" It responded with video, remote control, dashboard, shoes, small parcel boxes, monitors and sound. And then the user said, "Em, get me a Coke." The robot replied that he was going to take a look, identify the target item and execute the grab. It's a very real demonstration of the agentic architecture, or what we call the Modular-based architecture. The details here are that there's no hard code in the whole process, and everything is stacked up by the system. I've shared it before, in the simulation we've run out of AI's own authentication. Now, under real world, we've done a part of the validation, and the effect we can see is what the video shows.
The robot's intelligence is not as strong as humans, so our resoning is complicated. Human recognition is much better than I thought, and robots look at things like my near-sighted eyes: they start with a blurry vision, they build the whole semantic, they use a long chain to figure out what and where each room is. In the family context, it needs to know whether it is in the kitchen, in the bedroom or in the living room, and what does not change is furniture, such as tables, chairs, counters and even doors from the first day of the building. Above the door, we use a strategy like the eight fork tree to build relationships between all the things above. The process of finding things is like looking for things in the near-sighted eye: first looking at a vague position, looking at what is suspected of being a target, then doing template matching, multi-modular recognition, slots and micro-defining, generally knowing where it is, then "sights" close, find it and bring it back. The task chain was seen to be long, but success rate increased particularly. This is much better than the previous pure end-to-end results at this stage, and it is also a truly broad-based and limited-capacity approach to the task.
On the basis of these modular-based, we're also looking at how we can do a little better on the small skills, on the action floors. You might have known a lot about digger or husserl and a lot of things after intensive learning. What we're doing is we've gone over all the data strategies in the strategy. For example, we made our own UMI claws, which were not product-oriented, but were used in the lab; the one-armed strategy was done, the centridig strategy, the back of the strategy in the loop, and we built the whole set of simulations. We have tried to build a complete data-cleaning and upload chain on the cloud in the process of collecting data. The practical effect is that we've done a variety of trainings with a variety of data, and we've built up processes in real laboratories to build small VLAs, small VAMs, small skills. I have come to the conclusion that there are some things that have been done before that are very interesting, and that seven to two to one is one of the more true data that we currently have.
In the first half of this year we started looking at human central work. The way we do this is to get a first perspective, do some simple things with our hands, and then take out the assumption, and then start doing the next thing, and get rid of the parts that we don't need, and now we're going to model the engineering, IK, dynamics, and sports of the mechanical arm. The first thing to do was to enhance the imagination of the whole human center over the hidden objects. This one's for Mr. Lee's more important thing now. We have done a lot in this chain, and we want you to see the future together.
We're doing more on the demo floor, and early on we used an oppa max arm to do woodwork. When certain chains in the middle of the mission are found to be premature, success rate is low and leads in the wrong direction, people intervene. We need experts to introduce ideas into small spaces to see how to raise success rate even higher and then artificially intervene in the strategy of correcting robots. The so-called force feedback or proactive service is the manual intervention of corrective strategies in the same space. We did a lot of work under the demo system, grinding it into an in-house experimental platform. So, after collecting 20, 30 data, we clearly see that this fine-tight operation of small matter blocks, which used to be about 30 percent of success rate in simulations, and then about 60 percent in this operation. And then we did something better, and success rate came up with 90%.
We're actually going from demo to product in the whole exploration process. Many of the details of the work mentioned above are not, in essence, being done as a similitude demo, but rather as a more reliable product. We moved from work piles to field tillage, from technology to data-driven. It's the most real feeling in the last few years: there's no need for a lot of engineers to write about corner case, no need to spend a lot of time on every long question. AI did show a strong effect in simulation and real world. When you get a data back, and you build a better way of collecting data, and you improve your ability to collect data, AI can understand what it does, understand what it does wrong and find out how to do better. That's the quality change that the data drive brings.
Turning one-way intelligence to family symbioticity is another important shift. The robot will slowly become a member, and become a member that needs to grow with the robot. We want to finally make a bet, a partner, an offer to show up at home.
In engineering practice, the construction of a self-evolutionary system is a continuous iterative process. We have built automated authentication loops in simulations that allow robots to do their own evaluation and change codes, which not only enhances development efficiency, but also frees people from a lot of repetitive work. The part of people ' s participation is becoming less and the main decision-making is focused on the choice of which route of convergence will be faster. The system extends from simulation to reality, and although there are many details to be dealt with in the real world, a partial validation has been completed. It is foreseeable that, with the growing capacity of verification and evaluation, robots will become more and more stable in the environment, from being able to perform to being able to do so reliably.
From a larger perspective, the RSI of robotics is essentially building a system that can evolve itself. Target models, system models, development, development, civilization, evaluation form closed rings, and robots can only evolve if they are kept closed. The system cannot stop at a simulation stage, but must be connected to the real environment. While 100 per cent of the true success rate can be achieved, there are still a lot of long-term problems in the real environment. Our strategy is to use rule-based as a high-reliability loop, modular-based to achieve controllable intelligence, and then gradually to the end, both to ensure the stability of products at the current stage and to set the path for subsequent technological upgrading.
Being a robot is a very romantic thing, and it's something that makes me very purposeful. This is recognized by many of the company ' s small partners. What we want is a great robot, and we want you to use it to do great things, which is our most real promise. What we hope to see is a stable and useful platform for robots to evolve — imitating the evolution of the real world, to evolve in the real world, to truly evolve in the hands of the users. From work piles to scene tilling, from technology-driven to data-driven, from single-computer intelligence to household symbiotic, these three paths are the direction we are taking. We want to build the evolutionary ecology of the software with partners and developers so that every robot can be a contribution to the system.
Thermal continuity: December AICon Beijing station
On December 18-19, the Aicon Global Congress on Artificial Intelligence Development and Applications will be held in Beijing. The Congress, which focused on exploring the boundaries of application A, set up 13 major technical topics covering the core directions of LLM post-training, reasoning optimization, Agent engineering, intellectual engineering, AI agent assessment and security, self-evolution systems, AI software delivery, original architecture and organizational change, bringing together first-line technical experts to share in depth the real practices of AI technological breakthroughs and enterprise sizes.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.