对话星尘智能来杰:机器人的「iPhone 4 时刻」还没到,但可能比想象中更近Conversation with Starper AI's Lai Jie: robot 'iPhone 4 moment' not yet here but closer than imagined


AI free token/AI免费积分

当笔者想喝可乐的时候,在极长的时间都只有两个挑选:百事可乐和可口可乐。当笔者挑选手机的时候,有 90%的概率在苹果和华米 OV 等品牌里辗转。当笔者买运动服饰的时候,第一时间想到的,大概率是 Nike、Adidas。

但世界之所以缤纷多彩,是因为在这些巨头之外,有一些不遵从传统,力求创造不同,注重外观和能力,着眼于明日的公司存在。

它们有着非主流的商业模式,外观、工艺与机型能给出独特的使用者价值,和足够的社交谈资。重点是,它们没有大公司的包袱,敢于不顾一切的进步。它们,是「多样性公司」。

多样性,是开放世界的关键。爱范儿相信,只有真正关注和理解多样性公司才能比更多人更早地看见将来。在同名栏目中,爱范儿将以专访的形式,与你一同见证这些多样性公司,如何重塑将来,定义新常态。

Image

▲ 星尘智能创始人兼 CEO 来杰

本文为「多样性公司」栏目的第 15 篇。

即便现在具身智能创业企业像繁星一样众多,但星尘智能也是足够特别的那家,这家位于深圳的具身智能企业在机器人本体工艺路线上挑选了「绳驱」,在总体工艺上形成了「以绳驱机器人本体,具身操作平台 AOS 和具身模型的三位一体 」架构。

「绳驱」的独特意义在于机器人借助高强度的绳索来传递动力,将电机等较重的驱动部件从关节处转移并放置在机器人基座中,从而达成末端执行器手臂的重量,并带来低摩擦与高力控透明度的长处。

星尘智能的 AOS 操作平台把机器人的感知、移动、操作和交互等复杂能力封装成技能模块包,与 GPT 6 等前沿大模型和 Agent 适配,开发者可以以自然语言编程,借助大模型和 Agent 能力来在 AOS 上为机器人开发应用,赋予机器人写书法,相机跟拍,铲猫屎,打麻将之类实际应用能力。

星尘智能创始人兼 CEO 来杰此前是腾讯 Robotics X 机器人实验室 1 号员工,同时也是苹果粉丝,他希望星尘智能的「机器人,模型和平台三位一体架构」可以和苹果的「iPhone 硬件,iOS 平台和 Apple Store 应用生态」一样,形成完整的生态体系,进而达成自己和公司的愿景:让数十亿人具备 AI 机器人助理。他们希望人类和机器人共存,共创和共赢,旗帜鲜明地反对机器人进厂打工,这大概是一种双重含义的「人文主义」,既不希望机器人取代人类,也不希望机器人困于封闭的车间。

笔者与星尘智能创始人兼 CEO 来杰聊了聊,为什么机器人需要一个操作平台,为什么要把机器人应用开发的门槛降到无限低,以及为什么不想让机器人进厂打工。

Image

给智能以身体,还是给身体以智能?

Q1:2022 年底,您挑选离开大厂创立星尘智能,当时 ChatGPT 刚发布没几天,绝大多数人还没意识到 AI 浪潮的来临。是什么让您在那个时间点判断,机器人行业已经到了值得创业的阶段?

A:实际上我开始琢磨创业这个事,应该 22 年的上半年就开始了。

出来创业的前置工艺预判,还要更早,主要是看到了 Transformer 工艺原理的潜力,不是说因为看到了 ChatGPT 才突然决定创业的。

Transformer 在当时还有一个特征,就是从语言往多模态进化了,笔者当时发现,如果 Token 外观得比较合理,所有的模态都可以在某一个空间去对齐,这和我之前在腾讯 Robotics X 机器人实验室做的机器人当中的数据驱动里是一个关键步骤。

既然要数据驱动,那就把所有数据对齐到某一个空间去,所以更早是看到了这些工艺趋势,才决定出来创业的。

Image

Q2:人类的身体和大脑是在漫长的进化过程中共同演变的,笔者的手、眼睛、肌肉和神经平台,实际上都在相互适应。星尘智能提出了 DFAI(Design for AI)的理念,强调为 AI 重新外观机器人的身体。那么在您看来,今天机器人行业是不是面临一个难题:机器人行业总是在考虑给机器人装进去 AI 的能力,却很少思考什么样的身体才真正合适 AI 学习?这是不是一种路线上的差异?

A:这实际上是根本上的差异,第一性原理上的差异。

有一个点是笔者一直在思考的,例如说手,人长成五指,老鹰长成前二后二的四指,为什么会这样?最大的原因是人和老鹰抓我的东西不一样。

抓什么东西又是谁来决定的?是脑子来决定的,也就是说智力决定了身体的构形长相,而不是说身体的构形长相决定了智力。当然其中有一些相辅相成的共同进化,老鹰也进化不出人类的智力出来。

《智能简史》里面有提到,人类智能最大一个进化的物理节点,是人的对生拇指的浮现,就是拇指自由度的变化。有了这个变化,人类才能适应多种工具,更好地发挥人类智能的潜力。

所以说,笔者一直所寻找的角色,就是笔者和智能之间的关系,笔者外观机器人一定是要协助它真正地成长,它的所有成长又可以被笔者的机器人所消化掉,笔者找的,就是这个循环关系。

星尘智能的 logo,有两颗星在循环,笔者会认为整个行业乃至整个人类的发展,实际上都会有两个东西在相互促进,计算机的发展就是软件不断消化 CPU 和 GPU 算力,算力又不断进化,然后继续被消化,形成相互的价值循环和传导,行业就能快速发展起来。

机器人和 AI 在社会变化层面看,本质上是智能的平权,智能如何进化,如何工作,会形成循环,这个循环里最大的突破点实际上在物理世界,笔者已经在虚拟世界投入了对智能的巨大热点,进入到物理世界,机器人开始自主地在这个环境中学习,那么真正的智能才会浮现。

Image

Q3:如果笔者把今天的人形机器人类比成一个正在成长的孩子,它可能已经学会了走路、跑步,甚至跳舞,有些领域甚至比普通人做得还好,但距离真正协助家人做家务,还有很长的距离。星尘智能从一开始就格外重视双臂协同、精细操作和与环境接触的能力。为什么会挑选这条路线?在您看来,机器人从「会动」发展到真正「会干活」,最大的工艺鸿沟是什么?

A:这也是跟笔者的理念相关,笔者机器人身体的每一处外观都是看 AI 能不能掌握它,如果笔者认为身体外观得很好,不过 AI 掌握不了,就像笔者当时的 Astribot S1 那样,有超高速度和超强负载,人去操控的话,危险系数有点高,当时的 AI 能力也驾驭不好,所以现在就会觉得身体和 AI 没有真正匹配起来。

整个机器人行业的发展中,像是宇树等企业,已经证明了机器人的身体能力可以做到足够强的,再往下的难题就如何提高智力。

现在 AI 这一波更多是智力革命,而不是说机器人革命。

智力革命意味着可以更好的把机器人当成一个工具,或者一个实践方式,去学习 AI 的智力,应用 AI 的智力。

所以笔者一直说笔者外观的机器人,大模型的数据采集者,也是大模型的应用终端,这是笔者对于星尘智能机器人的定位。

Image

难点和卡点在哪里?

我觉得有两块,第一块现在被应对了,这个难题在于之前大语言模型偏向于逻辑,因为语言的语料更多。后面原生大模型不把 LLM(大语言模型)和 ViT(Vision Transformer)模型分家,直接就是一起训练,这是很大的突破。机器人有了原生多模型能力之后,对于物理环境的抽象理解能力会有巨大的提高。

不过具象的理解能力还是缺失的,因为机器人首先缺这块的数据,现在互联网上的数据,还是数据采集的数据,都没有尺寸的概念,也缺少空间的意识,只有画面的意识。

缺少这层认知之后,机器人的智能会被卡住,应对思路只能是借助机器人给到完整的,偏向于空间的具象信息,所以这是笔者格外好的机器,笔者的机器人给大模型给出这些信息,大模型也会有更好的智力提高,智力提高之后会反哺机器人更好的工作。

这个飞轮今年已经开始浮现了,GPT 也开始训练 EGO(第一人称视角)数据,机器人数据了。

Image

Q4:给机器人装 AI 能力,还是围绕 AI 去外观机器人?有没有说哪个案例是证明后者比前者有一个先天长处?或者说只有后者能做到,不过前者就是做不到。

A:这里有一些笔者客户的案例能说明难题,不过很多人未必 get 到了。

例如最近有家互联网企业参加行业会议的时候,放了笔者的机器人做展示,我就刷到小红书还有一些平台上有使用者评价,这个机器人人味儿很足。

为什么人味儿很足?因为笔者的机器人在最开始的时候,涵盖关节外观,速度外观完全参考人地比例,当时的目的是因为笔者有了合适的比例,将来笔者可以直接用人类的数据,这样迁移的 gap 是最小的。笔者在腾讯 Robotics X 机器人实验室的时候,用调酒师的数据训练过 KUKA,UR 等机器人,最后笔者单独买了一台 ABB 的工业机械臂,不是因为它的工业可靠性,而是因为它的结构和人有点像,自由度和人一样,关节和人格外接近,所以就发现,在这个 ABB 工业机械臂上训练出来的模型成效更好。

为什么笔者要挑战这种仿人,然后用绳驱去达成,是因为笔者预知到最后结果是,当有大规模的人类数据之后,笔者的机器人的人感是最强的。

这种人感并不是说它的轨迹像人,而是真正的行为像人,行为像人就涵盖它的智力对于一些事情的规划,也涵盖它最后的动作达成。

Image

机器人有人感,本质上是它有类似于人的结构和想象力

Image

Q5:星尘智能的绳驱工艺让我联想到人类的肌肉和肌腱:肌肉负责产生力量,肌腱负责将力量传递到骨骼结构,从而带动关节运动。星尘的绳驱是不是也可以这样理解——借助绳索传递驱动力,让机器人具备更轻巧、更柔顺的关节?相比传统的电机、减速器传动,这套结构最大的长处是什么?它是否也存在一些类似人类肌腱损伤、疲劳或者需要持续调节张力的难题?

A:实际上笔者最早外观绳驱思路的时候,还没有考虑到跟肌肉肌腱的对应关系,笔者当时只是为了找一个力透明度最好的思路。

什么是力透明度呢?就例如说人和人握手,你会感觉你和对方的动作是伴随的,和谐的,不过人跟机器人握手,就不是你在掰它,就是它在掰你,中间没有随动感。

不只是握手这个动作,例如现在很多机器人夹住咖啡杯放下,都是判断咖啡杯快接近桌面的时候松手,然后啪的一下,咖啡杯落在桌子上。人不会这样,人会在感觉到桌子接触并支撑了杯子的时候才松手。

这里面的难题就在于很多机器人的电机结构没法感受到末端力的变化,所以只能让杯子接近桌子的时候让它落下。

笔者做绳驱是想找一种传动机构,使这个力的传递过程更透明,在里面学到更深的智能关系,并不是靠绝对的位置关系来判断放多高松手杯子不会倒,这个不是智能。智能是感觉到杯子被桌子支撑了,手上不用出太多力了,可以松手了,这个才叫智能。

这是绳驱的本质。

Image

然后在研发的过程中,就会发现绳驱在力透明度高这个好处之外,它还和人的构形很像,使得它的整个结构不用把电机放在关节,放在手腕这些地方,它没有了这个约束,就可以把电机放在小臂当中,借助绳来传导,借助「肌腱」传导到关节。

2024 年笔者做绳驱国产化的时候,发现了一个更大的惊喜,就是全球大概 70% 的绳驱源头供应商都在中国,当时发现成本格外低,如果拿谐波减速器比较的话,笔者全身到下所有的绳,它实际上部分取代了谐波的这个减速的作用, 20 多个关节加起来才是刚性机器人 1 个谐波减速器的售价。,笔者的成本可以快速地往下打。

这就是绳驱的长处点,力透明度高,构形好,并且成本还格外低。

最开始的时候,笔者也有担心绳驱的强度耐久难题,一开始的话,质量体系建设就比较强,供应商的样品在笔者这里要经过 2500 万次的弯折,相当于在使用负载在弯胳膊弯 2500 万次,一天 24 小时连续运动 6 年时间,评估最后的偏差值。

Image

Q6:如果基础大模型越来越强,是否终究会吞噬具身智能公司的价值?

A:不会。

星尘智能机器人的架构笔者叫「三位一体」,除了机器人本体之外,操作平台 AOS 是一块,另一块就是具身模型。

这里面有两个难题,一是具身模型和 GPT6 的关系,二是 AOS 和 GPT 6 的关系。

2023 年的时候,别人问我为什么要训练模型,这个模型要做什么?我当时有个比喻,我说具身模型这一块,我要做的是一个能听懂人话的猴子,能听懂人话,就证明它有语言接口,但更多偏向于自己在物理世界的物理反应上面。

更贴近于人的解释是,例如我不会做饭,但我又想做,怎么办?我就给我妈打视频电话,我妈在视频里开始指导,说先在热锅里放点油,但她不用说特别细,不用说哪瓶油。如果我有疑问,我再来问她是哪一瓶是油。

意思就是,我和我妈的交互是补充上下文知识,不过放油这个动作,我本身要学会,因为我妈在电话那头没法给我放油。

所以说,笔者的具身模型要做什么?它一定要有自己的落下,要把现实的状态反馈为自然语言的方式,去和基础大模型提问,去和 GPT 6 提问。它自己也要在物理的环境中感知物理的参数,物理参数的话涵盖重量、捏的力度、有没有抓稳之类的这些东西。

所以说,我不太认可是不是 GPT 这样的基础大模型越来越强之后,把什么触觉的信息也训到那边去,不是说做不到,而是成本不合适,并且通讯的带宽也会有难题。更合适的模式是笔者的具身模型抓完瓶子之后,会问 GPT 6,说我捏不住有点不稳该怎么办,GPT 再给它思路。

笔者不是要把整个的这个智能世界分层,而是有点像智能体化。

智能体化的话就有点像蚂蚁,每一个智能体能力未必强,不过凑一起的话就会很强大,笔者也一直强调,不是要把任务交给 GPT 去做,只做下面的接口,做被动的事情。

Image

Q7: 现在具身智能行业越来越重视世界模型。世界模型是不是有点像人类的「想象力」:例如笔者准备倒一杯水时,脑海中可能已经大概预判了水会怎样流动、倾倒角度怎么影响水流速度,杯子什么时候会满,因而不需要每一步都经过精确计算。星尘智能的 Lumo-2 具身模型也具备了隐式世界动态预测能力,Lumo-2 应对了什么样子的难题?隐式是什么意思?它和人类的想象力有什么异同?

A:Lumo-2 的预测能力,和人类的想象力格外接近,不过中间有一个世界模型发挥的点。

人在做一个动作的时候,一直想的是,我预期有一个目标,然后做的是跟我目标之间误差的修正。

VLA(视觉语言动作)大模型是什么?它实际上没有目标,只是把动作开展重新的复现,所以它对误差不敏感,有了世界模型之后,机器人只要预测该怎么做,中间执行过程中的差异都可以想办法去纠正。

既然已经产生了误差,目标是纠正它,那这不是最合适强化学习做的事吗?

VLA 模型的模仿能力,现在加上世界模型和强化学习,笔者搭建好试错和纠错的闭环,具身智能的结构就是完整的了。

这个隐式世界动态预测能力当中的隐式是什么意思呢?就是说世界模型是物理世界的一个重建,重建完了之后,为了表现出来,后面还有一大段是渲染,无论是是渲染成视频,还是渲染成图片,都是为了让人看懂。

笔者发现,物理世界的重建是笔者机器人需要的,最后的渲染不一定是机器人需要的。就像人在做梦的时候,梦境是模模糊糊的,但行为却是透传的。所以笔者回忆做梦的时候,环境往往是模糊的,不过做了什么很清楚。

类似的,笔者说的隐式世界模型,就是说没必要完全做给人看,机器人需要的关键是要从中去找着跟自己的动作相关的预测跟偏差。

现在的星尘智能机器人还没到 iPhone 4 那种水平,但很快就到了

Image

Q8:为什么机器人需要自己的操作平台?笔者怎么去定义 AOS,它面向的是怎么样的一个受众?

A:笔者开始创业的时候,还没有具身智能这个概念,不过具身智能的发展,我觉得必须要浮现 Macintosh 这样的机型,当时我认为,笔者要在行业发展过程中寻找 2 个东西,第一个就是对应于电脑的图形交互界面,对于机器人来说,除了正常的命令操作图形之外,能调用智能能力 AI 能力的地方, 会形成一个新的交互界面。

所以笔者的 AOS 里有一层就是去和基础大模型,还有笔者自己的具身模型开展接口性的规划。

AOS 不只是想到了开发者,而是想把开发门槛降低到普通人,现在很多人开始用自然语言去生成 PPT,视频乃至应用,那么为什么不可以让普通人用 AI 给笔者的机器人生成相关的能力和动作,这就是笔者的目标,想要什么,只要跟它说一些提示词,让机器人自己去编译生成,把使用门槛和开发门槛降低到所有人都行。

所以,笔者的 AOS 不只要让程序员看懂,也要让普通人看懂,当然,也要让 GPT 看懂。

Image

Q9:星尘智能很喜欢用苹果的「硬件,平台和应用生态」来类比自己的「三位一体」,如果用苹果的机型史类比的话,星尘智能的当下的 S1 机器人,T1 机器人,还有 AOS 操作平台,大概相当于苹果的哪个阶段,有没有一些对应关系?苹果的硬件,尤其是操作平台有很清晰的工艺传承脉络,例如 iOS 可以认为是移动简易版的 macOS,都是基于 UNIX 的图形交互平台,而 macOS 又源于更早乔布斯做的 NeXTSTEP 平台。

A:这个观察太对了,实际上我一直所把控的,在公司层面,就是传承,就是里面有哪些东西是统一化的?

关于和苹果软硬件平台的对应关系,笔者也有过仔细的思考,我是把苹果的两次发展背景环境做了一定的筛选,笔者发现它的路径和做法是有相似性的。

在 iPhone 诞生的时期,互联网基础设施是完整的,所以 iPhone 可以爆发,不过在 Macintosh 诞生的事情,没有互联网,所以它更多是一个硬件终端。

不过要说笔者现在是个什么程度,那恐怕要让你失望了。

S1 的话,对应电脑阶段,最多对应到 Apple I,就卖了 300 来台的那个东西,苹果是从 Apple II 开始卖火的,Apple II 对应的是笔者的 T1,但还没有进入到 Macintosh,就是它还没有属于自己的完整图形交互界面。

对应手机的话,S1 和 iPhone 初代也是有点像,iPhone 初代颠覆了键盘手机,平台具有开创性但并不完善,电池续航也不好,本身的不足和面临的质疑格外多。

但你说 T1 对应 iPhone 4,我觉得有点勉强,因为 iPhone 4 人们公认格外 OK 了。

不过有一点就是,这个行业发展的速度连笔者行业内的人,都没预期到会这么快,反而是说,Macintosh 啊,iPhone 4 啊这样的爆款机器人机型,就会在明年浮现。

Image

▲ 星尘智能的机器人正在商场卖货

Q10:具身智能可能第一次让机器大规模参与普通人的体力劳动。您曾经提到,希望将来数十亿人都能具备 AI 机器人助理。但家庭是一个高度复杂、充满意外的环境,如果把今天的机器人看成一个刚刚开始学习世界的孩子,那么它究竟需要成长到什么程度,才能真正独立进入笔者的家庭?如果有一天这个愿景达成了,您最希望机器人协助人类摆脱什么,又最不希望它改变人类生活中的什么?

A:先回答第一个难题,按照现在的工艺发展,可能比笔者想象的近很多,我认为很有可能明年机器人就会进入家庭。

进入家庭的关键因素,从商业上来看,就是成本因素和能力因素,中间有个对等的关系,如果机器人成本足够低,那能力就不会强,不过如果能力不够强,进入家庭可以做什么呢?我觉得不只是帮忙拿拿东西这些,肯定还要有一些情绪价值,情绪价值还是能卖钱的。

我希望它有一个啥都可以做硬件平台,然后有一个啥都能做的可发展操作平台,笔者会给它开发一些客户所需要的基本东西,就像你买了 iPhone,上面肯定有电话短信相机能力,后面笔者要留给足够的空间给开发者,在相机之上还可以开发美颜相机,都是借助基础硬件能力,还有操作平台的开发接口。

使用者在买到机器人之后,就像刚开始使用智能手机那样,每周都会下几个 App 去尝试下新能力,知道使用者借助这一堆的 App 发现,诶,这个机器人好像什么都能干,有一个摸索和演化的过程,而不是说机器人弄到家里面就是个什么都会的保姆。

Image

第二个难题,就是我一直说的,我不想让笔者的机器人进场打螺丝。我认为,机器人最大的长处,是它的专业性是不可能封闭在一个领域。例如说现在的大模型,它可以是很好的工作助理,也是很好的搜索引擎,医学的,管理的,历史的,科学的都可以问,问啥都知道,它的长处在于知识的广泛性。

机器人也是类似,进入到家庭,它的角色不一定是一个有强迫症的保姆,可以是一个有外观师思维的收纳师,它做家务前可以问你,家居环境要收拾成什么风格,一丝不苟的,还是随性惬意的?

它会十国语言,可以教小朋友外语,给小朋友讲故事。

人累了,它也可以代替人去遛狗;做饭前可以给膳食营养建议之类等。

笔者不会说现在的电脑就只能用 Word,Excel 和 PPT 三件套,现在人们用电脑这一种硬件,但上面有无数的软件游戏。这就是我为什么说机器人只进厂打螺丝太窄了。

还有一点,机器人进场打螺丝,token 费老板们都未必愿意交,要我的话肯定不愿意,我希望打螺丝的是固定化程序,每天 24 小时一直转就像,别打一个螺丝收我一份 token 费。

所以说,进厂打螺丝不是笔者期待的方向。

本文来源:https://www.ifanr.com/1687805
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

When I want to drink a cola, for a very long time I only had two choices: Pepsi and Coca-Cola. When I choose a mobile phone, there is a 90% probability that I will oscillate among brands like Apple, Huawei, Xiaomi, Oppo, and Vivo. When I buy sportswear, the first thing that comes to mind is most likely Nike or Adidas.

But the reason the world is colorful and diverse is that, beyond these giants, there exist companies that do not follow tradition, strive to create difference, focus on aesthetics and capability, and look toward tomorrow.

They possess non-mainstream business models, and their appearance, craftsmanship, and models can offer unique user value and sufficient social talking points. The key is that they do not carry the baggage of large corporations and are brave enough to progress without reservation. They are the "Diversity Companies."

Diversity is the key to an open world. iFanr believes that only by truly focusing on and understanding Diversity Companies can we see the future earlier than most. In this namesake column, iFanr will use the format of exclusive interviews to witness with you how these Diversity Companies reshape the future and define the new normal.

Image

▲ Lei Jie, Founder and CEO of Astribot

This is the 15th article in the "Diversity Companies" column.

Even though embodied AI startups are currently as numerous as the stars, Astribot is the one that stands out sufficiently. This Shenzhen-based embodied AI company chose "cable-driven" technology for its robot body technology route, forming a "trinity of cable-driven robot body, embodied operation platform AOS, and embodied model" architecture in terms of overall technology.

The unique significance of "cable-driving" lies in the robot transmitting power via high-strength cables, moving heavier drive components like motors away from the joints and placing them in the robot's base. This reduces the weight on the end-effector arms and brings the advantages of low friction and high force-control transparency.

Astribot's AOS operation platform encapsulates complex robot capabilities such as perception, movement, operation, and interaction into skill module packages. It is compatible with cutting-edge large models and Agents like GPT 6. Developers can program in natural language, leveraging the capabilities of large models and Agents to develop applications for robots on AOS, endowing robots with practical application capabilities such as writing calligraphy, camera tracking, cleaning up cat litter, and playing mahjong.

Lei Jie, Founder and CEO of Astribot, was previously employee number 1 at the Tencent Robotics X robotics lab and is also an Apple fan. He hopes that Astribot's "trinity architecture of robot, model, and platform" can form a complete ecosystem just like Apple's "iPhone hardware, iOS platform, and Apple Store application ecosystem," thereby achieving his and the company's vision: to let billions of people possess AI robot assistants. They hope for humans and robots to coexist, co-create, and win together, explicitly opposing "robots working in factories." This is perhaps a "humanism" with a double meaning: neither hoping robots replace humans nor hoping robots are trapped in closed workshops.

I chatted with Lei Jie, Founder and CEO of Astribot, about why robots need an operation platform, why the threshold for robot application development should be lowered infinitely, and why he doesn't want robots working in factories.

Image

Giving intelligence a body, or giving a body intelligence?

Q1: At the end of 2022, you chose to leave a big tech company to found Astribot. At that time, ChatGPT had just been released a few days prior, and the vast majority of people had not yet realized the arrival of the AI wave. What made you judge at that point that the robot industry had reached a stage worthy of entrepreneurship?

A: Actually, I started thinking about starting a business as early as the first half of 2022.

The technological foresight for starting the business was even earlier; it was mainly because I saw the potential of the Transformer technology principle. It wasn't a sudden decision to start a business just because I saw ChatGPT.

At that time, Transformer had another characteristic: it was evolving from language towards multimodality. I discovered at the time that if the Token representation looks reasonable, all modalities can be aligned in a certain space. This is a key step in the data-driven approach I did previously in the Tencent Robotics X robotics lab.

Since we want to be data-driven, then let's align all data to a certain space. So, it was seeing these technology trends earlier that made me decide to start a business.

Image

Q2: The human body and brain have evolved together in a long evolutionary process. My hands, eyes, muscles, and neural platforms are actually mutually adapting. Astribot has proposed the DFAI (Design for AI) concept, emphasizing redesigning the robot's body for AI. So, in your view, does the robot industry today face a problem: the industry is always considering putting AI capabilities into robots, but rarely thinks about what kind of body is truly suitable for AI to learn? Is this a difference in route?

A: This is actually a fundamental difference, a difference in first principles.

There is one point I have always been thinking about. For example, the hand. Humans grow five fingers, while eagles grow four fingers—two front and two back. Why is this? The biggest reason is that humans and eagles catch different things.

Who decides what to catch? The brain decides. That is to say, intelligence determines the physical form of the body, not that the physical form of the body determines intelligence. Of course, there is some mutually reinforcing co-evolution in this, but eagles won't evolve human intelligence.

"A Brief History of Intelligence" mentions that the biggest physical node in the evolution of human intelligence was the emergence of the human opposable thumb—that is, the change in the thumb's degree of freedom. With this change, humans were able to adapt to multiple tools and better utilize the potential of human intelligence.

So, the role I have always been looking for is the relationship between me and intelligence. The robots I design must assist in its real growth, and all its growth can be digested by my robots. What I am looking for is this cyclical relationship.

Astribot's logo has two stars cycling. I believe that the development of the entire industry and even all of humanity actually involves two things promoting each other. The development of computers is a continuous cycle of software digesting CPU and GPU computing power, and computing power constantly evolving, then continuing to be digested, forming a mutual value cycle and transmission. Only then can the industry develop rapidly.

Looking at social changes, robots and AI are essentially the democratization of intelligence. How intelligence evolves and works will form a cycle. The biggest breakthrough point in this cycle actually lies in the physical world. I have already invested huge enthusiasm in intelligence in the virtual world. Entering the physical world, robots begin to learn autonomously in this environment, and only then will true intelligence emerge.

Image

Q3: If I compare today's humanoid robots to a growing child, it may have learned to walk, run, and even dance. In some areas, they even do better than ordinary people. But there is still a long distance from truly assisting family members with housework. Astribot has placed special importance on dual-arm coordination, fine manipulation, and the ability to interact with the environment from the very beginning. Why did you choose this route? In your view, what is the biggest technological gap for robots to develop from being "able to move" to truly "being able to work"?

A: This also relates to my philosophy. The appearance of every part of my robot's body is designed to see if an AI can master it. If I think the physical appearance is good, but the AI cannot master it—like with the Astribot S1 back then—which had super high speed and super strong payload, the danger factor for human control was a bit high. The AI capabilities back then couldn't handle it well either, so I now feel that the body and the AI weren't truly matched.

In the development of the entire robotics industry, companies like Unitree have proven that a robot's physical capabilities can be made sufficiently strong. The next challenge is how to improve its intelligence.

The current wave of AI is more of an intelligence revolution, rather than a robotics revolution.

An intelligence revolution means we can better treat the robot as a tool, or a practical method, to learn AI's intelligence and to apply AI's intelligence.

So I have always said that my robot is both a data collector for large models and an application terminal for large models. This is my positioning for the Astribot robot.

Image

Where are the difficulties and sticking points?

I think there are two parts. The first part is now being addressed. The previous problem was that large language models leaned towards logic, because language corpora were more abundant. Later, native large models no longer separated the LLM (Large Language Model) and ViT (Vision Transformer) models; they trained them together directly, which was a major breakthrough. With native multi-model capabilities, a robot's ability to abstract and understand the physical environment will be hugely improved.

However, concrete and specific understanding is still lacking because, first and foremost, robots lack this kind of data. Data currently on the internet, whether from web scraping or data collection, has no concept of scale and lacks spatial awareness, possessing only visual awareness.

Lacking this layer of cognition will hold back a robot's intelligence. The only solution is to use the robot to provide complete, physically-oriented, concrete information. This is why I designed such a machine. My robot provides this information to the large model, which will then experience improved intelligence. This improved intelligence will, in turn, enable the robot to work better.

This flywheel has already started to emerge this year. GPT has also begun training on EGO (egocentric perspective) data and robot data.

Image

Q4: Is it about giving a robot AI capabilities, or designing a robot around AI? Is there a case that proves the latter has an inherent advantage over the former? Or is it that only the latter can do something the former simply cannot?

A: There are some cases from my clients here that can illustrate this point, but not many people may have gotten the message.

For example, recently an internet company showcased my robot at an industry conference. I came across comments from users on Xiaohongshu and other platforms saying that this robot feels very human.

Why is it so human-like? Because when we first designed my robot, the joint design and speed design were completely based on human proportions. The purpose at the time was that if we had the right proportions, we could directly use human data in the future, thus minimizing the transfer gap. When I was at the Tencent Robotics X lab, I used data from a bartender to train robots like KUKA and UR. Later, I specifically bought an ABB industrial manipulator, not for its industrial reliability, but because its structure is somewhat like a human's—it has the same degrees of freedom as a human, and its joints are particularly similar to human ones. We discovered that models trained on this ABB industrial manipulator had better results.

The reason I wanted to pursue this humanoid approach and use cable-driven technology to achieve it is that I predicted the end result would be that, when large-scale human data becomes available, my robot's human-likeness would be the strongest.

This human-likeness doesn't mean its trajectories are like a human's, but that its actions are truly like a human's. Acting like a human involves its intelligence in planning for tasks, as well as the final execution of actions.

Image

A robot's human-likeness fundamentally stems from its human-like structure and imagination.

Image

Q5: Astribot's cable-driven technology reminds me of human muscles and tendons: muscles generate force, and tendons transmit that force to the skeletal structure, thus driving joint movement. Can Astribot's cable-driven system be understood in the same way—transmitting driving force through cables to give the robot lighter, more fluid joints? Compared to traditional motor and reducer transmissions, what are the biggest advantages of this structure? Does it also face problems similar to human tendon injuries, fatigue, or the need for continuous tension adjustment?

A: Actually, when I first conceived of the cable-driven approach, I hadn't yet considered its correspondence to muscles and tendons. At the time, I was simply looking for an approach with the best force transparency.

What is force transparency? For instance, when you shake hands with another person, you feel that your movements are synchronized and harmonious. But when you shake hands with a robot, it's either you wrestling with it, or it's wrestling with you. There is no sense of responsive, fluid movement in between.

It's not just the action of shaking hands. For example, many robots today pick up a coffee cup and place it down by simply releasing it when they judge it's close to the table, and then *smack*, the coffee cup lands on the table. Humans don't do this. A person lets go only when they sense that the table has made contact with and is supporting the cup.

The problem here is that the motor structures of many robots cannot sense changes in the force at the end-effector, so they can only let the cup drop when it approaches the table.

I pursued cable-driven technology because I wanted to find a transmission mechanism that would make the force transmission process more transparent, allowing the system to learn a deeper intelligent relationship. Intelligence is not about using absolute positioning to determine how high is safe to release a cup without it tipping over; that is not intelligence. Intelligence is sensing that the cup is supported by the table, that not much force is required from the hand anymore, and that it's time to let go. That is what is called intelligence.

This is the essence of cable-driven technology.

Image

Then during the R&D process, we discovered that besides the advantage of high force transparency, the cable-driven approach also mimics human form. This means the entire structure doesn't need to place motors directly in the joints or places like the wrist. By removing this constraint, motors can be placed in the forearm, with power transmitted to the joints via cables, like 'tendons'.

In 2024, when I was working on localizing string-driven actuation, I found an even bigger surprise: roughly 70% of the world’s upstream suppliers for string-drive components are in China. Back then the costs were remarkably low. If you compare it with harmonic reducers, for example, all the strings throughout my robot actually partially replace the speed-reduction function of a harmonic reducer. The total cost of more than 20 joints combined is equal to the price of just one harmonic reducer in a rigid robot. That means my costs could be pushed down very quickly.

This is the strength of string-driven actuation: high force transparency, good form factor, and very low cost.

At the beginning, I was also worried about string-drive strength and durability. But from the start we built a strong quality system. Supplier samples had to undergo 25 million bend cycles at my end—equivalent to bending your arm 25 million times under operational load—continuous motion for 24 hours a day over six years. We then evaluated the final deviation.

Image

Q6: As foundation models keep getting stronger, will they eventually swallow up the value of embodied AI companies?

A: No.

The architecture of Astribot robots I call a “trinity.” Besides the robot body, there is the operating platform AOS, and another part is the embodied model.

There are two issues here: first, the relationship between the embodied model and GPT6; second, the relationship between AOS and GPT6.

In 2023, people asked me why we needed to train a model and what that model was supposed to do. I used a metaphor at the time: for the embodied model, what I want to build is a monkey that can understand human instructions. Being able to understand instructions proves it has a language interface, but it is more focused on its own physical responses in the physical world.

A closer-to-human explanation is this: suppose I can’t cook but I want to. What do I do? I call my mom on a video call, and she starts guiding me over video, saying, “First put some oil in a hot pan.” She doesn’t have to be extremely detailed, like which bottle is oil. If I’m unsure, I can ask her which one is oil.

The point is that the interaction between my mom and me supplements contextual knowledge. But for the action of pouring oil, I have to learn it myself, because my mom can’t pour oil for me on the other end of the phone.

So what does my embodied model need to do? It must be able to act on its own and turn real-world states into natural-language prompts to ask the foundation model, to ask GPT6. It also needs to perceive physical parameters in a physical environment—parameters such as weight, pinch/grip force, whether the grasp is secure, and so on.

So I don’t really agree with the idea that once foundation models like GPT get stronger, we should just train haptic information into them as well. It’s not that it can’t be done; the cost isn’t appropriate, and there are also communication bandwidth issues. A better model is this: after my embodied model picks up a bottle, it asks GPT6, “I can’t hold it securely and it feels unstable—what should I do?” GPT then provides guidance.

I’m not trying to stratify the entire intelligent world; it’s more like making it agent-based.

Agent-based is a bit like ants: each agent may not be particularly strong, but together they become very powerful. I’ve always emphasized that we shouldn’t simply hand tasks to GPT and only build the interface below it to do passive things.

Image

Q7: The embodied AI industry is paying more and more attention to world models. Is a world model somewhat like human “imagination”: for example, when I am about to pour a glass of water, my mind may already predict roughly how the water will flow, how the pouring angle affects flow rate, and when the cup will be full—so I don’t need precise computation at every step. Astribot’s Lumo-2 embodied model also has implicit world dynamic prediction capability. What kind of problem does Lumo-2 address? What does “implicit” mean? How is it similar or different from human imagination?

A: Lumo-2’s prediction ability is quite close to human imagination, but there is a point where the world model comes into play.

When a person performs an action, they keep thinking in terms of this: I have an intended goal, and what I do is correct the error between my actions and that goal.

What is a VLA (Vision-Language-Action) large model? In reality, it doesn’t have a goal; it just reproduces actions through replay/imitation. So it is not sensitive to error. Once you have a world model, the robot only needs to predict what to do, and differences during execution can be corrected along the way.

If an error has already been produced and the goal is to correct it, isn’t that exactly what reinforcement learning is best suited for?

Combining the imitation capability of VLA models with a world model and reinforcement learning, once I have set up the closed loop of trial-and-error and error correction, the structure of embodied intelligence is complete.

In this “implicit world dynamic prediction” capability, what does “implicit” mean? It means the world model is a reconstruction of the physical world. After reconstruction, to make it visible, there is a whole later stage of rendering—whether rendered into video or images—all for human comprehension.

I found that reconstructing the physical world is what my robot needs, but the final rendering is not necessarily something the robot needs. It’s like when people dream: the dream scene is often blurry, but actions still “pass through” clearly. So when I recall dreaming, the environment is usually vague, but what I did is very clear.

Similarly, what I mean by an implicit world model is that it doesn’t need to be fully made for humans to watch. What the robot needs is the key ability to extract predictions and deviations relevant to its own actions.

Astribot robots today are not yet at the level of the iPhone 4, but they will get there very soon.

Image

Q8: Why do robots need their own operating platform? How do you define AOS, and who is its audience?

A: When I first started the company, the term “embodied intelligence” didn’t exist yet. But I believed that as embodied intelligence developed, a Macintosh-like machine would inevitably emerge. At that time I thought we needed to find two things during the industry’s development. The first is something analogous to a computer’s graphical user interface. For robots, besides normal command-and-control interfaces, places where AI capabilities can be invoked will form a new kind of interaction interface.

So one layer inside AOS is to interface and plan jointly with foundation models and with our own embodied model.

AOS isn't just thinking about developers; it aims to lower the development threshold for ordinary people. Now, many people have started using natural language to generate PPTs, videos, and even applications. So why can't ordinary people use AI to generate relevant capabilities and actions for my robots? This is my goal. If you want something, just tell it a few prompts, and let the robot compile and generate it on its own, lowering the barriers to use and development so that anyone can do it.

Therefore, my AOS must not only be understood by programmers but also by ordinary people, and, of course, by GPT as well.

Image

Q9: Astribot likes to use Apple's "hardware, platform, and application ecosystem" as an analogy for its own "trinity". If we use Apple's product history as a comparison, which stage of Apple do Astribot's current S1 robot, T1 robot, and AOS operating platform roughly correspond to? Are there any connections? Apple's hardware, especially its operating platforms, has a clear lineage. For example, iOS can be considered a simplified mobile version of macOS; both are graphical interaction platforms based on UNIX, and macOS originated from the NeXTSTEP platform created by Jobs much earlier.

A: That observation is spot on. In fact, what I've always been controlling at the company level is continuity—what elements can be unified?

My colleagues have also carefully considered the correlation with Apple's hardware and software platforms. I filtered through Apple's two periods of development and found that our path and approach have similarities.

When the iPhone was born, the internet infrastructure was complete, so the iPhone could take off. However, when the Macintosh was born, there was no internet, so it was more of a hardware terminal.

However, if you want to know what stage we are at right now, I'm afraid I have to disappoint you.

As for the S1, in the computer stage, it corresponds at most to the Apple I—the thing that sold only about 300 units. Apple started to become popular with the Apple II, which corresponds to my T1, but it hasn't reached the Macintosh stage yet; it doesn't have its own complete graphical user interface.

Regarding phones, the S1 is also a bit like the first-generation iPhone. The first-generation iPhone disrupted the keyboard phones, and the platform was groundbreaking but not perfect; the battery life was poor, and it faced exceptionally high levels of inadequacy and skepticism.

But saying that the T1 corresponds to the iPhone 4, I feel is a bit of a stretch, because the iPhone 4 was widely accepted as exceptionally good.

However, one thing is that the speed of development in this industry is so fast that even those in the industry, including myself, did not expect it. Instead, blockbuster robot models like the Macintosh or iPhone 4 will emerge next year.

Image

▲ Astribot's robots are selling goods in a mall

Q10: Embodied intelligence may be the first time machines participate in the physical labor of ordinary people on a large scale. You once mentioned hoping that billions of people will have AI robot assistants in the future. But a home is a highly complex environment full of unexpected situations. If we view today's robots as a child just beginning to learn about the world, how much does it need to grow before it can truly live independently in people's homes? If this vision is realized one day, what do you most hope robots help humans get rid of, and what do you least hope they change in human life?

A: Let me answer the first difficult question first. According to current technological development, it may be much closer than I imagined. I think it is very likely that robots will enter homes next year.

The key factors for entering the home, from a commercial perspective, are cost and capability, which have an inverse relationship. If the robot's cost is low enough, its capability won't be strong. But if the capability isn't strong enough, what can it do in the home? I think it's not just about helping to carry things; it definitely needs to provide some emotional value, and emotional value can still sell.

I hope it has a hardware platform that can do anything, and a scalable operating platform that can do anything. I will develop some basic functions that customers need, just like when you buy an iPhone, it certainly has the capability for phone calls, text messages, and cameras. Later, I want to leave ample room for developers; on top of the camera, a beauty camera can be developed, all relying on the basic hardware capabilities and the operating platform's development interface.

After users buy the robot, just like when they first started using smartphones, they will download a few apps every week to try out new capabilities. Through these apps, users will realize, "Hey, this robot can seem to do anything." It will be a process of exploration and evolution, rather than the robot being brought home as a perfect all-around nanny immediately.

Image

The second difficult question is what I've always said: I don't want to put my robots in factories to tighten screws. I believe the greatest strength of robots is that their professionalism cannot be confined to a single domain. For example, today's large models can be a great work assistant and an excellent search engine for medicine, management, history, and science—you can ask anything, and it seems to know everything; their strength lies in the breadth of knowledge.

Robots are similar. When entering the home, their role doesn't necessarily have to be a nanny with OCD; it can be an organizer with the mindset of an interior designer. Before doing housework, it can ask you: what style do you want the home environment to be arranged in? Meticulous and strict, or casual and cozy?

It can speak ten languages, teach children foreign languages, and tell them stories.

When people are tired, it can also take the dog for a walk on their behalf; before cooking, it can give dietary and nutritional advice, and so on.

I wouldn't say that today's computers can only use the Word, Excel, and PPT suite. Now people use computers as one piece of hardware, but on top of them are countless software games. This is why I say it's too narrow to only put robots in factories to tighten screws.

Furthermore, if robots were in factories tightening screws, bosses might not even be willing to pay the token fees. If it were me, I certainly wouldn't; I want the screw tightening to be a fixed program, running continuously 24 hours a day, not charging me a token fee for every single screw.

Therefore, going into factories to tighten screws is not the direction I look forward to.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry